J Neurol Neurosurg Psychiatry: first published as 10.1136/jnnp-2020-325347 on 9 June 2021. Downloaded from http://jnnp.bmj.com/ on September 25, 2023 by guest.

Protected by copyright.
General neurology

Review

Retinal imaging in Alzheimer’s disease

Carol Y Cheung ‍ ‍,1 Vincent Mok,2 Paul J Foster,3 Emanuele Trucco,4
Christopher Chen ‍ ‍,5,6 Tien Yin Wong7,8

►► Additional supplemental ABSTRACT vivo by abnormal biomarkers of cerebral amyloid-β

material is published online Identifying biomarkers of Alzheimer’s disease (AD) will deposition and pathologic tau and treats cognitive
only. To view, please visit
the journal online (http://​dx.​ accelerate the understanding of its pathophysiology, impairment as a symptom or sign of the disease.9
doi.​org/​10.​1136/​jnnp-​2020-​ facilitate screening and risk stratification, and aid in This implies the cognitive function of a person
325347). developing new therapies. Developments in non-­invasive affected by AD can thus range from cognitively
retinal imaging technologies, including optical coherence unimpaired (preclinical AD) to mild cognitive
For numbered affiliations see
tomography (OCT), OCT angiography and digital retinal impairment (MCI) (AD MCI), to dementia (AD
end of article.
photography, have provided a means to study neuronal dementia).9 Although this latest framework encour-
Correspondence to and vascular structures in the retina in people with ages in vivo detection and study of AD at an earlier
Dr Carol Y Cheung, Department AD. Both qualitative and quantitative measurements stage (eg, preclinical AD or AD MCI), current tech-
of Ophthalmology and Visual from these retinal imaging technologies (eg, thinning nologies to detect amyloid-β and tau pathology
Sciences, The Chinese University of peripapillary retinal nerve fibre layer, inner retinal using positron emission tomography (PET) brain
of Hong Kong, Hong Kong,
layer, and choroidal layer, reduced capillary density, imaging or cerebrospinal fluid (CSF) examina-
China; c​ arolcheung@​cuhk.​
edu.h​ k abnormal vasodilatory response) have been shown to tion are limited due to their high cost, technical
be associated with cognitive function impairment and complexity, invasiveness of the procedures and/
Received 7 October 2020 risk of AD. The development of computer algorithms for or the necessity of using radioactive tracers. Thus,
Accepted 27 May 2021 respective retinal imaging methods has further enhanced
Published Online First 9 June identifying alternative, more accessible technol-
2021 the potential of retinal imaging as a viable tool for rapid, ogies and biomarkers of preclinical AD prior to
early detection and screening of AD. In this review, we onset of cognitive impairment may accelerate the
present an update of current retinal imaging techniques understanding of the pathogenesis of AD, facilitate
and their potential applications in AD research. We also screening and stratification of risk, and ultimately
discuss the newer retinal imaging techniques and future aid in the discovery, development and testing of
directions in this expanding field. new treatments or preventive therapies in clinical
trials.3
The retina, a neurosensory layered tissue lining
the back of the eye and directly connected to the
INTRODUCTION
brain via the optic nerve, receives light that the lens
Alzheimer’s disease (AD), the most common form
has focused, converts the light into neural signals
of dementia, is a major public health and clinical
and sends these signals on to the brain for visual
challenge globally.1 Despite decades in research,
recognition. The retina has long been considered a
the pathophysiology of AD remains unclear. The
‘platform’ to study disorders in the central nervous
current thinking is that the neuropathology of
AD, as characterised by accumulation of misfolded system (CNS), as it is an accessible extension of the
amyloid-β and tau protein, begins years before the brain in terms of embryology, anatomy and phys-
onset of clinical symptoms. Given this long natural iology (box 1).12 Figure 1 shows the optic nerve
history, there are opportunities for early disease head, macular area, nerve fibre layer, arterioles
detection and thus timely intervention.2 3 Indeed, and venules captured from a retinal photograph.
recent clinical trials have suggested the efficacy Similar to the neurovascular unit (NVU) in the
of certain measures (eg, lifestyle interventions or CNS, the retinal NVU contains neurons (ganglion
medication) in improving symptoms or slowing cells, amacrine cells, horizontal and bipolar cells),
progression of AD.4 glial cells (Müller cells and astrocytes) and vascular
The definition of AD has also evolved in the cells (endothelial cells and pericytes).13 Evidence
past decade with the discovery of novel in vivo of retinal involvement in AD dementia has been
biomarkers for AD.5 6 It has been shown that clin- shown in histopathological studies of postmortem
ically diagnosed cases of dementia presumably due specimens.14,S1 Associations of AD dementia with
to AD may be amyloid-­negative in up to 25% of common eye diseases with overt clinical signs,
cases.7 8 Thus, biomarker confirmation has been such as age-­related macular degeneration (AMD),
© Author(s) (or their
proposed to improve the precision of AD diagnosis diabetic retinopathy (DR) and glaucoma, have also
employer(s)) 2021. No
commercial re-­use. See rights and now even biomarkers are indispensable for an been reported. In addition to digital retinal photog-
and permissions. Published AD diagnosis.9–11 The evolution in definition and raphy (figure 1), recent advances in non-­invasive
by BMJ. diagnostic criteria of AD and other dementias may retinal imaging technologies allow more detailed
To cite: Cheung CY, Mok V, account for some of the variations and differences interrogation of the different retinal layers, and
Foster PJ, et al. J Neurol seen between studies discussed in this review. The even deeper structures beyond the retina, such as
Neurosurg Psychiatry 2018 National Institute on Aging and Alzheimer’s the choroid including the choroidal vasculature.
2021;92:983–994. Association Research Framework defines AD in These retinal imaging technologies, which include
Cheung CY, et al. J Neurol Neurosurg Psychiatry 2021;92:983–994. doi:10.1136/jnnp-2020-325347 983
 J Neurol Neurosurg Psychiatry: first published as 10.1136/jnnp-2020-325347 on 9 June 2021. Downloaded from http://jnnp.bmj.com/ on September 25, 2023 by guest. Protected by copyright.
General neurology
optical coherence tomography (OCT) (figure 2) and OCT angi- RETINAL IMAGING MEASURES AND AD
ography (OCT-­ A) (figure 3), have provided researchers with Retinal neuronal layer changes at optic disc and macula
further access to detailed retinal neuronal structure (eg, nerve AD is classically characterised by loss of neurons and synapses
fibre layer, ganglion cell layer and inner-­plexiform layer) and in the cerebral cortex and specific subcortical regions. Previous
microvasculature (eg, foveal avascular zone), respectively. histological studies have demonstrated that patients with AD also
Table 1 shows a glossary of retinal imaging technologies and
parameters used in studying AD. Details of these technologies
are summarised in online supplemental section 1. In compar- Box 1 Similarities between the retina and the brain
ison to brain imaging technology, retinal imaging has the advan-
tages of being non-­invasive, comparatively low cost, increasingly Embryological
widely available in non-­tertiary (eg, primary care and commu- ►► During embryonic development, the retina and optic nerve
nity) settings, and having different variables for quantifying originate from the diencephalon. The retina maintains its
the structures of retina. The development of next-­generation connection with the brain via the optic nerve after birth,
computational techniques such as artificial intelligence (AI) and being an integral component of the CNS.12
deep learning (DL) algorithms has further enhanced the poten-
Anatomical
tial of data-­rich retinal imaging as a promising tool and a source ►► Anatomically, the layered cytological and vascular structures
of biomarker for AD, particularly for individuals at preclinical and the presence of a blood barrier are similar in the
AD stage. However, most of the retinal imaging technologies retina and the brain. The retinal layers are composed of
are still specialised equipment and the interpretation of data several types of neurons including the retinal ganglion cell
requires expertise from ophthalmologists or visual scientists. (RGC) which comprises a cell body, dendrites and an axon,
In this review, we present an update of the current retinal similar to neurons in the CNS. In addition, the RGC axons,
imaging technologies, recent research findings and future which collectively form the optic nerve, are myelinated by
research applications in the study of AD. oligodendrocytes and are ensheathed by all three layers
of meninges. Furthermore, the retinal tissue is isolated
by the blood–retinal barrier which maintains a distinct
COMMON EYE DISEASES AND AD immunological and physiological environment, similar to the
Vision itself may be an important stimulus for the maintenance blood–brain barrier.S153 In terms of vascular structure, both
of cognitive health or reflect possible relationships between AD the cerebral and the retinal microvasculature components
with underlying eye diseases.S2-­S7 Population-­based and clinical are surrounded by a single layer of endothelial cells, which
studies have consistently shown that visual impairmentS8-­S14 are non-­fenestrated and possess similar intercellular tight
and a range of common eye diseasesS15-­S19 are associated with junctional complexes.S154 Both are also surrounded by the
dementia and impaired cognitive function. For example, in one perivascular end feet from astrocytes.S155 The choroid, the
study, older persons with visual impairment were twice as likely primary vascular supply for the outer retina, is sandwiched
to have cognitive dysfunction than those with good vision.S11 between the retina and the sclera. This has one of the highest
Online supplemental table 1 summarises research on the rela- blood flows per volume unit of any structure in the body.
tionship between common eye diseases, including AMD,S2-­S5
Physiological
DRS5 and open-­angle glaucoma,S5-­S7 with AD dementia. These
►► There are many physiological similarities between retina
epidemiological relationships suggest shared risk factors (eg,
and brain. First, a NVU is present in both retina and brain,
hypertension, smoking) and possibly pathogenic pathways (eg,
widely known as the ‘blood–retinal barrier’ and ‘blood–
neurodegeneration, amyloid-β deposits, chronic microvascular brain barrier’, respectively.13, S156 The NVU allows functional
insults) between these ocular diseases and AD dementia. However, coupling and interdependency of neurons, glia and the
the associations have not been consistently observed in the liter- vasculature, for example, regulating blood flow in response
ature, particularly in studies at a population level using large to neural activity or metabolic demands. S157,S158 Retinal
data linkage.S2,S3,S7 For example, a linkage study using English vascular autoregulation is achieved by retinal glial-­synaptic
National Health Service data (AD cohort n=251 703, reference interaction.S158 Second, similar to the CNS neurons, the
cohort n>2.5M), found no associations between AMD (relative RGCs produce an identical response to insults, including
risk 0.86 (0.67–1.08), compared with the reference cohort) and axonal degeneration, myelin destruction, scar formation
glaucoma (rate ratio 1.01 (0.96, 1.06), compared with the refer- and secondary degeneration.S159-­S161 In addition, the RGCs
ence cohort) with AD dementia.S2,S7 Lee et al however suggested have limited regenerative ability after injury.S162 Third, the
that the patterns of associations between eye diseases and AD retina is considered an immune-­privileged site and contain
dementia may be different when the ocular conditions are cate- similar collection of cell-­surface molecules, immunoregulatory
gorised as recent (diagnosed within 0–5 years) or established molecules and cytokines.S163,S164 Moreover, both cerebral
(>5 years) diagnoses.S5 They found only established AMD (HR and retinal microglia show phagocytic properties and
1.50 (1.25, 1.8)) and recent glaucoma (HR 1.46 (1.08, 1.91)) phagocytose injured neurons.S165,S166 The cerebral vasculature
are associated with AD dementia, while both established DR is devoid of autonomic innervation beyond the pia vessels.S167
(HR 1.50 (1.05, 2.15)) and recent DR (HR 1.67 (1.01, 2.74)) Similarly, there is no autonomic innervation to the retinal
are associated with AD dementia.S5 A recent meta-­analysis (21 vasculature beyond the level of the lamina cribrosa (except
studies, 7 876 499 study subjects) reported that patients with AD the choroidal circulation).S168,S169 Finally, both the inner
dementia are at greater risk for AMD (OR 2.22, I2=50%), and retinal and the cerebral circulation are under the fine control
patients with AMD are also at increased risk of AD dementia/ of the autoregulatory mechanism, which consists of myogenic
cognitive dementia (OR 2.42, I2=38%).15 Hence, it is likely that and metabolic components.S154
AD dementia and common eye diseases are linked via complex,
CNS, central nervous system; NVU, neurovascular unit.
interlinked, multimechanistic pathophysiology and pathways.
984 Cheung CY, et al. J Neurol Neurosurg Psychiatry 2021;92:983–994. doi:10.1136/jnnp-2020-325347
 J Neurol Neurosurg Psychiatry: first published as 10.1136/jnnp-2020-325347 on 9 June 2021. Downloaded from http://jnnp.bmj.com/ on September 25, 2023 by guest. Protected by copyright.
General neurology

Figure 1 (A) Retinal photograph showing the optic nerve head, macular area, nerve fibre layer, arterioles and venules. (B) Examples of a cognitively
normal subject; and (C) a subject with Alzheimer’s disease (AD) dementia (Global Clinical Dementia Rating Score of 1). Retinal photographs assessed
quantitatively with Singapore I Vessel Assessment software (SIVA, V.4.0; National University of Singapore, Singapore). Arterioles are in red and venules are in
blue. The measured area is standardised and defined as the region 0.5 to 2.0 disc diameters away from the disc margin. The subject with AD dementia was
also diagnosed with age-­related macular degeneration. The subject with AD dementia had a sparser retinal vascular network (arteriolar fractal dimension:
1.246 vs 1.316; venular fractal dimension: 1.253 vs 1.273) and more tortuous retinal vessels (arteriolar tortuosity (×104): 0.61 vs 0.48; venular tortuosity
(×104): 1.4 vs 0.50), compared with the cognitively normal subject.

have loss of retinal ganglion cells (RGCs) and their axons.S20,S21 Of the two SD-­OCT measurements, some studies suggested
A more recent postmortem study further suggested that the that macular GC-­IPL may be more sensitive than peripapillary
number of melanopsin RGCs, photoreceptors driving circadian RNFL for assessing neurodegeneration related to AD.20 21 For
photoentrainment, may be reduced in AD.14 These observations example, Cheung et al reported that macular GC-­IPL has a
have been the basis of clinical studies using OCT to determine better performance to discriminate AD from normal controls
the relationship between different retinal layers and AD. than that of peripapillary RNFL (area under receiver operating
The retinal nerve fibre layer (RNFL) surrounding the optic characteristic curves (AUROCs) 0.685 vs 0.601), adjusting for
disc (peripapillary RNFL) reflects RGC axons (figure 1). age and gender.20 Figure 2B,C shows an example of GC-­IPL
RNFL thickness can be measured by techniques including and RNFL measurement in a MCI subject with a positive cere-
time-­domain OCT (first generation of OCT)S22-­S32 as well as bral amyloid PET imaging. Finally, these OCT studies are also
confocal scanning laser ophthalmoscopyS33 and scanning laser consistent with research on other neurodegenerative diseases,
polarimetry.S34 These studies showed that in patients with AD such as patients with Parkinson disease and Lewy body
dementia, there is a significant reduction in RNFL thickness dementia, who also have thinner RNFL and thinner inner
compared with age-­matched cognitively normal controls. retinal layers.S40,S41 Furthermore, thinner RNFL, GC-­ IPL
Spectral-­domain OCT (SD-­OCT) and swept-­source OCT, and ganglion cell layer are also associated with reduced cere-
a newer generation of OCT, provides information on inner bral grey matter and white matter volumes and brain volume
retinal layers with greater resolution, such as the ganglion cell measured from MRI.22–26
layer and the inner-­plexiform layer (figure 2).S35 SD-­OCT is Several possibilities have been proposed to explain the above
now routinely used to not only to measure peripapillary RNFL findings on thinning of the retinal neuronal layer.18 27 First,
but also to assess RGC cell body and dendrites together by the cerebral pathology of AD may affect the neuronal connec-
segmenting and quantifying the thickness of the ganglion cell tions of the visual tract and cause retrograde degeneration of
inner plexiform layer (GC-­IPL, a combination of the ganglion the optic nerve and retinal layers, resulting in thinner retinal
cell layer and the inner plexiform layer) at the macula, neuronal and axonal layers including RNFL and GC-­IPL.27
since this region contains more than 50% of the total RGCs However, peripapillary RNFL could not discriminate controls
volume.S36 Numerous studies have investigated the association from AD patients with posterior cortical atrophy, a clinical
between SD-­OCT measures and AD (online supplemental table variant of AD with dominant involvement of parieto-­occipital
2). These studies first showed that SD-­OCT measurements (ie, visual) cortex, where one would explicitly expect this
of both peripapillary RNFL (inter-­ visit intraclass correla- retrograde degeneration to occur.S42 Alternatively, it is specu-
tion coefficient (ICC) 0.927 (range 0.845–0.961) and coef- lated that cerebral signs of AD pathology including amyloid-β
ficients of variation (CoV) 3.83% (range 2.71%–5.25%)), and plaques, fibrillar tau and signs of neuroinflammation occur
GC-­IPL (inter-­visit ICC was 0.968 (0.941–0.985), and CoV simultaneously both in the brain and the retina, underlining
was 1.91% (range 1.24%–2.32%)) are reproducible in patients a common pathogenesis linking retinal neuronal and axonal
with cognitive impairment.S37 Second, while a few studies layer changes and AD.S1 S43-­S47 The less common observation of
have reported thicker retinal layers in eyes of patients with thickened RNFL in ADS38,S39 may be explained by occurrence
AD dementia compared with controls,S38,S39 or no significant of reactive gliosis in inner retina, an inflammatory response,
thickness differences,16 17 the majority of studies, including a during early stages of AD, which may precede retinal neuronal
meta-­analysis, indicate that patients with AD dementia have layer thinning or mask underlying subtle retinal neuronal layer
thinner peripapillary RNFL (standardised mean difference thinning on OCT.S48
(SMD)=−0.67; I2=89%) and macular GC-­IPL (SMD=−0.46, There are fewer prospective studies on the longitudinal
I2=71%) compared with controls,18 consistent with a human relationship of SD-­ O CT measures and development of
postmortem study.19 cognitive function deterioration and AD dementia.27–31 The
Cheung CY, et al. J Neurol Neurosurg Psychiatry 2021;92:983–994. doi:10.1136/jnnp-2020-325347 985
 J Neurol Neurosurg Psychiatry: first published as 10.1136/jnnp-2020-325347 on 9 June 2021. Downloaded from http://jnnp.bmj.com/ on September 25, 2023 by guest. Protected by copyright.
General neurology

Figure 2 (A) Cross-­sectional view of retina captured by optical coherence tomography (OCT). (B) Assessment of macular ganglion cell-­inner plexiform
layer (GC-­IPL) and (C) peripapillary retinal nerve fibre layer (RNFL) of a subject who presented with mild cognitive impairment having a positive cerebral
amyloid positron emission tomography (PET) imaging, measured with Cirrus HD-­OCT (Carl Zeiss Meditec, Dublin, California, USA) (OD=right eye;
OS=left eye). No retinal disorders were observed in this subject. The GC-­IPL thickness map (B) and peripapillary RNFL thickness map (C) use a false colour
coding with warm colours represent high and cool colours represent low thickness values. The software further compares the measured thickness to the
device’s internal normative age-­matched database and generates a deviation map. Thinner GC-­IPL and RNFL thicknesses are observed in this example in
prodromal stage. (D) A cross-­sectional view of choroidal vasculature in an example of a cognitively normal subject; and (E) in a subject with Alzheimer’s
disease (AD) dementia (Global Clinical Dementia Rating Score of 1). The choroid is indicated by red arrows. Choroidal-­scleral interface can be clearly
identified. The subject with AD dementia had a thinner choroidal layer compared with the cognitively normal subject.

986 Cheung CY, et al. J Neurol Neurosurg Psychiatry 2021;92:983–994. doi:10.1136/jnnp-2020-325347

 J Neurol Neurosurg Psychiatry: first published as 10.1136/jnnp-2020-325347 on 9 June 2021. Downloaded from http://jnnp.bmj.com/ on September 25, 2023 by guest. Protected by copyright.
General neurology

Figure 3 Imaging of retinal capillary network using optical coherence tomography angiography (OCTA), which is not visualised using conventional retinal
camera.

Rotterdam Study (n=3289) showed that a thinner peripap- those in the top RNFL quintile.28 In regard to GC-­IPL, data
illary RNFL is associated with a higher risk of developing from the Rotterdam Study showed that thinner GC-­IPL is
dementia (per 1 µm decrease HR 1.02 (1.01, 1.04)), including only associated with prevalent dementia (per 1 µm decrease
AD dementia (per 1 µm decrease HR 1.02 (1.01, 1.04)), OR 1.03 (1.00, 1.09)), but not with incident dementia
independent of cardiovascular risk factors.27 In addition, (per 1 µm decrease HR 1.02 (0.99, 1.05)) or incident AD
longitudinal data from the UK Biobank (n=32 038) enrolling dementia (per 1 µm decrease HR 1.02 (0.99, 1.05)). 27 The
healthy community dwelling participants also showed that Rotterdam Study group speculated that there may be a time
thinner RNFL is a precursor of future decline in cognitive delay between optic nerve degeneration (reflected by RNFL
function. 28 Specifically, they found that those in the lowest thinning) and RGC loss (reflected by GC-­IPL thinning) as
two quintiles of baseline peripapillary RNFL distribution the damage to the optic nerve may cause swelling or gliosis
had twice the likelihood of a developing a decline in cogni- formation of the RGC axons (ie, RNFL). Therefore, the
tive function over a 3-­year follow-­u p interval compared with neurodegenerative process may manifest itself in the retina

Table 1 Glossary of retinal imaging technologies and parameters used in studying Alzheimer’s disease
Parameters measured from retinal
Retinal imaging technologies imaging Descriptions of the parameters
Spectral-­domain optical coherence Peripapillary retinal nerve fibre layer (RNFL) A thickness of RNFL layer in peripapillary region to assess retinal ganglion cell (RGC) axon.
tomography (SD-­OCT) thickness Traditionally, peripapillary RNFL thickness is calculated along a 3.4 mm circle around optic disc.
  Macular ganglion cell inner plexiform layer A combined thickness of ganglion cell layer and inner plexiform layer in macular region to assess
(GC-­IPL) thickness RGC cell bodies and dendrites.
  Macular ganglion cell complex (GCC) layer A combined thickness of RNFL, ganglion cell layer and inner plexiform layer in macular region to
assess RGC axon, cell bodies and dendrites.
  Macular thickness or volume A thickness or volume between inner limiting membrane and the retinal pigment epithelium at
macular region.
Quantitative retinal vasculature analysis Central retinal arteriole equivalent (CRAE) A summary index reflecting the average width of retinal arterioles.
with retinal fundus photography
  Central retinal venule equivalent (CRVE) A summary index reflecting the average width of retinal venules.
  Fractal dimension A measure of a fractal structure that exhibits the property of self-­similarity, characterising the
distribution of the branching retinal vasculature.
  Tortuosity A measure of the straightness/curliness of the retinal vessels.
  Length-­diameter ratio A measure of the ratio of the length between two branching points to trunk vessel width.
Optical coherence tomography Vessel density A measure of the area occupied by blood vessels (including capillaries) over total area within the
angiography (OCT-­A) interested region.
  Foveal avascular zone (FAZ) area A capillary-­free area in the central macula with highest cone photoreceptor density and oxygen
consumption.
Optical coherence tomography with Choroidal thickness or volume A thickness or volume between the outer border of retina pigment epithelium and sclera-­choroidal
enhanced depth imaging (OCT-­EDI) interface.
Dynamic vessel analyser Flicker-­induced vessel dilation An average percentage increase in the vessel diameter in response to the flickering-­light during
measurement cycles, relative to the baseline diameter size.
Retinal oximetry Retinal vessel oxygen saturation A measure of the oxygen saturation in retinal arterioles and venules to detect changes in oxygen
metabolism.
Ultra-­widefield retinal photography Presence of peripheral hard drusen Presence of small and distinct yellow deposits under retina in the periphery.
Fluorescence lifetime imaging Fluorescence lifetime A measure to describe the average time a single fluorophore remains in its electronically excited
ophthalmoscopy (FLIO) state after absorbing the energy of a photon and before returning to the ground state by emitting a
photon.

Cheung CY, et al. J Neurol Neurosurg Psychiatry 2021;92:983–994. doi:10.1136/jnnp-2020-325347 987

 J Neurol Neurosurg Psychiatry: first published as 10.1136/jnnp-2020-325347 on 9 June 2021. Downloaded from http://jnnp.bmj.com/ on September 25, 2023 by guest. Protected by copyright.
General neurology
initially as thinner peripapillary RNFL, after which thinning associated with AD dementia.40 These findings suggest that
of GC-­IPL follows.27 changes in the retinal vascular fractal dimension may also reflect
a departure from optimal integrity of the cerebral microcircula-
tion (eg, rarefaction) related to cognitive impairment.46 Studies
Retinal arteriolar and venular changes also showed that narrower retinal venular calibre is associated
There is also substantial evidence indicating a vascular disease with AD dementia, which may be related to an increased venous
component in AD pathophysiology. Clinical and epidemiolog- wall thickness due to collagen deposition in cerebral veins.39 40
ical studies show that vascular diseases and their risk factors Although some studies observed significant associations between
commonly accompanies AD.32 Vascular risk factors are associ- retinal vascular tortuosity and AD dementia, the relationship
ated with higher cerebral Αβ burden.33 Comorbidity of cere- remains equivocal. Cheung et al40 found that both increased
brovascular disease and amyloid-β is associated with cognitive retinal arteriolar (per-­SD increase OR 1.80 (1.40, 2.31)) and
decline and neurodegeneration.34 In particular, microvascular or venular tortuosity (per-­SD increase OR 1.94 (1.48, 2.53)) are
small vessel disease is now thought to be a major contributor to associated with AD dementia, while Williams et al41 found that
dementia and cognitive decline.S49-­S52 For example, an autopsy-­ decreased retinal arteriolar (per-­ SD increase OR 0.78 (0.63,
based neuropathological study showed that a large majority of 0.97)) tortuosity is associated with AD dementia. It is noted that
patients diagnosed with AD without clinical evidence of mixed increased retinal vascular tortuosity is associated with higher
(vascular) dementia had microvascular pathology including blood pressure and diabetes,S59,S62 and difference between
lacunes, cerebral microbleeds and multiple microinfarcts indica- studies may reflect differences in prevalence of hypertension and
tive of small vessel disease.35 diabetes (eg, participants in the study by Cheung et al had higher
The retinal circulation of arterioles and venules, measuring prevalence of both hypertension and diabetes than participants
100–300 µm in size, are the only optically accessible small in studies by Williams and Frost). On the other hand, if different
blood vessels in the human body. The retinal vasculature can be pathophysiological mechanisms occur at different stages of
imaged by either conventional retinal photography (figure 1A) AD,47 associations with retinal parameters would be expected
or dye-­based fluorescein angiography. However, because dye-­ to change correspondingly. Thus, clinical and demographic data
based fluorescein angiography is invasive, conventional retinal and stages of AD should be taken into account when interpreting
photography has been the most commonly used retinal imaging findings across studies.
technique to capture clinical vascular disease signs, such as those
typically seen in patients with diabetes or hypertension (eg, retinal
haemorrhage, cotton wool spots, microaneurysms, arteriovenous Retinal capillary changes
nicking, enhanced arteriolar light reflex). In addition to these In addition to the arterioles and venules, changes at the capil-
qualitative signs, computerised algorithms have been developed lary level (5–15 µm) may also be studied for their relationship
to measure quantitative changes in the retinal vasculature, for to AD.48 49 The retinal capillary network can now be imaged by
instance, the calibre of arterioles and venules.36–38 Furthermore, dye-­free OCT-­angiography (OCT-­A) which visualises capillary
geometric patterns of the retinal vasculature may also provide levels at different levels and sites of the retina: the superficial
information on microvascular health. Based on Murray’s prin- capillary plexus, the deep capillary plexus and the radial peri-
ciple of minimum work, the human circulatory system is a papillary capillary plexus. Images captured by OCT-­A (figure 3)
branching system that conforms to optimum design principal to have helped to identify and quantify capillary level abnormalities
minimise the energy required to maintain blood flow.S53 Algo- in primary retinal diseases such as DR and AMD.S63-­S66 Online
rithms estimating a number of retinal geometric parameters supplemental figure 1 shows image processing steps adopted to
such as fractal dimension, tortuosity and bifurcation have been quantify the capillary networks from OCT-­A images.S63 Recent
further reported.S54-­S61 These algorithms capture deviations from reports (online supplemental table 2) have identified changes
the normal optimal architecture of the retinal vascular network. in the retinal capillary networks using OCT-­A in AD dementia
Several algorithms can be used to analyse retinal photographs and preclinical AD. S67-­S75 For example, Bulut et al observed
taken by conventional retinal cameras. Figure 1B,C shows exam- decreased capillary network density (45.5% vs 48.7%, p=0.002)
ples of quantitative retinal vasculature analysis using a widely and an enlarged foveal avascular zone area (0.47 mm2 vs 0.33
used software, the Singapore I Vessel Assessment (SIVA), in a mm2, p=0.001) in 26 patients with AD dementia, compared
cognitively normal subject and a subject with AD, respectively. with 26 age-­matched and sex-­matched controls.S67 O’Bryhim
Other software and algorithms include IVAN (Integrative Vessel et al further demonstrated that enlarged foveal avascular zone
Analysis),36 VAMPIRE (Vascular Assessment and Measurement measured from OCT-­A is associated with preclinical AD (0.364
Platform for Images of the REtina)S57 and QUARTZ (QUantita- mm2 vs 0.275 mm2, p=0.002), as defined by the presence of
tive Analysis of Retinal vessel Topology and siZe).S58 amyloid-β biomarkers from PET or CSF, compared with those
Online supplemental table 2 summarises the clinical studies amyloid-β-negative control subjects (n=32 participants).S74
reporting relationship of quantitative retinal vasculature anal- In addition, they reported that foveal avascular zone area can
ysis from retinal photographs with AD. In general, these discriminate participants with biomarker positive and biomarker
studies showed a sparser retinal vascular network (indicated by negative with an AUROC of 0.801.S74 These observations are
decreased retinal vascular fractal dimension) is associated with in line with disturbances in the morphology and function of
AD dementia,39–42 poorer cognitive test score performances43 44 cerebral capillary networks observed as antecedents to neuro-
and MRI markers of cerebral small vessel disease.45 For example, degenerative changes associated with AD in animal models and
Frost et al found that decreased arteriolar fractal dimension postmortem studies. S76-­S78 However, the current literature is not
(1.201 vs 1.235, p=0.008) and venular fractal dimension entirely consistent. For example, Querques et al and den Haan
(1.171 vs 1.210, p<0.001) in AD dimension, compared with et al did not observe any differences in capillary network density
controls.39 Cheung et al found that decreased arteriolar fractal and foveal avascular zone between patients with AD dementia
dimension (per-­SD decrease OR 1.35 (1.08, 1.68)) and venular and controls.S71,S72 Van de Kreeke et al have reported that an
fractal dimension (per-­SD decrease OR 1.47 (1.17, 1.84)) are increased capillary network density in patients with preclinical
988 Cheung CY, et al. J Neurol Neurosurg Psychiatry 2021;92:983–994. doi:10.1136/jnnp-2020-325347
 J Neurol Neurosurg Psychiatry: first published as 10.1136/jnnp-2020-325347 on 9 June 2021. Downloaded from http://jnnp.bmj.com/ on September 25, 2023 by guest. Protected by copyright.
General neurology
AD, instead of a decreased one, compared with controls (inner correlations with dynamic flicker-­induced retinal arteriolar or
ring macula difference: 0.81%, p=0.002; outer ring macula venular dilation in an elderly cohort.51 It is noteworthy that
difference: 0.50%, p=0.024; and around optic nerve head patients with diabetes and DR have reduced flicker-­ induced
difference: 0.83%, p=0.015), which may be due to an inflam- vasodilatory response.S85,S86 Kotliar’s and Golzan’s studies did
matory state of the retina in the early stages of amyloid-β accu- not specifically exclude those with diabetes and with ocular
mulation.S75 It is noteworthy that similar to the literature on diseases, which may confound the results reported. Querques
retinal imaging technologies as mentioned above, the diagnosis et al recently conducted a similar study but excluded subjects
of AD dementia differs between different OCT-­A studies, and with diabetes and optic nerve or retinal diseases and reported
most of studies are currently limited by small sample size and the that retinal arteriolar dilation in response to flickering light is
inability to adequately account for potential confounding factors reduced in the AD dementia group, compared with controls
(eg, diabetes, axial length of the eyeball). (0.77% vs 3.53%, p=0.002) and the MCI group (0.77% vs
2.84%, p=0.045).S71 This study suggests that the neurovascular
coupling of retinal vessels in AD dementia is impaired, especially
Choroidal vasculature changes
in retinal arterioles, and the effect is independent of diabetes.
Deep to the retina, the choroid contributes blood supply to the
This observation suggests that decreased retinal and cerebral
outer retina. In addition to the retinal vasculature (arterioles,
blood flow (ie, hypoperfusion) might impair the endothelial
venules, capillary network), SD-­ OCT with enhanced depth
function and the production of nitric oxide, which is vital for
imaging, or swept-­source OCT, has now made it possible to
the vasodilatation process.52, S71
image the choroidal vasculature. Figure 2D,E shows a cross-­
sectional view of the choroidal vasculature imaged by SS-­OCT
in a cognitively normal subject and a subject with AD, respec- Retinal vessel oxygen saturation changes
tively. A few case–control studies have observed thinning of the The retina is one of the most metabolically active tissues in the
choroidal layer as assessed by SD-­OCT with enhanced depth human body. Retinal vessel oxygen saturation can now also be
imaging in AD dementia (online supplemental table 2).S67,S79,S80 measured non-­invasively, based on estimation of haemoglobin
For example, Gharbiya et al first reported reduced subfoveal oxygen saturation in retinal vessels by taking two simultaneous
choroidal thicknesses (200.9 µm vs 266.1 µm, p=0.001) in 21 retinal photographs with 570 and 600 nm light.S87 Studies
patients with mild to moderate AD dementia compared with have reported that retinal oxygen saturation in arterioles and
controls. It was postulated that choroidal thinning indicates an venules is significantly higher in both AD dementia and MCI
abnormal choroidal blood supply associated with vasoregression eyes compared with controls.S88, S89 Einarsdottir et al showed
or atrophic changes related to a series of pathologic events (eg, that retinal arterioles have 94.2%±5.4% oxygen saturation in
inflammatory cascade) triggered by amyloid-β deposition in the moderate AD compared with 90.5%±3.1% in healthy subjects
brain.S80,S81 In a recent prospective study (n=78), a larger reduc- (p=0.028). Retinal venules were 51.9%±6.0% saturated in
tion in choroidal thicknesses is observed in AD dementia over a moderate AD compared with 49.7%±7.0% in healthy subjects
12-­month follow-­up (changes in subfoveal choroidal thickness: (p=0.02).S88 Olafsdottir et al further showed that arteriolar and
−10.47 µm vs −2.0 µm), compared with controls.S81 This finding venular oxygen saturation was increased in patients with MCI
is consistent with a report on postmortem eyes from patients compared with healthy individuals (93.1%±3.7% vs 91.1±3.4%,
with AD dementia and animal models of AD.S82 Furthermore, a p=0.01; 59.6±6.1% vs 54.9±6.4%, p=0.001, respectively).S89
population-­based study with more than 3000 participants found These studies demonstrate a decreased metabolic activity in the
reduced subfoveal choroidal thickness is significantly associated retina and this exploratory finding may provide new insight into
with lower Mini Mental Status Examination score, in line with the pathophysiology of AD related to hypometabolism.
the findings in AD dementia eyes.S83
Peripheral retinal changes
Vasodilatory response changes Drusen deposits are the hallmark of AMD. Amyloid-β has been
The retinal NVU contains neurons, glial cells and vascular cells, found in drusen deposits in the retina.S90 In a recent pilot study,
similar to the NVU in the CNS.13 Flickering light stimulates Csincsik et al used ultra-­widefield scanning laser ophthalmos-
activity of the neural retina and leads to retinal vessel dilation copy and examined AMD-­like drusen deposits in a wider field of
as a result of the release of vasodilating factors, especially nitric the peripheral retina, in addition to the macular region.42 They
oxide, from endothelial and neural cells. This dynamic reaction found that patients with AD dementia are more likely to have
of retinal vessels to flickering light is influenced by neurovas- drusen deposits in the peripheral retina compared with controls
cular coupling and can therefore be used to assess the func- (25.4% vs 4.2%, p=0.04), especially in the superior nasal
tion of NVU in the retina.S84 The flicker-­induced vasodilatory quadrant.42 Online supplemental figure 2 shows an example of
response can now be measured non-­invasively using a dynamic presence of drusen deposits in the peripheral retina imaged by
vessel analyser (DVA). A few studies have explored to use DVA ultra-­widefield scanning laser ophthalmoscopy in a subject with
to investigate the flicker-­induced vasodilatory response in AD, AD dementia. In addition, Ukalovin et al also found that the
but the current findings are still preliminary and inconclusive. severity of cerebral amyloid angiopathy, a vascular feature asso-
Kotliar et al observed that overall flicker-­induced vasodilatory ciated with AD dementia and cognitive decline,S91 is correlated
response is delayed in AD dementia (arterial time to reach 30% with the number of drusen in the peripheral retina from post-
of maximum dilation: 7.0 s vs 5.0 s, p<0.001) compared with mortem AD eyes (r=0.78, p<0.05).S92 The preliminary findings
controls, suggesting delayed arterial reaction in AD.50 They also from these exploratory studies suggest that retinal abnormalities
demonstrated that this DVA parameter can discriminate AD related to AD are also present in peripheral retina, in addition
dementia with an AUROC of 0.853.50 However, Golzan et al to retinal changes measured from conventional retinal imaging
only found a positive correlation between neocortical amyloid-β areas focused centrally (ie, macular and optic disc regions).
standardised uptake value ratio measured by PET with the Further studies are required to determine whether the peripheral
amplitude of retinal arterial pulsations, but did not observe any retinal changes have similar or additional predictive value for
Cheung CY, et al. J Neurol Neurosurg Psychiatry 2021;92:983–994. doi:10.1136/jnnp-2020-325347 989
 J Neurol Neurosurg Psychiatry: first published as 10.1136/jnnp-2020-325347 on 9 June 2021. Downloaded from http://jnnp.bmj.com/ on September 25, 2023 by guest. Protected by copyright.
General neurology
AD dementia or cognitive decline, and to develop better under- plaques qualitatively resemble those found in the cerebrum of
standing of the significance of these associations, particularly subjects with AD dementia and were only detected in minimal
since peripheral changes can also occur in the intermediate to quantities in age-­matched non-­AD controls.14, S1 Moreover, it is
advanced stage of AMD.S93 possible that accumulation of retinal amyloid-β plaques occurs
earlier than in the brain and the amount increases with disease
Retinal fluorescence lifetime changes progression.S44 Similar to those in the brain, retinal amyloid-β
The retina has several kinds of endogenous fluorophores, deposition also showed to be associated with marked neurode-
including lipofuscin, advanced glycation end products, collagen, generation. La Morgia et al demonstrated that there is amyloid-β
melanin and elastin. Retinal fluorophores can be excited by deposition in and around degenerating melanopsin RGCs,
monochromatic light (eg, laser) and gain a higher level of energy suggesting amyloid-β is toxic to retinal neurons.14 Koronyo et
before returning to their ground state by emitting photons of al also reported that, when compared with matched controls,
longer wavelengths than the exciting light (ie, fluorescence reduced cell counts in the RGC layer, inner nuclear layer and
lifetime).S94 Each fluorophore possesses a characteristic fluo- outer nuclear layer are observed in patients with AD dementia
rescence lifetime. Fluorescence lifetime imaging ophthalmos- along with accumulation of amyloid-β.S1 In line with these find-
copy (FLIO) is a technique for measuring fluorescence lifetime, ings, the neurotoxicity of amyloid-β to retinal neurons has also
which is calculated by the average amount of time a fluoro- been shown in both cell-­line studiesS107,S108 and animal model
phore remains in the excited state following excitation, to studies.S43,S109,S110 These findings from animal studies are there-
allow detecting metabolic alterations in the retina for various fore consistent with the observations of clinical studies reporting
retinal disease such as AMD at an early stage.S94 Jentsch et al thinner RNFL and GC-­ IPL in patients with AD dementia
found that a longer retinal fluorescence lifetime measured by discussed above. Apart from retinal amyloid-β deposition, animal
FLIO is correlated with higher tau protein concentration in CSF studies also showed that there are tau aggregates in the retina
in patients with AD dementia.S95 In a pilot study, Sadda et al of transgenic mice with a tau mutation.S111 These mice display
also found that subjects with preclinical AD, compared with abnormal neurotrophic factor signalling, increased susceptibility
controls, have longer retinal fluorescence lifetime (593.9±93.3, to excitotoxic damage, early axonopathy and functional deficits
454.4±38.6 ps; 475.0±71.6, 394.1±28.2 ps in short spectral of RGCs.S111,S112
channel and long spectral channel of AD and control groups, However, it is noteworthy that the evidence of AD pathology
respectively, p=0.036 and 0.024), which is also correlated with in the retina is not consistent across the literature, and presence
GC-­IPL thickness, amyloid-β and tau protein in CSF.S96 The of amyloid-β or tau in the retina still remains controversial.S113
preliminary findings from these exploratory studies suggest that For example, Schön et al reported that the postmortem retinas
the fluorophore composition in the retina may be related to AD. of patients with AD dementia only contain hyperphosphory-
lated tau but not amyloid-β or fibrillar tau aggregates.S45 Several
Age-related effect on the retina studies did not find any AD-­related pathology in AD retinas
Age-­related changes are well reported in the retinal neuronal and could not replicate previous findings.S21,S114,S115 Haan et al
and vascular structures as well as in the choroid. RNFL and found that despite amyloid-β/amyloid precursor protein being
GC-­ IPL thicknesses decrease with increasing age in healthy present in postmortem AD retinas, there are no amyloid-β/
adults.S97-­S99 In a longitudinal analysis, the mean rate of change amyloid precursor protein-­related differences, but rather tau-­
of average RNFL is −0.52 µm/year in normal individuals.S100 related changes, between AD and control retinas.S116 It is noted
Similarly, narrower retinal vessel calibre, straighter retinal that adoption of different tissue processing methods and immu-
vessels, sparser retinal vasculature and decreased choroidal nostaining protocols, and comorbidity of ocular diseases and AD
thickness are correlated with increasing age in healthy adults as may possibly explain some discrepancies. On the other hand,
reported from population-­based studies.S101-­S105 In the Singapore although amyloid is proclaimed as a toxic substance, it triggers
Malay Eye Study-2, for each 1-­year increase in age, subfoveal the pathological process in the very initial state of the disease.
choroidal thickness decreased by 3.10 μm.S104 These age-­related For instance, hippocampal volume is characteristically reduced
effects on retinal imaging measures are considered as a phys- in AD with congruent tau pathology in the corresponding region,
iological process; however, the current measurements are not whereas it lacks amyloid pathology in hippocampal region.S117
standardised by age. Although most of the reported studies were Findings on the presence of amyloid-β or tau in the retina should
case–control study design matched with age or included age as be further replicated before we assume that AD pathology in the
a confounding factor in the statistical models, how to taking retina is present. Furthermore, studies on relationships among
ageing effect into account during image interpretation in clinical amyloid, tau and retinal changes in the pathological study of AD
practice should be further addressed. are warranted.

Evidence of AD pathology in the retina is equivocal

Amyloid-β protein plaques and neurofibrillary tangles comprised FUTURE RESEARCH
of hyperphosphorylated tau protein are pathological hallmarks Studies to date suggest a range of retinal imaging technologies
of AD. Several histopathological studies have identified retinal can be used to study different stages of AD. However, more
amyloid-β deposits in animal models of ADS82,S106 as well research is clearly needed. For example, studies on relationships
as human subjects with definite AD14, S1 and suspected early between retinal imaging measures and ‘standard’ biomarkers
AD.S44 The majority of amyloid-β deposits are in the GC-­IPL,S46 of AD pathophysiology, as well as direct comparisons between
and some show perivascular clustering.14, S1 A clinical study different retinal imaging measures to determine relative sensi-
by Koronyo et al also demonstrated an increased presence of tivity/specificity of each measure for AD clinical diagnosis or
amyloid-β in the retina of patients with AD dementia compared biomarker status are warranted. Furthermore, some neuronal
with controls, especially in the peripheral superior quadrant, and vascular changes in the retina are similar between AD and
often clustered along blood vessels.S1 These retinal amyloid-β ocular disease. Some of these researches are discussed below.
990 Cheung CY, et al. J Neurol Neurosurg Psychiatry 2021;92:983–994. doi:10.1136/jnnp-2020-325347
 J Neurol Neurosurg Psychiatry: first published as 10.1136/jnnp-2020-325347 on 9 June 2021. Downloaded from http://jnnp.bmj.com/ on September 25, 2023 by guest. Protected by copyright.
General neurology
Framework of development and validation of retinal imaging with AD dementia only. Moreover, studies on preclinical AD
as a source of biomarkers are largely lacking. Table 2 summarises the findings on retinal
Disease-­modifying treatments for AD are most likely to be neuronal changes, retinal capillary changes, retinal arteriolar and
successful if initiated early in the disease process, ideally before venular changes, and choroidal vasculature changes. In addition,
irreversible neurodegeneration and functional decline set in.53 54 some of the reported associations may not account for potential
Biomarkers may aid in risk profiling to identify those at greatest confounding factors (eg, ocular diseases and systemic conditions)
risk, detection of pathology at the earliest possible stage, and by and clinical stages of AD. The current retinal imaging features
providing endpoints for trials that identify benefit earlier in the (eg, RNFL loss) are also non-­specific to AD and not designed
natural history of the disease, thus accelerating the development to fully describe and characterise the spectrum of AD-­related
of new treatments. Therefore, discovering effective biomarkers disorders. For example, there is a lack of consensus on how best
of AD is a priority for development of new treatment. to differentiate between retinal changes seen in glaucoma and
A biomarker is defined as an objective substance, character- AD dementia. Typically, in patients with glaucoma, the reduction
istic or other parameter of a biological process that enables the in peripapillary RNFL occurs in inferior and superior sectorsS118
assessment of a disease risk or prognosis and provides guidance but the studies in AD dementia have also shown similar patterns
for diagnosis or monitoring of treatment.10 Different groups of RNFL reduction.18 Many prior studies have also not eval-
have proposed frameworks or roadmaps for biomarker devel- uated retinal imaging for a specific clinical purpose such as
opment and validation for AD.9 10 55 56 For example, Frisoni et screening, diagnosis, prognosis or monitoring of AD dementia.
al proposed a five-­phase framework to foster the clinical valida- Appropriate statistical analysis of biomarker validation should
tion of biomarkers for an early diagnosis of AD, adapted from be performed (eg, performing area under the receiver operating
the approach for cancer biomarkers. These include preclin- characteristic curve to evaluate discriminative performance).
ical exploratory studies, clinical assay development for AD Finally, the definition of AD in the reports did not consistently
pathology, retrospective studies using longitudinal data avail- include confirmation from current available biomarkers. Thus,
able in repositories, prospective diagnostic accuracy studies and there is inadequate evidence on the clinical usefulness of retinal
disease burden reduction studies.10 A consistent framework to imaging measures as a biomarker for AD.
assess the validity of biomarkers for different clinical purposes is We suggest the following framework of research (online
essentially important for implementation and clinical use in AD. supplemental table 3). First, studies should be designed and
Currently, retinal imaging technologies cannot yet be consid- conducted to validate biomarkers from retinal images with a
ered a source of biomarker of AD as many areas of research specific clinical purpose (eg, screening or risk stratification).
remain conflicting and evidence-­based guidelines on the use and Second, diagnosis of AD should be defined in a consistent
interpretation of retinal imaging are currently lacking. Most manner with the latest available criteria (eg, confirmation with
published studies have focused on demonstrating measure differ- current biomarkers). Third, standard statistical measures of accu-
ence compared with controls and associations (linear or logistic racy of biomarkers should be reported. For example, if a study
regression models) between a set of retinal imaging measures evaluates measures from retinal images as a potential screening

Table 2 The summary of current findings on retinal neuronal changes, retinal capillary changes, retinal arteriolar and venular changes, and
choroidal vasculature changes
Preclinical AD or MCI AD dementia
Retinal neuronal changes
Peripapillary RNFL thickness ►► Some studies showed reduced peripapillary RNFL, ►► Majority of studies and meta-­analysis showed reduced peripapillary RNFL,
macular GC-­IPL and GCC thicknesses, compared with macular GC-­IPL and macular thicknesses, compared with controls
controls ►► But a few studies showed no differences in the measures, compare with
►► But a few studies showed no differences in the controls
measures, compared with controls and AD dementia ►► A few studies also showed thickened peripapillary RNFL and macular GC-­IPL
►► A few studies also showed thickened peripapillary thicknesses, compared with controls
RNFL and macular GC-­IPL thicknesses in MCI, ►► In a population-­based study, reduced peripapillary RNFL associated with
compared with controls incident AD dementia
Retinal capillary changes
Vessel density and foveal avascular ►► Majority of studies showed decreased vessel density ►► Majority of studies showed decreased vessel density and enlarged foveal
zone and enlarged foveal avascular zone area, compared avascular zone area, compared with controls
with controls ►► But a few studies showed no differences in the measures.
►► But a few studies showed no differences in the
measures, and one study showed increased vessel
density in preclinical AD, compared with controls.
Retinal arteriolar and venular changes
Vessel calibre (CRAE and CRVE) ►► No data ►► Majority of studies showed narrower CRVE, compared with controls
Fractal dimension ►► No data ►► Majority of studies showed decreased arteriolar and venular fractal
dimensions, compared with controls
Tortuosity ►► No data ►► The association is equivocal. Both increased and decreased retinal vascular
tortuosity were observed
Choroidal vasculature changes
Choroidal thickness ►► No data ►► Reduced choroidal thicknesses, compared with controls
►► A larger reduction in choroidal thicknesses over a 12-­month follow-­up,
compared with controls
AD, Alzheimer's disease; CRAE, central retinal artery equivalent; CRVE, central retinal vein equivalent; GCC, ganglion cell complex; GC-­IPL, ganglion cell-­inner plexiform layer;
MCI, mild cognitive impairment; RNFL, retinal nerve fibre layer.

Cheung CY, et al. J Neurol Neurosurg Psychiatry 2021;92:983–994. doi:10.1136/jnnp-2020-325347 991

 J Neurol Neurosurg Psychiatry: first published as 10.1136/jnnp-2020-325347 on 9 June 2021. Downloaded from http://jnnp.bmj.com/ on September 25, 2023 by guest. Protected by copyright.
General neurology
tool, the sensitivity, specificity, false-­positive and false-­negative many layers of linear (convolutional) and non-­linear operations
rates of retinal imaging should be reported. The degree of abnor- trained on previously unfeasible amounts of data.
mality and the prevalence (ie, how frequent is the abnormality In retinal imaging, AI and DL technology have been developed
found in AD) should be described. Fourth, studies should be in several areas, the two most prominent ones being first, in the
designed carefully to take into account many age-­related ocular assessment of retinal photographs for detection and screening
conditions. The effects of ocular factors (eg, axial length of the of DR,S132-­S137 AMD,S138-­S140 glaucoma,S141, S142 and retinopathy
eyeball) on retinal imaging measures and the associations with of prematurity,S143,S144 and second, the segmentation and assess-
common eye diseases (eg, glaucoma and AMD) should be deter- ment of OCT images for diagnosis and screening of major retinal
mined and considered in the analysis and interpretation. Fifth, diseases.S145-­S148 These studies demonstrate the promise of DL
the reproducibility of the retinal imaging measures should be for discovering discriminative latent information associated
determined, particularly if retinal imaging is used for disease with AD as well as neurodegenerative disease and cerebrovas-
diagnosis and monitoring on multiple follow-­up visits. Finally, cular disease from retinal images.S149,S150 For example, using DL,
the incremental benefit and cost-­effectiveness of retinal imaging, target-­specific features are automatically learnt by DL algorithm
as well as acceptability to patients in different settings, should be in the feature extraction stage and numerous unconventional
evaluated. These areas should be reported in the framework of features that are neither noticed by human previously nor exam-
retinal imaging studies for AD. ined by appropriate clinical study will also be assessed. DL could
be used to recognise specific pattern of retinal changes secondary
to AD pathology (ie, ‘retinal fingerprint’ of AD) potentially.
Other new retinal imaging technologies related to AD How might AI-­ based retinal imaging be used potentially?
Several new retinal imaging technologies are also being explored Online supplemental figure 3 shows a proposed pathway of
which might provide additional value in this field. screening AD using retinal imaging. By providing a simple two-­
Retinal hyperspectral imaging obtains a series of hyperspectral tier risk stratification output, this algorithm could assist physi-
reflectance images, combing both spectral and spatial informa- cians to identify asymptomatic individuals who are more likely
tion, by scanning the retina with a continuous range of wave- to have AD in the community. The availability of retinal imaging
lengths of light.S119 More et al showed that amyloid-β exerts a in eye clinics for assessing ocular diseases allows opportunistic
characteristic influence on the reflectance of light as assessed screening for AD on a large scale. Introducing retinal imaging in
by retinal hyperspectral imaging and that the magnitude of this neurology clinics for subjects with memory issues would add a
effect varies in proportion with the amount of amyloid-β in the complementary risk profiling tool for assessing the risk of AD.
retina of AD mouse.S120,S121 This finding was further validated in Higher-­risk patients could then benefit from subsequent more
recent in vivo human clinical studies by showing optical density intensive and specific (but expensive) examinations (eg, PET
spectral profiles are different between AD and controls,S122 and imaging of CSP analysis for identification of underlying disease
cases who are amyloid-β positive and negative on PET can be pathologies). This would potentially benefit the treatment
discriminated from the reflectance of hyperspectral retinal workflow of AD if a disease-­modifying therapy is successfully
images with a machine learning–based model for the classifi- developed.
cation.S123,S124 Hyperspectral imaging technology is also being
investigated to identify brain cancerS125 as well as to estimate
Clinical trials and outcome monitoring
cerebral metabolism and haemodynamics from brain tissues.S126
While retinal imaging cannot fully replace current tests such as
Adaptive optics, improving the performance of optical systems
PET imaging or CSF analysis for detecting AD pathology (eg,
by reducing the effects of optical aberrations, can be employed
amyloid-β and tau accumulation), retinal imaging offers several
in scanning laser ophthalmoscopy to achieve very high resolu-
unique advantages over current biomarkers. First, retinal imaging
tion (~2 µm) in the human retina resulting in the direct visualisa-
offers lower cost methods to identify appropriate study cohorts
tion of nerve fibre bundles and other minute retinal features.S127
(ie, cognitively normal individuals with AD-­ related retinal
Zhang et al found that individuals with MCI have a significantly
characteristics) for recruitment into clinical trials of new treat-
higher number of hyperreflective granular membranes at the
ments for dementia (eg, anti-­amyloid therapy to delay cognitive
peripapillary area covered the RNFL as assessed by adaptive
decline). Measurements from retinal imaging (eg, neuronal and
optics scanning laser ophthalmoscopy.S48 The authors speculated
vascular changes) may also be used to assess optimal or subop-
that the hyperreflective granular membranes are due to inner
timal therapeutic response to medical intervention. For example,
retinal gliosis which supports a previously established asso-
the ENVIS-­ion study, which aims to determine the effective-
ciation between AD and glial cell activation in the brain and
ness of low-­dose aspirin in reducing the development of white
retina.S128,S129
matter lesion and silent brain infraction, is also validating retinal
Given that the retina is an easily accessible window and
vascular changes as potential treatment outcomes.57 In addition,
connected to the CNS, it is believed that further advance in
blood-­based biomarkers, a less invasive and potentially cheaper
retinal imaging as well as multimodal, composite biomarkers for
approach, are being explored for aiding AD detection at early
AD will be continuously developed.
stage. 58 59,S151-­S152 Combining both retinal imaging and blood-­
based biomarkers (ie, ‘multiple marker approach’) may increase
the accuracy to identify appropriate study cohorts for recruit-
Artificial intelligence
ment into the clinical trials, compared with using only a single
Another major area of future research in the analysis of retinal
marker.
images is AI. The current retinal imaging measures (eg, reduc-
tion in RNFL thickness) are not necessarily specific to AD and
not designed to fully describe and characterise the spectrum of CONCLUSIONS
AD-­related disorders. Recent developments of AI, particularly in There is an increasing body of research using current and
DL, have potential to transform imaging technology in health- emerging retinal imaging technology to study AD. Newer retinal
care.S130,S131 DL is based on deep neural networks, involving imaging technologies are increasingly available, are non-­invasive,
992 Cheung CY, et al. J Neurol Neurosurg Psychiatry 2021;92:983–994. doi:10.1136/jnnp-2020-325347
 J Neurol Neurosurg Psychiatry: first published as 10.1136/jnnp-2020-325347 on 9 June 2021. Downloaded from http://jnnp.bmj.com/ on September 25, 2023 by guest. Protected by copyright.
General neurology
and comparatively low cost and easy to use for clinical and 2 Jack CR, Knopman DS, Jagust WJ, et al. Tracking pathophysiological processes in
population studies. While current research shows promising Alzheimer’s disease: an updated hypothetical model of dynamic biomarkers. Lancet
Neurol 2013;12:207–16.
evidence that many retinal imaging measures show associations 3 Sperling R, Mormino E, Johnson K. The evolution of preclinical Alzheimer’s disease:
with AD, longitudinal studies are lacking and larger replication implications for prevention trials. Neuron 2014;84:608–22.
studies are necessary. A framework for retinal imaging develop- 4 Cummings J, Passmore P, McGuinness B, et al. Souvenaid in the management of mild
ment and validation in AD should be developed and followed by cognitive impairment: an expert consensus opinion. Alzheimers Res Ther 2019;11:73.
5 Dubois B, Padovani A, Scheltens P, et al. Timely diagnosis for Alzheimer’s disease: a
future studies to allow consistent comparison of findings. Newer
literature review on benefits and challenges. J Alzheimers Dis 2016;49:617–31.
computational technology, such as AI hold promise to use retinal 6 Albert MS, DeKosky ST, Dickson D, et al. The diagnosis of mild cognitive impairment
imaging as a ‘point of care; level test for screening, early risk due to Alzheimer’s disease: recommendations from the National Institute on Aging-­
assessment and stratification. Alzheimer’s association workgroups on diagnostic guidelines for Alzheimer’s disease.
Alzheimers Dement 2011;7:270–9.
7 Siemers ER, Sundell KL, Carlson C, et al. Phase 3 solanezumab trials: secondary
Author affiliations
1 outcomes in mild Alzheimer’s disease patients. Alzheimers Dement 2016;12:110–20.
Department of Ophthalmology and Visual Sciences, The Chinese University of Hong
8 Beach TG, Monsell SE, Phillips LE, et al. Accuracy of the clinical diagnosis of Alzheimer
Kong, Hong Kong, China
2 disease at National Institute on aging Alzheimer disease centers, 2005-2010. J
Gerald Choa Neuroscience Centre, Therese Pei Fong Chow Research Centre for
Neuropathol Exp Neurol 2012;71:266–73.
Prevention of Dementia, Lui Che Woo Institute of Innovative Medicine, Department
9 Jack CR, Bennett DA, Blennow K, et al. NIA-­AA research framework: toward a
of Medicine and Therapeutics, The Chinese University of Hong Kong, Hong Kong,
biological definition of Alzheimer’s disease. Alzheimers Dement 2018;14:535–62.
Hong Kong
3 10 Frisoni GB, Boccardi M, Barkhof F, et al. Strategic roadmap for an early diagnosis of
National Institute for Health Research Biomedical Research Centre, Moorfields
Alzheimer’s disease based on biomarkers. Lancet Neurol 2017;16:661–76.
Eye Hospital NHS Foundation Trust, NHS Foundation Trust, UCL Institute of
11 McKhann GM, Knopman DS, Chertkow H, et al. The diagnosis of dementia due
Ophthalmology, London, UK
4 to Alzheimer’s disease: recommendations from the National Institute on Aging-­
VAMPIRE project, Computing, School of Science and Engineering, University of
Alzheimer’s association workgroups on diagnostic guidelines for Alzheimer’s disease.
Dundee, Dundee, UK
5 Alzheimers Dement 2011;7:263–9.
Pharmacology, National University Singapore Yong Loo Lin School of Medicine,
12 London A, Benhar I, Schwartz M. The retina as a window to the brain-­from eye
Singapore
6 research to CNS disorders. Nat Rev Neurol 2013;9:44–53.
Memory Aging and Cognition Centre, National University Health System, Singapore
7 13 Antonetti DA, Klein R, Gardner TW. Diabetic retinopathy. N Engl J Med
Singapore Eye Research Institute, Singapore National Eye Centre, Singapore
8 2012;366:1227–39.
Ophthalmology and Visual Sciences Academic Clinical Programme, Duke-­NUS
14 La Morgia C, Ross-­Cisneros FN, Koronyo Y, et al. Melanopsin retinal ganglion cell loss
Medical School, Singapore
in Alzheimer disease. Ann Neurol 2016;79:90–109.
15 Rong SS, Lee BY, Kuk AK, et al. Comorbidity of dementia and age-­related macular
Acknowledgements We would like to acknowledge Dr Victor Chan from the degeneration calls for clinical awareness: a meta-­analysis. Br J Ophthalmol
Chinese University of Hong Kong, Hong Kong, China; Dr Saima Hilal from National 2019;103:bjophthalmol-2018-313277–83.
University of Singapore, Singapore; Dr Narayanaswamy Venketasubramanian from 16 Haan JD, van de Kreeke JA, Konijnenberg E. Retinal thickness as a potential biomarker
the Raffles Hospital, Singapore; and Dr M. Kamran Ikram from Erasmus Medical in patients with amyloid-­proven early- and late-­onset Alzheimer’s disease. Alzheimers
Centre, The Netherlands. Dement 2019;11.
Contributors CYC, CPLHC and TYW conceptualised and designed the article. 17 Sánchez D, Castilla-­Marti M, Rodríguez-­Gómez O, et al. Usefulness of peripapillary
CYC drafted the article. CYC, VM, PJF, ET, CPLHC, TYW and CPLHC were responsible nerve fiber layer thickness assessed by optical coherence tomography as a biomarker
for revising it critically for important intellectual content. All authors approved this for Alzheimer’s disease. Sci Rep 2018;8:16345.
version to be published. 18 Chan VTT, Sun Z, Tang S, et al. Spectral-­Domain OCT measurements in Alzheimer’s
disease: a systematic review and meta-­analysis. Ophthalmology 2019;126:497–510.
Funding CYC has received grant funding from the Health and Medical Research 19 Asanad S, Ross-­Cisneros FN, Nassisi M, et al. The retina in Alzheimer’s disease:
Fund, Hong Kong (Grant Number: 04153506) and Bright Focus Foundation histomorphometric analysis of an ophthalmologic biomarker. Invest Ophthalmol Vis Sci
(Reference Number: A2018093S). VM has received grant funding from SEEDS 2019;60:1491–500.
Foundation. PJF has received grant support for this work from The Richard Desmond 20 Cheung CY-­lui, Ong YT, Hilal S, et al. Retinal ganglion cell analysis using high-­
Charitable Foundation via Fight for Sight, London, UK (Grant code 1965), The definition optical coherence tomography in patients with mild cognitive impairment
International Glaucoma Association, Ashford, UK and The Alcon Research Institute, and Alzheimer’s disease. J Alzheimers Dis 2015;45:45–56.
Fort Worth, Texas, USA. CPLHC has received grant funding from the National Medical 21 Ito Y, Sasaki M, Takahashi H, et al. Quantitative assessment of the retina using OCT
Research Council Singapore (NMRC/CG/NUHS/2010 and NMRC/CG/013/2013). and associations with cognitive function. Ophthalmology 2020;127:107–18.
ET has received grant funding from the EPSRC (M005976/1) and the NIHR Global 22 Ong Y-­T, Hilal S, Cheung CY, et al. Retinal neurodegeneration on optical coherence
Health Research Unit “INSPIRED” (16/136/102). TYW has received grant funding tomography and cerebral atrophy. Neurosci Lett 2015;584:12–16.
from the National Medical Research Council Singapore (NMRC/STaR/016/2013 23 Casaletto KB, Ward ME, Baker NS, et al. Retinal thinning is uniquely associated with
and NMRC/OFLCG/001 c/2017) and Duke-­NUS Medical School (DUKE-­NUS/ medial temporal lobe atrophy in neurologically normal older adults. Neurobiol Aging
RSF/2014/0001). 2017;51:141–7.
Competing interests None declared. 24 Mutlu U, Bonnemaijer PWM, Ikram MA, et al. Retinal neurodegeneration and brain
MRI markers: the Rotterdam study. Neurobiol Aging 2017;60:183–91.
Patient consent for publication Not required.
25 Mutlu U, Ikram MK, Roshchupkin GV, et al. Thinner retinal layers are associated
Provenance and peer review Not commissioned; externally peer reviewed. with changes in the visual pathway: a population-­based study. Hum Brain Mapp
2018;39:4290–301.
Supplemental material This content has been supplied by the author(s). It
26 Tao R, Lu Z, Ding D. Perifovea retinal thickness as an ophthalmic biomarker for
has not been vetted by BMJ Publishing Group Limited (BMJ) and may not have
been peer-­reviewed. Any opinions or recommendations discussed are solely those mild cognitive impairment and early Alzheimer’s disease. Alzheimers Dement
of the author(s) and are not endorsed by BMJ. BMJ disclaims all liability and 2019;11:405–14.
responsibility arising from any reliance placed on the content. Where the content 27 Mutlu U, Colijn JM, Ikram MA, et al. Association of retinal neurodegeneration on
includes any translated material, BMJ does not warrant the accuracy and reliability optical coherence tomography with dementia: a population-­based study. JAMA
of the translations (including but not limited to local regulations, clinical guidelines, Neurol 2018;75:1256–63.
terminology, drug names and drug dosages), and is not responsible for any error 28 Ko F, Muthy ZA, Gallacher J, et al. Association of retinal nerve fiber layer thinning with
and/or omissions arising from translation and adaptation or otherwise. current and future cognitive decline: a study using optical coherence tomography.
JAMA Neurol 2018;75:1198–205.
ORCID iDs 29 Shi Z, Wu Y, Wang M, et al. Greater attenuation of retinal nerve fiber layer thickness in
Carol Y Cheung http://​orcid.​org/​0000-​0002-​9672-​1819 Alzheimer’s disease patients. J Alzheimers Dis 2014;40:277–83.
Christopher Chen http://​orcid.​org/0​ 000-​0002-​1047-​9225 30 Choi SH, Park SJ, Kim NR. Macular ganglion cell -Inner plexiform layer thickness is
associated with clinical progression in mild cognitive impairment and Alzheimers
disease. PLoS One 2016;11:e0162202.
REFERENCES 31 Trebbastoni A, D’Antonio F, Bruscolini A, et al. Retinal nerve fibre layer thickness
1 Alzheimer’s Association. 2016 Alzheimer’s disease facts and figures. Alzheimers changes in Alzheimer’s disease: results from a 12-­month prospective case series.
Dement 2016;12:459–509. Neurosci Lett 2016;629:165–70.

Cheung CY, et al. J Neurol Neurosurg Psychiatry 2021;92:983–994. doi:10.1136/jnnp-2020-325347 993

 J Neurol Neurosurg Psychiatry: first published as 10.1136/jnnp-2020-325347 on 9 June 2021. Downloaded from http://jnnp.bmj.com/ on September 25, 2023 by guest. Protected by copyright.
General neurology
32 Sweeney MD, Montagne A, Sagare AP, et al. Vascular dysfunction-­The disregarded 47 Jack CR, Holtzman DM. Biomarker modeling of Alzheimer’s disease. Neuron
partner of Alzheimer’s disease. Alzheimers Dement 2019;15:158–67. 2013;80:1347–58.
33 Gottesman RF, Schneider ALC, Zhou Y, et al. Association between midlife vascular risk 48 Østergaard L, Engedal TS, Moreton F, et al. Cerebral small vessel disease:
factors and estimated brain amyloid deposition. JAMA 2017;317:1443–50. capillary pathways to stroke and cognitive decline. J Cereb Blood Flow Metab
34 Yassi N, Hilal S, Xia Y, et al. Influence of comorbidity of cerebrovascular disease and 2016;36:302–25.
amyloid-β on Alzheimer’s disease. J Alzheimers Dis 2020;73:897–907. 49 Østergaard L, Aamand R, Gutiérrez-­Jiménez E, et al. The capillary dysfunction
35 Toledo JB, Arnold SE, Raible K, et al. Contribution of cerebrovascular disease in hypothesis of Alzheimer’s disease. Neurobiol Aging 2013;34:1018–31.
autopsy confirmed neurodegenerative disease cases in the National Alzheimer’s 50 Kotliar K, Hauser C, Ortner M, et al. Altered neurovascular coupling as measured by
coordinating centre. Brain 2013;136:2697–706. optical imaging: a biomarker for Alzheimer’s disease. Sci Rep 2017;7:12906.
36 Wong TY, Knudtson MD, Klein R, et al. Computer-­assisted measurement of retinal 51 Golzan SM, Goozee K, Georgevsky D, et al. Retinal vascular and structural changes
vessel diameters in the Beaver dam eye study: methodology, correlation between are associated with amyloid burden in the elderly: ophthalmic biomarkers of
eyes, and effect of refractive errors. Ophthalmology 2004;111:1183–90. preclinical Alzheimer’s disease. Alzheimers Res Ther 2017;9:13.
37 Cheung CY-­L, Hsu W, Lee ML, et al. A new method to measure peripheral retinal 52 de la Torre JC, Stefano GB. Evidence that Alzheimer’s disease is a microvascular
vascular caliber over an extended area. Microcirculation 2010;17:495–503. disorder: the role of constitutive nitric oxide. Brain Res Brain Res Rev
38 Hubbard LD, Brothers RJ, King WN, et al. Methods for evaluation of retinal 2000;34:119–36.
microvascular abnormalities associated with hypertension/sclerosis in the 53 Yassine HN. Targeting prodromal Alzheimer’s disease: too late for prevention? Lancet
Atherosclerosis risk in Communities study. Ophthalmology 1999;106:2269–80. Neurol 2017;16:946–7.
39 Frost S, Kanagasingam Y, Sohrabi H, et al. Retinal vascular biomarkers for early 54 Livingston G, Sommerlad A, Orgeta V, et al. Dementia prevention, intervention, and
detection and monitoring of Alzheimer’s disease. Transl Psychiatry 2013;3:e233. care. Lancet 2017;390:2673–734.
40 Cheung CY, Ong YT, Ikram MK, et al. Microvascular network alterations in the retina 55 Smith EE, Biessels GJ, De Guio F, et al. Harmonizing brain magnetic resonance
of patients with Alzheimer’s disease. Alzheimers Dement 2014;10:135–42. imaging methods for vascular contributions to neurodegeneration. Alzheimers
41 Williams MA, McGowan AJ, Cardwell CR, et al. Retinal microvascular network Dement 2019;11:191–204.
attenuation in Alzheimer’s disease. Alzheimers Dement 2015;1:229–35. 56 Humpel C. Identifying and validating biomarkers for Alzheimer’s disease. Trends
42 Csincsik L, MacGillivray TJ, Flynn E, et al. Peripheral retinal imaging biomarkers for Biotechnol 2011;29:26–32.
Alzheimer’s disease: a pilot study. Ophthalmic Res 2018;59:182–92. 57 Reid CM, Storey E, Wong TY, et al. Aspirin for the prevention of cognitive decline in
43 Cheung CY, Ong S, Ikram MK, et al. Retinal vascular fractal dimension is associated the elderly: rationale and design of a neuro-­vascular imaging study (ENVIS-­ion). BMC
with cognitive dysfunction. J Stroke Cerebrovasc Dis 2014;23:43–50. Neurol 2012;12:3.
44 Ong YT, Hilal S, Cheung CY, et al. Retinal vascular fractals and cognitive impairment. 58 Palmqvist S, Janelidze S, Quiroz YT, et al. Discriminative accuracy of plasma
Dement Geriatr Cogn Dis Extra 2014;4:305–13. Phospho-­tau217 for Alzheimer disease vs other neurodegenerative disorders. JAMA
45 Hilal S, Ong YT, Cheung CY, et al. Microvascular network alterations in retina of 2020;324:772–81.
subjects with cerebral small vessel disease. Neurosci Lett 2014;577:95–100. 59 Karikari TK, Pascoal TA, Ashton NJ, et al. Blood phosphorylated tau 181 as a
46 Cheung CY, Chen C, Wong TY. Ocular fundus photography as a tool to study stroke biomarker for Alzheimer’s disease: a diagnostic performance and prediction modelling
and dementia. Semin Neurol 2015;35:481–90. study using data from four prospective cohorts. Lancet Neurol 2020;19:422–33.

994 Cheung CY, et al. J Neurol Neurosurg Psychiatry 2021;92:983–994. doi:10.1136/jnnp-2020-325347

 medRxiv preprint doi: https://doi.org/10.1101/2021.09.27.21264168; this version posted September 29, 2021. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

It is made available under a CC-BY-NC-ND 4.0 International license .

A Novel Retinal Vascular Feature and Machine Learning-based Brain White Matter Lesion Prediction
Model

Alauddin Bhuiyan1, Pallab Kanti Roy2, Tasin Bhuiyan3, Elsdon Storey4, Walter P Abhayaratna5, Mandip
Dhamoon6, R. Theodore Smith1, and Kotagiri Ramamohanarao2

1 Department of Ophthalmology, Icahn School of Medicine at Mount Sinai, New York, USA
2 Department of Computing and Information Systems, The University of Melbourne, VIC, Australia.
3 College of Arts and Science, New York University, USA
4 Department of Medicine (Neuroscience), Monash University, Melbourne, Australia. E-mail:
[email protected]
5 College of Medicine, Biology and Environment, Australian National University, Canberra, Australia. E-
mail: [email protected]
6 Department of Neurology, Icahn School of Medicine at Mount Sinai, New York, USA

*Corresponding author: R. Theodore Smith MD PhD

Professor of Ophthalmology and Neuroscience
Department of Ophthalmology
Icahn School of Medicine at Mount Sinai
[email protected]

Page 1 of 14
NOTE: This preprint reports new research that has not been certified by peer review and should not be used to guide clinical practice.
 medRxiv preprint doi: https://doi.org/10.1101/2021.09.27.21264168; this version posted September 29, 2021. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

It is made available under a CC-BY-NC-ND 4.0 International license .

Abstract

White matter lesion (WML) is one of the common cerebral abnormalities, it indicates changes in the white
matter of human brain and have shown significant association with stroke, dementia and deaths. Magnetic
resonance imaging (MRI) of the brain is frequently used to diagnose white matter lesion (WML) volume.
Regular screening can detect WML in early stage and save from severe consequences. Current option of
MRI based diagnosis is impractical for regular screening because of its high expense and unavailability.
Thus, earlier screening and prediction of the WML volume/load specially in the rural and remote areas
becomes extremely difficult. Research has shown that changes in the retinal micro vascular system reflect
changes in the cerebral micro vascular system. Using this information, we have proposed a retinal image
based WML volume and severity prediction model which is very convenient and easy to operate.

Our proposed model can help the physicians to detect the patients who need immediate and further MRI
based detail diagnosis of WML. Our model uses quantified measurement of retinal micro-vascular signs
(such as arteriovenular nicking (AVN), Opacity (OP) and focal arteriolar narrowing (FAN)) as input and
estimate the WML volume/load and classify its severity. We evaluate our proposed model on a dataset of
111 patients taken from the ENVISion study which have retinal and MRI images for each patient. Our model
shows high accuracy in estimating the WML volume, mean square error (MSE) between our predicted WML
load and manually annotated WML load is 0.15. The proposed model achieves an F1 score of 0.92 in
classifying the patients having mild and severe WML load.

The preliminary results of our study indicate that quantified measurement of retinal micro-vascular features
(AVN, OP and FAN) can more accurately identify the patients who have high risk of cardio-vascular
diseases and dementia.

Introduction

The retina is the inner surface of the eye which covers the largest part of the fundus. The main structures
of a retina include the optic disc (i.e., optic nerve), the blood vessels (arteries and veins), and the macula.
Retinal images are captured using a fundus camera, Figure 1 shows a color fundus image of the human
retina. Retinal pathologies includes different types of retinopathy signs, arteriovenous nicking (AVN), focal
arteriolar narrowing (FAN), generalized arteriolar narrowing, venular dilation, etc.; detail descriptions of
these retinal pathologies are present in [1] and [2]. Retina is an extension of the inter-brain [3] and retinal
micro-vasculature share anatomic, embryologic and physiologic characteristics with the cerebral or brain
microvasculature [4]. Association between retinal microvascular changes and different types of cerebral
abnormalities (such as stroke and dementia) have been reviewed in [5-7]. Table. 1 shows the association
between retinal and brain imaging abnormalities or pathologies.

Three largest population-based studies: AGES [8], ARIC [9] and CHS [10] have reported the association of
retinopathy signs and retinal microvascular abnormalities with brain imaging abnormalities. Among them,
only ARIC found significant association between retinopathy signs and brain imaging abnormalities which
includes infarcts, WMLs and atrophy. Associations of arteriolar narrowing and venular dilation with the brain
imaging abnormalities were generally weaker as reported by two population-based studies as shown in
Table 1. In contrast to that, all three population-based studies [8, 10] have reported that AVN and FAN have
statistically significant association with the brain abnormalities. Therefore, among different retinal
pathologies, we mainly focus on the AVN and FAN for designing a prediction model of WML load. Among
different brain abnormalities, we select WML because WML has a strong association with stroke, dementia
and deaths as shown in Table 2.

Table 1. Summary of the associations between retinal imaging pathologies and brain imaging pathologies
[4] based on cross-sectional population based studies.

Retinal imaging Study Name Brain Imaging pathologies

Page 2 of 14
 medRxiv preprint doi: https://doi.org/10.1101/2021.09.27.21264168; this version posted September 29, 2021. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

It is made available under a CC-BY-NC-ND 4.0 International license .

pathologies
Odds Ratio (95% CI)
Periventricular
Any Lacunar Subcortical Subcortical Cortical
Infarct Infarct WML WML atrophy Atrophy

AGES [1, 2] 1.06 1.10 1.11 1.12 .. ..

Any retinopathy (0.89;1.26)(0.86;1.40) (0.88;1.41) (0.91;1.37)
ARIC [1, 3] 2.95* .. 2.50* .. 1.50* 1.90*
(1.30;6.59) (1.50;4.00) (1.00;2.30) (1.20;3.00)
CHS [1, 4] 0.94 .. .. .. .. ..
(0.63;1.41)

AGES [1, 2] 1.07 1.33* 1.64* 1.54* .. ..

Arteriovenous
nicking (0.93;1.22)(1.09;1.62) (1.36;1.97) (1.31;1.81)
ARIC [1, 3] 1.90* .. 2.10* 1.20 1.20 ..
(1.25;2.88) (1.40;3.20) (0.80;1.60) (0.90;1.70)
CHS [1, 4] 1.84* .. .. .. .. ..
(1.23;2.76)

AGES [1, 2] 1.08 1.19 1.40* 1.40* .. ..

Focal narrowing (0.90;1.29)(0.92;1.54) (1.09;1.79) (1.12;1.74)
ARIC [1, 3] 1.89* .. 2.10* 1.20 1.10 ..
(1.22;2.92) (1.40;3.10) (0.90;1.70) (0.80;1.60)
CHS [1, 4] 1.20 .. .. .. .. ..
(0.81;1.78)

ARIC [1, 2] 1.74 .. 1.20 1.10 1.00 ..

Arteriolar narrowing (0.95;3.21) (0.80;1.90) (0.80;1.50) (0.70;1.40)
CHS [1, 4] 1.15 .. .. .. .. ..
(0.99;1.26)
Rotterdam [1,
5] 1.14 .. .. .. .. ..
(0.91;1.44)

CHS [1, 4] 1.05 .. .. .. .. ..

Venular dilation
(0.93;1.19)
Rotterdam [1,
5] 1.07 .. .. .. .. ..
(0.85;1.35)

Note: Here * means p <= 0.05.

Table 2. Association of white matter lesion (WML) with the stroke, dementia and deaths based on the
outcome of 46 longitudinal studies [11].
Disease Hazard Ratio 95% Confidence Interval
Stroke 3.3 2.6 to 4.4
Dementia 1.9 1.3 to 2.8

Page 3 of 14
 medRxiv preprint doi: https://doi.org/10.1101/2021.09.27.21264168; this version posted September 29, 2021. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

It is made available under a CC-BY-NC-ND 4.0 International license .

Death 2.0 1.6 to 2.7

White matter lesions (WMLs) are commonly found on the MRI scans of the brain. The prevalence of WMLs
ranges from 11-21% in adults age and around 64% to 94% at the older age of 80 or more [11]. WMLs
appear hyper-intensed in the T2 weighted and fluid attenuated inversion recovery (FLAIR) MRI scans as
shown in Figure 2. The underlying pathology of these lesions mostly reflects demyelination and axonal loss
as a consequence of chronic ischemia caused by cerebral small vessel disease (microangiopathy) [12].
Early diagnosis of WMLs can help the physicians to consider the clinical events associated with it for the
preventive treatment. Regular MRI screening of a patient can be done to detect WML in early stage, but it
is not practical because of its high expense and inadequacy in the rural areas. Retinal color fundus imaging
can be an efficient technology to resolve this problem as it can visualize retinal vascular system non-
invasively. Research studies have shown that, retinal microvascular features: AVN and FAN has significant
association with WML as mentioned in Table 1. The previous studies establish the association between
WML, and retinal microvascular features based on their qualitative measurement. Since manual qualitative
assessments are highly subjective and less reproducible. Therefore, in this study, we have analyzed the
association between WML, and two prominent retinal vascular features (AVN and FAN) based on their
quantified measurements and proposed a retinal vascular features based WML prediction model.

In our proposed model, we use retinal micro-vascular sign to predict the severity of WML load. Focal
arteriolar narrowing (FAN) and AVN are two prominent micro-vascular signs in the retina. Arteriovenous
nicking (AVN) (as shown in Figure 3) can be defined as the narrowing of venular caliber by a stiff artery at
their crossing point and FAN (as shown in Figure 4) is the sudden narrowing of arteriolar width in the
retina. FAN and AVN have shown significant association with WMLs load in the brain as mentioned in
Table.2. Retinal arteriolar opacity or copper wiring showed association with hypertension which is a risk
factor for heart attack and stroke [4, 9, 13]. Therefore, severity of AVN, FAN and Opacification (OP) can
be an important parameter for predicting the severity of WML load. In previous studies, manual qualitative
grading of AVN and FAN is used to compute their association with WML load. Manual grading can be
affected by grader’s fatigue and lack of concentration. As a result, manual grading produces high inter
and intra-grader variability and it is highly dependent on the expertise of the grader. Therefore, we
hypothesize that computer based quantitative gradings of AVN, FAN and OP can represent their
association with WML load more accurately and precisely compared to their manual grading.

In this paper, our main contribution is the proposal of a model to estimate and classify the severity of
WML load based on the computer aided quantified measurement of AVN, FAN and OP. To best of our
knowledge, no previous work has been done on the prediction of WML volume based on the quantified
AVN, FAN and OP measurement. In addition, we compare quantitative and qualitative AVN and FAN
measurement-based state of the art machine learning prediction algorithms to find out the benefit of the
computer aided quantification of AVN and FAN. We evaluate our proposed model on a dataset of 111
patients of ENVISion [14] study. For each of these patients, we have multi-channel MRI images (T1, T2
and FLAIR) and corresponding disc and macula centered retinal images.

The rest of the paper is organized as follows. Section 1 details the methodology used in this study. The
experimental results and discussions are presented in section 2, and section 3 concludes the paper.

1. Method

The block diagram of the proposed WML volume prediction model is presented in Figure 5. In the
proposed WML volume prediction model, quantified retinal micro-vascular signs (AVN,OP and FAN) are
extracted as features from a color retinal image. The extracted features are then used to train machine
learning models such as Random Forest, artifical Neural Network, Linear Regression and Support Vector
regressor to estimate the WML volume. The following subsections contains the description of our dataset,

Page 4 of 14
 medRxiv preprint doi: https://doi.org/10.1101/2021.09.27.21264168; this version posted September 29, 2021. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

It is made available under a CC-BY-NC-ND 4.0 International license .

feature extraction step and training machine learning models: Random Forest, artifical Neural Network, ,
Linear Regression and support vector for WML volume prediction.

1.1 Dataset

Brain MRI and retinal vascular imaging data used in this work are obtained from the ENVISion [14] study.
All participants have given written consent separately for the ENVIS-ion trial. Protocols of ENVIS-ion
study have been approved by the Monash University health and research ethics committee. Our study is
approved as a low-risk project under the ENVIS-ion study by the Alfred hospital ethics committee,
Melbourne, Victoria. Therefore, we have access to only imaging data of this study.

A 111 standardized Brain MRI sequences are acquired to allow morphologic, microstructure and
functional assessment. First, a volumetric 3D T1-weighted sequence is used to obtain high resolution
anatomical information. The size of each subject’s voxel set is 166x256x256 and the resolution is 1:2x
0:94x 0:94 mm3. This is followed by the FLAIR and T2 sequences used to assess inflammation and
signal alteration in brain tissues. For them size of each subject’s voxel set is 256x256x36 and resolution
is 0:94 0:94 4 mm3.

On the other-hand, two color retinal photographs, centered on the optic disc and macula, respectively, is
taken from both left and right eye of the participant. The photographs are taken after 5 minutes of dark
adaption, without pharmacological pupil dilation, using a Canon NMR 45 digital fundus camera. The
resolution of retinal images is 4752 pixelsx3168 pixels.

The manual segmentation of WML is time consuming therefore, we use an automated method proposed
in [16] to segment WML. In [15] a wide range of features are used which contain multi-channel local
intensity, textural and spatial information to train a Random Forest (RF) classifier to compute the posterior
probability of a voxel as a lesion. Then the RF computed lesion probability of each voxel is incorporated
into a Markov Random Field (MRF) to obtain the final segmentation of white matter lesion. After the
automatic segmentation of white matter lesions, the false positive and missing lesions are manually
corrected by an expert grader using ITK-SNAP [16]. After the manual correction of WML, their volume is
computed by using the voxel’s resolution and size.

Table 3. Grading Levels for arteriovenous nicking (AVN) and focal arteriolar narrowing (FAN).

Severity Label AVN FAN

0 Absent Absent

1 Questionable Questionable

2 Moderate -

3 Severe Severe

1.2 Feature extraction

As mentioned in section 1.1, our dataset contains retinal images and MRI images of 111 patients. For
each of these patients, AVN is graded into four levels (0=absent, 1=questionable, 2=moderate and
3=severe) and FAN is graded into 3 levels (0=absent, 1=questionable and 3=severe) by two expert
graders of the Centre for Eye Research Australia (CERA). The manually graded AVN and FAN segments
are then quantified by two semi-automatic methods proposed in [17] and [18].

Page 5 of 14
 medRxiv preprint doi: https://doi.org/10.1101/2021.09.27.21264168; this version posted September 29, 2021. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

It is made available under a CC-BY-NC-ND 4.0 International license .

We have created the feature set using the quantified measurement of AVN and FAN.

The details of the feature extraction step are given in the following subsections.

1.2.1 Quantified score of arteriovenous nicking (AVN)

We compute the amount of narrowing in the vein near the artery vein (AV) crossover point to quantify the
severity of AVN using the method proposed in [18]. The procedures of retinal vessel segmentation, AV
crossover point detection, region of interest selection, artery vein classification, separation of venular
segment, vein edge and centerline refinement and width computation are explained in [18] (as shown in
Figure 6). We quantify the AVN by using the characteristics of the vein widths as shown in Figure. 7.
Suppose that the vein segment 1 in Figure 7 is made up of n vessel cross-sectional (each white line)
widths, w = {wi i Î1,..., n} (the vessel widths wi are indexed and ordered so that w1 is the width closest
to the crossover point, while wn is the furthest). W is approximated as the median value of its
measurements, or W = median(w). Suppose that N is the number of successive cross-sections, starting
from the first measurement w1, with the measurements lower than normal vessel width W, WC is
computed as the minimum value of the first N vessel widths, or WC = min{wi i Î1,..., N }. Then the AVN
score for vein segment 1 is computed by the following equation:

æ W ö
avns = ç1 - C ÷ ´ 100%
è W ø (1)

Since each crossover point has two venular segment, therefore, we will get two avns score such as avns1
for vein segment 1 and avns2 for vein segment 2. The max (avns1, avns2) is chosen as the final avn score
for the respective AV crossover point.

We select maximum of the AVN scores because it has shown highest correlation with the corresponding
manual grading of AVN severity compared to the summation and average of the AVN scores of two vein
segments correspond to an AV cross-over point.

1.2.2 Quantified score of focal arteriolar narrowing (FAN)

We adopted a semi-automatic approach proposed in [18] to quantify the FAN. The region of interest is
selected by clicking four points around the vessel as shown in Figure 8. Then we use the method
proposed in [17] to crop the selected region in such a way that the vessel will be always located at the
center of the ROI. Following that the cropped image is affinely transformed as mentioned in [18] to
horizontally align the vessel position with the center of the image as shown in Fig. 8(b)

This spatial normalization of the vessel makes the mapping of edges easier. We apply an edge
preserving smoothing filter proposed in [19] to suppress the noise (Fig. 8(c)). After smoothing the vessel
region, a Canny edge detector [20] is applied in the spatially normalized ROI to detect the edges (Fig.
8(d)). Vessel edges are mapped from the detected edges (Fig. 8(e)) by using Dijkstra’s shortest path
algorithm [21]. After identifying vessel edges, its width (as shown in Fig. 8(g)) is measured through
mapping edge pixel pairs from both edges as proposed in [17].

For each arteriolar segment, we compute fans from the vessel widths of that segment to measure the
amount of narrowing. Suppose that the segment is made up of n arteriolar width measurements as shown
in Figure 9, w = {wi i Î1,..., n}. The average of the width measurements is computed as follows:

Page 6 of 14
 medRxiv preprint doi: https://doi.org/10.1101/2021.09.27.21264168; this version posted September 29, 2021. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

It is made available under a CC-BY-NC-ND 4.0 International license .

"
𝑑! = ! ∑!" 𝑤# (2)

æ ö
Let j is the index of the minimum width measurement ç j = arg min w j ÷then the j th width measurement
ç ÷
è 1£ j £ n ø
is smoothed by using a boxfilter of size n as follows:
n
1
ds = å w j +k
2n + 1 k = - n
(3)

Finally, the narrowing in percentage is computed as follows:

ds
fans = (1 - ) ´ 100% (4)
dn
For each subject, we compute the number of AVN and FAN cases and statistical information (mean,
median, standard deviation, summation) of the corresponding quantified AVN and FAN scores (avns and
fans ) as shown in Table 4. Illustration of the feature set construction for a patient is shown in Figure. 10.
Here, patient-1 has 7 AVN and 5 FAN affected vessels.

Table 4. Description of the feature set.

Name Description
#AVN #AVN is the number of AVN cases
#FAN #FAN is the number of FAN cases
#OP #OP is the number of OP cases
maxAVN/ maxFAN / maxOP Maximum of the quantified AVN / FAN / OP scores
meanAVN/ meanFAN / Mean of the quantified AVN / FAN / OP scores
meanOP
medianAVN/ medianFAN / Median of the quantified AVN / FAN / OP scores
medianOP
sumAVN/ sumFAN / sumOP Summation of the quantified AVN / FAN / OP
scores
stdAVN/ stdFAN / stdOP Standard deviation of the quantified AVN / FAN /
OP scores

1.3 Prediction and classification of white matter lesion volume

We have compared the state of the art machine learning algorithms to obtain the optimal WML prediction
model using the feature set proposed in Table 4. A brief description of these models are given in the
following sub sections.

1.3.1 Support Vector machine (SVM)

SVM [25] is a learning system that uses higher dimensional feature space. A support vector machine
constructs a hyperplane in a higher dimensional space which can be used for classification or regression.
Intuitively, a good separation is achieved by the hyperplane that has the largest distance to the nearest

Page 7 of 14
 medRxiv preprint doi: https://doi.org/10.1101/2021.09.27.21264168; this version posted September 29, 2021. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

It is made available under a CC-BY-NC-ND 4.0 International license .

training data point of any class (so-called functional margin), since in general the larger the margin the
lower the generalization error of the classifier. In our implementation, we used the kernel radial basis
function (rbf), with a C value of 1.0, and a degree 3.

1.3.2 Random Forest (RF)

A Random Forest [22] consists of a various number of individual decision trees. It is an ensemble
estimator which gives lower variance. We used the scikit-learn implementation of random forest machine
learning model. The most important parameter in this case is the number of estimators i.e., the number of
individual decision trees. We experimented with different number of decision trees ranging from 10 to
600. We got the best accuracy when we used 400 trees. All other parameters such as criterion and
max_depth were set as default value of scikit-learn implementation. We used 400 trees, allowed nodes to
expand until all leaves are pure, minimum samples split of 2, and minimum samples required to be at leaf
node to be 1.

1.3.3 Artifical Neural Network (ANN)

Artifical Neural Network (ANN) [22] is a supervised machine learning algorithm that has the capability to
learn non-linear models. It can have one or more non-linear layers in between the input layer and the
output layer. These non-linear layers are called hidden layers. Our ANN has 3 hidden layers each
containing of 100 hidden units called neurons. We used the scikit-learn implementation of artifical neural
network. We used ‘lbfgs’ solver for weight optimization as our dataset was small. Because in case of
small datasets, ‘lbfgs’ converges faster and performs better [22]. Rectified Linear Unit (ReLU) was used
as activation function. All other parameters were set as default.

1.3.4 Linear Regression (LR)

Linear Regression is an approach to model the relationship between two variables, dependent and
independent variables, by fitting a linear equation to the data. The relationships are modeled using linear
predictor functions whose parameters are estimated using the observed data. We used the scikit-learn
implementation of linear regression in our study. The regressors would be normalized before regression
by subtracting the mean and dividing by the “l2-norm”.

1.4 Evaluation

We use mean square error (MSE) as shown in equation 6 to compute the accuracy of the prediction of
WML volume. Here, Yˆi represents the predicted volume and Yi

å {Yˆ - Y }
n
1 2
MSE = i i
(6)
N i =1

We select a cut-off value of 5mL to classify the severity of WML volume. The severity of the WML load is
usually classified into four levels [23] as shown in Table 5. In our data set, the number of subjects is
comparatively small in severity level 3 and 4, therefore, we combine severity level 2, 3 and 4 into one
class named as present and severity level 1 is considered as absent class. We apply the same cut-off
value (5 mL) in the predicted WML volume to classify them into absent and present class. The distribution
of WML over the study sample is presented in Figure 10. Following that we compute sensitivity (SEN),
specificity (SPE), positive predictive value (PPV), negative predictive value (NPV) and

Table 5. Classification of the severity of white matter lesion volume [23]

Severity Level Lesion volume (LV) Subject Class

Page 8 of 14
 medRxiv preprint doi: https://doi.org/10.1101/2021.09.27.21264168; this version posted September 29, 2021. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

It is made available under a CC-BY-NC-ND 4.0 International license .

1 LV < 5 mL 52 absent
2 5 mL < LV < 10 mL 31
3 10 mL < LV < 15 mL 12 Present
4 LV > 15 mL 16

F1 score to evaluate the classification accuracy. All these measurement metrics are computed by using
TP (true positive), FP (false positive), TN (true negative) and FN (false negative) as shown in the
following equations:

TP
SEN =
TP + FN
TN
SPE =
TN + FP
TP
PPV =
TP + FP
TN
NPV =
TN + FN
PPV ´ SEN
F1 = 2 ´
PPV + SEN 1
As mentioned earlier, we use ten-fold cross validation to evaluate our prediction of WML volume. To
remove the training data bias, we shuffle the subjects randomly to produce ten different permutations of
the subjects. For each random permutation, we run ten-fold cross validation resulting 100 evaluations.
The mean MSE, SEN, SPE, PPV, NPV and F1 score are reported in the result section.

Table 6. Prediction and classification accuracy of artificial neural network (multi-layer perceptron),
support vector classifier, logistic regression classifier and random forest classifier using
combination of quantitative arteriovenular narrowing (AVN), Opacity (OP) and focal narrowing
scores(FAN)as Input Parameter.

Classifier Input Parameter Accuracy ( %) SEN SPE PPV NPV F1

RF AVN,FAN & OP 91.6 (85.5 to 95.7)% 0.84 0.97 0.96 0.89 0.92

ANN AVN,FAN & OP 91.6 (85.5 to 95.7)% 0.83 0.99 0.98 0.88 0.92

SVM AVN,FAN & OP 77.1 (69.0 to 84.0)% 0.68 0.83 0.73 0.79 0.77

LR AVN,FAN & OP 71.8 (63.2 to 79.3)% 0.65 0.75 0.53 0.83 0.71

Table 7. Prediction and classification accuracy of Artificial Neural Network (ANN) ,support vector
classifier, logistic regression classifier and random forest classifier using combinations of 2
features among quantitative arteriovenular narrowing (AVN),opacity (OP) and focal narrowing
scores(FAN) as Input Parameter at a time and using AVN, OP and FAN individually as Input
Parameter.

Classifier Input Parameter Accuracy ( %) SEN SPE PPV NPV F1

RF AVN, FAN 88.5 (81.8 to 93.5)% 0.80 0.95 0.92 0.87 0.89

Page 9 of 14
 medRxiv preprint doi: https://doi.org/10.1101/2021.09.27.21264168; this version posted September 29, 2021. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

It is made available under a CC-BY-NC-ND 4.0 International license .

AVN, OP 85.5 (78.3 to 91.0)% 0.81 0.88 0.80 0.89 0.85

FAN,OP 79.4 (71.5 to 86.0)% 0.66 0.94 0.92 0.72 0.79

AVN, FAN 88.5 (81.8 to 93.5)% 0.79 0.96 0.94 0.85 0.89

ANN
AVN, OP 85.5 (78.3 to 91.0)% 0.81 0.88 0.80 0.89 0.85

FAN,OP 79.4 (71.5 to 86.0)% 0.67 0.92 0.90 0.73 0.80

AVN, FAN 74.1 (65.7 to 81.3)% 0.66 0.79 0.64 0.80 0.74

SVM
AVN, OP 68.0 (59.2 to 75.8)% 0.58 0.74 0.55 0.76 0.68

FAN,OP 66.4 (57.6 to 74.4)% 0.56 0.72 0.49 0.77 0.66

AVN, FAN 74.8 (66.5 to 82.0)% 0.69 0.78 0.59 0.84 0.74

LR
AVN, OP 65.7 (56.9 to 73.7)% 0.56 0.69 0.39 0.82 0.64

FAN,OP 66.4 (57.6 to 74.4)% 0.55 0.73 0.53 0.74 0.66

AVN 72.5 (64.0 to 80.0)% 0.60 0.85 0.80 0.68 0.73

RF FAN 76.3 (68.1 to 83.3)% 0.64 0.89 0.86 0.71 0.77

OP 75.5 (67.3 to 82.6)% 0.90 0.73 0.39 0.98 0.73

AVN 72.5 (64.0 to 80.0)% 0.60 0.85 0.80 0.68 0.73

ANN FAN 76.3 (68.1 to 83.3)% 0.63 0.90 0.88 0.70 0.77

OP 75.6 (67.3 to 82.7)% 0.90 0.73 0.39 0.98 0.72

AVN 67.9 (59.2 to 75.8)% 0.56 0.76 0.63 0.71 0.68

SVM FAN 63.4 (54.5 to 71.6)% 0.67 0.63 0.04 0.99 0.51

OP 62.6 (53.7 to 70.9)% undefined 0.63 0.0 1.0 0.48

AVN 66.4 (57.6 to 74.4)% 0.55 0.73 0.55 0.73 0.66

LR FAN 64.1 (55.3 to 72.3)% 0.67 0.64 0.08 0.98 0.54

OP 64.1 (55.3 to 72.3)% 1.00 0.64 0.04 1.00 0.52

2. Results and Discussion

Page 10 of 14
 medRxiv preprint doi: https://doi.org/10.1101/2021.09.27.21264168; this version posted September 29, 2021. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

It is made available under a CC-BY-NC-ND 4.0 International license .

Table 6 and Table 7 shows the prediction accuracy of the proposed method for different input parameter
sets. We use four state of the art classifier models: random forest (RF) classifier, artifical neural network
(ANN) classifier, support vector machine (SVM) classifier and logistic regression (LR) classifier to predict
the WML volume absent or present classification. Table 6 and Table7 show that, all of these models
(SVM, LR, ANN and RF) are showing better prediction accuracy for combination of all three of
quantitative AVN , FAN OP scores compared to combination of any two features/input parameters of
AVN, OP and FAN scores or using AVN, OP and FAN individually. We apply a cut-off value of 5 mL as
mentioned in section 1.4 to classify the actual and predicted WML volume into absent and present class
and compute the classification accuracy. Among the four models, quantitative AVN, OP and FAN score-
based RF classifier is showing the highest sensitivity of 0.84 with a PPV score of 0.96. In contrast to that,
the maximum specificity of 0.99 is shown by quantitative AVN, OP and FAN score-based ANN with a
negative predictive value of 0.88. The overall classification accuracy is computed by F1 score, where
quantitative AVN, OP and FAN score-based RF and ANN classifier both are showing highest F1 score of
0.92. From Table 6 and Table 7 we can draw a conclusion that proposed quantitative AVN, OP and FAN
score-based models are showing better prediction accuracy compared to the qualitative AVN and FAN
score-based models.

Finally, it can be noted that quantified AVN , FAN and OP scores are showing better accuracy in
predicting the WML volume class compared to the corresponding qualitative grading, which confirms a
number of research studies [24-26] where quantitative grading shows better accuracy than qualitative
grading in disease prediction. Since the retinal imaging technology is a cheaper option, and it can be
installed in low socio-economic counties and rural areas, a retinal image based accurate white matter
prediction model can be helpful for clinicians to identify those patients who need further detailed
diagnosis. Our future goal is to evaluate our model on a large scale and longitudinal dataset.

We have shown that we can predict WML with higher accuracy using quantified retinal FAN and AVN
than the qualitative FAN and AVN. We compute the features to train the machine-learning classifier (i.e.,
SVM) to predict the brain WML severity by WML volume. WML severity levels are classified as: none
(<5ml), mild (>5ml and <10ml), moderate (>10ml and <15ml), and severe (>15ml). We evaluate our
proposed model on a dataset of 111 patients taken from the ENVIS-ion study which has retinal and MRI
images for each patient. Our model shows high accuracy in estimating the WML volume/severity (none,
mild, moderate, and severe); the mean square error (MSE) between our predicted WML load and
manually annotated WML load is 0.13. Proposed WML prediction model shows high classification
accuracy with AUC 0.80 (the expert graded i.e., qualitative AVN and FAN score based WML
prediction AUC 0.69) in classifying the healthy patients having mild and severe WML load. The ROC in
Fig. 13 shows that we can screen more than 90% of individuals with WML while picking the 50% normal
individuals at high risk. We note that we only considered two features and did not consider the
proposed parameters (age, gender, etc. which were not provided yet).

3. Conclusion

In this paper, a quantified AVN, FAN and OP score-based prediction model is proposed to predict the
severity of WML volume in the brain. The prediction model is evaluated on a dataset of 111 patients of
ENVISion study which has MRI and retinal image for each subject. Quantification of WML volume is done
by using our proposed automatic software, which is further refined by manual correction. Arteriovenous
nicking (AVN) and focal arteriolar narrowing (FAN) are graded by the expert graders of Center for Eye
Research Australia. Then each manually graded AVN and FAN affected vessel segment is quantified by
using our proposed computer aided diagnostic (CAD) methods. We adopted four state of the art
classifiers to classify the severity of WML volume. Proposed quantified AVN, OP and FAN scores
significantly improve the classification accuracy compared to the expert graded qualitative AVN, OP and

Page 11 of 14
 medRxiv preprint doi: https://doi.org/10.1101/2021.09.27.21264168; this version posted September 29, 2021. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

It is made available under a CC-BY-NC-ND 4.0 International license .

FAN scores for each of these classifiers. We obtain the optimal classification accuracy ( F1 score = 0.92)
by using both RF and ANN classifier. Our future goal is to include patients meta data information (age,
sex, gender, ethnic origin, etc.) into our model and evaluate the performance on a large-scale and
longitudinal dataset.

Proposed WML prediction model shows high classification accuracy with AUC 0.80 (compared to the
expert graded i.e., qualitative AVN and FAN score based WML prediction, AUC 0.69) in classifying
the healthy patients having mild and severe WML load.

WML may be an indicator for stroke and Alzheimer’s disease prediction, thus our model can help
screening for early detection of indicuviduals at risk of stroke and Alzheimer’s disease. At present, there
is affordable little opportunity for treatment of Alzheimer’s outside of palliative care. Recent clinical
findings suggest that successful treatment needs to start in the prodromal stages of the disease[5].
Doctors can’t definitely diagnose AD until after death when they can closely examine the brain under a
microscope [6]. However, regular Magnetic Resonance Imaging (MRI) of the brain can be used to find out
symptoms of AD [7], but regular MRI is impractical for the screening of a patient because of its high cost
and inaccessibility in remote areas. Therefore, retinal imaging can be a key for screening in the primary
care level. Similarly, we can utilize the screening tool for identification of individuals at risk of incident
stroke.

Figure Legends

Figure 1. A color fundus retinal image showing different anatomical structure of retina.

Figure 2. Example of axial fluid attenuated inversion recovery (FLAIR) MRI with (a) minor white matter
lesion and (b) extensive WML.

Figure 3. Example of arteriovenus nicking (AVN) affected artery-vein crossover (AV) point (a) and healthy
AV crossover point (b).

Figure 4. Color retinal image with focal arteriolar narrowing and light reflex (opacification or bright copper
wiring) in the center of the artery as sown by arrow.

Figure 5. Block diagram of the proposed white matter lesion (WML) volume prediction model.

Figure 6. Output images of the different steps of image processing module for arteriovenous nicking
(AVN) quantification (a) original image; (b) region of interest; (c) simplified segmentation of the artery vein
crossover point; (d) detection of venular segments.

Figure 7. Automatically computed width for vein segment 1 and vein segment 2.

Figure 8. (a) Original retinal image with user selected region of interest points marked by yellow cross point,
(b) zoomed original image, (c) cropped region of interest (marked by yellow box),(d) affine translation of the
region of interest and (e) segmented vessel region.

Figure 9. (a) Vessel segment marked with width measurements, (b) plot of the cross-sectional widths of
the measurement.

Figure 10. Distribution of WML over the study sample

Figure 11. ROC curve for quantitative feature based artifical neural network (ANN), support vector
machine (SVM), random forest, (RF), and linear regression (LR) classifiers.

Fig. 12. ROC curve for SVM random forest (RF) and linear regression (LR) classifiers (top 3 classifiers).

Page 12 of 14
 medRxiv preprint doi: https://doi.org/10.1101/2021.09.27.21264168; this version posted September 29, 2021. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

It is made available under a CC-BY-NC-ND 4.0 International license .

References

[1] C. Wilkinson et al., "Proposed international clinical diabetic retinopathy and diabetic macular
edema disease severity scales," Ophthalmology, vol. 110, no. 9, pp. 1677-1682, 2003.
[2] T. Wong and P. Mitchell, "Hypertensive retinopathy," N Engl J Med, vol. 351(22), pp. 2310-7,
2004.
[3] P. M. D'onofrio and P. D. Koeberle, "What can we learn about stroke from retinal ischemia
models?," Acta Pharmacologica Sinica, vol. 34, no. 1, pp. 91-103, 2013.
[4] S. M. Heringa, W. H. Bouvy, E. Van Den Berg, A. C. Moll, L. J. Kappelle, and G. J. Biessels,
"Associations between retinal microvascular changes and dementia, cognitive functioning, and
brain imaging abnormalities: a systematic review," Journal of Cerebral Blood Flow & Metabolism,
vol. 33, no. 7, pp. 983-995, 2013.
[5] T. Y. Wong and R. McIntosh, "Systemic associations of retinal microvascular signs: a review of
recent population-based studies," Ophthalmic and Physiological Optics, vol. 25, no. 3, pp. 195-
204, 2005.
[6] A. R. Sharrett, "A review of population-based retinal studies of the microvascular contribution to
cerebrovascular diseases," Ophthalmic epidemiology, vol. 14, no. 4, pp. 238-242, 2007.
[7] A. Bhuiyan, R. Kawasaki, E. Lamoureux, T. Y. Wong, and K. Ramamohanarao, "Retinal Vascular
Features for Cardio Vascular Disease Prediction: Review," Recent Patents on Computer Science,
vol. 3(3), pp. 164-175, 2010.
[8] C. C. Qiu et al., "Cerebral microbleeds, retinopathy, and dementia the ages-reykjavik
study," Neurology, vol. 75 (24), pp. 2221–2228, 2010.
[9] L. S. Cooper, "Retinal microvascular abnormalities and MRI-defined subclinical cerebral infarction
- The atherosclerosis risk in communities study," (in English), Stroke, vol. 37, no. 1, pp. 82-86,
2006.
[10] M. L. Baker et al., "Retinal microvascular signs, cognitive function, and dementia in older persons
the cardiovascular health study," Stroke, vol. 38(7), pp. 2041–2047, 2007.
[11] S. Debette and H. Markus, "The clinical importance of white matter hyperintensities on brain
magnetic resonance imaging: systematic review and meta-analysis," BMJ, vol. 341:c3666, 2010.
[12] M. Mortamais et al., "Spatial distribution of cerebral white matter lesions predicts progression to
mild cognitive impairment and dementia," PloS one, vol. 8, no. 2, p. e56972, 2013.
[13] C. Qiu et al., "Cerebral microbleeds, retinopathy, and dementia the ages-reykjavikstudy,"
Neurology, vol. 75(24) pp. 2221–2228, 2010.
[14] C. M. Reid et al., "Aspirin for the prevention of cognitive decline in the elderly: rationale and
design of a neuro-vascular imaging study (ENVIS-ion)," BMC Neurology, vol. 12:3, pp. 1-9, 2012.
[15] P. K. Roy et al., "Automated segmentation of white matter lesions using global neighbourhood
given contrast feature-based random forest and Markov random field," in 2014 IEEE International
Conference on Healthcare Informatics, 2014: IEEE, pp. 1-6.
[16] P. A. Yushkevich et al., "User-guided 3D active contour segmentation of anatomical structures:
significantly improved efficiency and reliability," Neuroimage, vol. 31, no. 3, pp. 1116-1128, 2006.
[17] P. K. Roy, A. Hussain, A. Bhuiyan, R. Kawasaki, and K. Ramamohanarao, "A robust and reliable
quantification method for Focal Arteriolar Narrowing in color retinal image," in 2015 IEEE 12th
International Symposium on Biomedical Imaging (ISBI), 2015: IEEE, pp. 1510-1513.
[18] P. K. Roy, U. T. Nguyen, A. Bhuiyan, and K. Ramamohanarao, "An effective automated system
for grading severity of retinal arteriovenous nicking in colour retinal images," in 2014 36th Annual
International Conference of the IEEE Engineering in Medicine and Biology Society, 2014: IEEE,
pp. 6324-6327.
[19] A. Bhuiyan et al., "Retinal artery and venular caliber grading: A semi-automated evaluation tool,"
Computers in biology and medicine, vol. 44, pp. 1-9, 2014.
[20] K. He, J. Sun, and X. Tang, "Guided image filtering," in European conference on computer vision,
2010: Springer, pp. 1-14.
[21] J. Canny, "A computational approach to edge detection," IEEE Transactions on pattern analysis
and machine intelligence, no. 6, pp. 679-698, 1986.

Page 13 of 14
 medRxiv preprint doi: https://doi.org/10.1101/2021.09.27.21264168; this version posted September 29, 2021. The copyright holder for this
preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

It is made available under a CC-BY-NC-ND 4.0 International license .

[22] C. Shackelford, G. Long, J. Wolf, C. Okerberg, and R. Herbert, "Qualitative and quantitative
analysis of nonneoplastic lesions in toxicology studies," Toxicologic pathology, vol. 30, no. 1, pp.
93-96, 2002.
[23] C.-C. Chang and C.-J. Lin, "LIBSVM: a library for support vector machines," ACM transactions on
intelligent systems and technology (TIST), vol. 2, no. 3, pp. 1-27, 2011.
[24] M. D. Steenwijk et al., "Accurate white matter lesion segmentation by k nearest neighbor
classification with tissue type priors (kNN-TTPs)," NeuroImage: Clinical, vol. 3, pp. 462-469,
2013.
[25] H. K. Genant et al., "Comparison of semiquantitative visual and quantitative morphometric
assessment of prevalent and incident vertebral fractures in osteoporosis," Journal of Bone and
Mineral Research, vol. 11, no. 7, pp. 984-996, 1996.
[26] H. Weissleder and R. Weissleder, "Lymphedema: evaluation of qualitative and quantitative
lymphoscintigraphy in 238 patients," Radiology, vol. 167, no. 3, pp. 729-735, 1988.

Page 14 of 14
Hindawi Publishing Corporation
BioMed Research International
Volume 2015, Article ID 636548, 8 pages
http://dx.doi.org/10.1155/2015/636548

Research Article
Analysis of Retinal Peripapillary Segmentation in Early
Alzheimer’s Disease Patients

Elena Salobrar-Garcia,1 Irene Hoyas,1,2 Mercedes Leal,1,2 Rosa de Hoz,1,3

Blanca Rojas,1,2 Ana I. Ramirez,1,3 Juan J. Salazar,1,3 Raquel Yubero,4 Pedro Gil,2,4
Alberto Triviño,1,2 and José M. Ramirez1,2
1
Instituto de Investigaciones Oftalmológicas Ramón Castroviejo, Universidad Complutense de Madrid (UCM),
28040 Madrid, Spain
2
Departamento de Oftalmologı́a y ORL, Facultad de Medicina, UCM, 28040 Madrid, Spain
3
Departamento de Oftalmologı́a y ORL, Facultad de Óptica y Optometrı́a UCM, 28040 Madrid, Spain
4
Servicio de Geriatrı́a, Hospital Universitario Clı́nico San Carlos, 28040 Madrid, Spain

Correspondence should be addressed to José M. Ramirez; [email protected]

Received 23 March 2015; Accepted 8 June 2015

Academic Editor: Jose F. Arevalo

Copyright © 2015 Elena Salobrar-Garcia et al. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly
cited.

Decreased thickness of the retinal nerve fiber layer (RNFL) may reflect retinal neuronal-ganglion cell death. A decrease in the RNFL
has been demonstrated in Alzheimer’s disease (AD) in addition to aging by optical coherence tomography (OCT). Twenty-three
mild-AD patients and 28 age-matched control subjects with mean Mini-Mental State Examination 23.3 and 28.2, respectively, with
no ocular disease or systemic disorders affecting vision, were considered for study. OCT peripapillary and macular segmentation
thickness were examined in the right eye of each patient. Compared to controls, eyes of patients with mild-AD patients showed no
statistical difference in peripapillary RNFL thickness (𝑃 > 0.05); however, sectors 2, 3, 4, 8, 9, and 11 of the papilla showed thinning,
while in sectors 1, 5, 6, 7, and 10 there was thickening. Total macular volume and RNFL thickness of the fovea in all four inner
quadrants and in the outer temporal quadrants proved to be significantly decreased (𝑃 < 0.01). Despite the fact that peripapillary
RNFL thickness did not statistically differ in comparison to control eyes, the increase in peripapillary thickness in our mild-AD
patients could correspond to an early neurodegeneration stage and may entail the existence of an inflammatory process that could
lead to progressive peripapillary fiber damage.

1. Introduction retinal degenerations have been described [5, 6]. The retinal
nerve fiber layer (RNFL) is composed of retinal-ganglion cell
Alzheimer’s disease (AD), the most common cause of demen- axons that form the optic nerve. Decreased thickness of the
tia, afflicts 67 of every 1000 persons over age 65. Its prevalence RNFL can reflect retinal neuronal-ganglion cell death and
and incidence increase exponentially with age [1, 2]. In 2006, axonal loss in the optic nerve [7, 8].
the worldwide prevalence of Alzheimer’s was 26.6 million, The RNFL reportedly thins with aging [9, 10]. Some stud-
and by 2050, the prevalence will quadruple, meaning that by ies have also shown a decrease of the RNFL in AD in addition
that time 1 in 85 persons worldwide will be living with the to aging [7, 8, 11–16]. Hinton et al. [17] were the first to show
disease [2]. histopathological evidence of retinal-ganglion cell loss and
AD is characterized by a decline in cognitive function, optic-nerve degeneration in patients with AD. These findings
loss of learning and memory, and the formation of neuritic were then confirmed in several follow-up studies [18–21].
plaques and neurofibrillary tangles, primarily in the cerebral Indeed, the large magnocellular cell axon degeneration in AD
cortex [3, 4]. has been documented [19, 22]. Other histopathology studies
The retina is a projection of the brain, and a number [23–28], however, have failed to confirm these findings and
of similarities between AD pathology and several distinct suggest that methodological differences were responsible for
2 BioMed Research International

the contradictory results, due to a different postmortem delay in their medical record). Of the 29 mild-AD patients and 37
in axon count or difficulties in obtaining well-preserved age-matched control subjects selected (normal MMSE
myelinated axons. scores), 6 mild-AD patients and 9 age-matched control
Currently it is thought that retinal ganglion cell (RGC) subjects were subsequently excluded due to posterior
loss in AD might result from amyloid deposits in the eye pole pathology including macular degeneration, drusen,
and/or retina. Amyloid-beta plaques as well as oligomers have suspicion of glaucoma, glaucoma, epiretinal membrane, or
been reported in postmortem retinal tissue from patients cataract that prevented ocular examination. Because of this
with AD and in a mouse model of AD, as well as in human selection, 23 patients with mild AD and 28 age-matched
retinal tissue in vivo [29]. Therefore, amyloid accumulation control subjects were considered for the study. Informed
in the eye or retina of patients with AD may result in consent was obtained from both groups. The research
the degeneration of RGC in parallel to amyloid-beta-related followed the tenets of the Declaration of Helsinki, and the
neurodegeneration in the brain [29]. protocol was approved by the local ethics committee.
Diagnosis and progression of AD, especially early cases,
are complicated because of imprecise neuropsychological 2.2. Methods. For the ophthalmological part of the study, the
testing, sophisticated but expensive neuroimaging tech- right eye of each patient was analyzed. All participants met
niques, and invasive sampling of cerebrospinal fluid [30, 31]. the following inclusion criteria: being free of ocular disease,
Improved methods for screening and early detection are AREDS Clinical Lens Standards <2, retinal drusen, and
essential to identify cognitively normal individuals who have systemic disorders affecting vision; having a best corrected
a high risk of developing AD, so that treatment can be devel- VA of 20/40; having a ±5 spherocylindrical refractive error;
oped to delay the progression of the disease [32]. Currently, and having intraocular pressure of less than 20 mmHg. For
there is no definitive antemortem diagnosis for AD, and new screening, all AD patients and control subjects underwent a
biomarkers for diagnosis are therefore needed. Over the last complete ophthalmologic examination, including assessment
few decades, very accurate tools for analyzing the eye fundus of VA, refraction, anterior segment biomicroscopy, appla-
have been developed (i.e., OCT, laser polarimetry), opening nation tonometry (Perkins MKII tonometer, Haag Streit-
new ways of examining the retina in vivo. Reliance Medical, Switzerland), dilated fundus examination,
OCT is a reliable noninvasive technique, routinely used and OCT. The RNFL thickness and macular thickness were
in ophthalmology to visualize and quantify the layers of the measured by OCT Model 3D OCT-1000 (Topcon, Japan)
retina. OCT enables quantitative cross-sectional imaging of after pupil dilatation. The RNFL thickness was scanned 3
the RNFL and macular volume. A recent study published consecutive times per patient in each area studied. The mean
by our group [33] has shown that in mild-AD patients values were considered for statistical analysis. All tests were
the first affected area of the retina is the macular area. As performed by the same optometrist (ESG) under the same
the neurodegeneration progresses, a significant decline in conditions. These tests were selected considering that in
peripapillary RNFL thickness will become apparent. this developmental stage of the disease the results were not
The goal of the present study was to examine in detail influenced by the patient’s cognitive impairment.
peripapillary and macular segmentation in order to deter- The peripapillary RNFL thickness parameters evaluated
mine which is the earliest thinned area in patients with mild in this study were average thickness (360∘ measurement),
AD which may be used, in the future, as a predictive tool. thickness for each 12-o’clock hour position with the 3-o’clock
position as nasal, 6-o’clock position as inferior, 9-o’clock posi-
tion as temporal, and 12-o’clock position as superior. Macular
2. Material and Methods RNFL thickness data were displayed in three concentric rings
centered in the foveola that were distributed as follows: a
2.1. Subjects. To select patients, we reviewed the Database central macular ring, 1 mm away from the fovea; an inner
of the Memory Unit of the Hospital Clinico San Carlos in macular ring, 3 mm away from the fovea; and an outer
Madrid (Spain), consisting of a total of 2635 patients. First, macular ring, 6 mm away from the fovea. As a result, the total
we excluded the patients with a Global Deterioration Scale area studied made up a 6 mm macular map. In addition, the
(GDS) over 4 and then those with a mood or psychiatric
inner and outer rings were each divided into four quadrants
disorder. Next, we took into account 87 patients with mild
(superior, inferior, nasal, and temporal) (Figure 1). The total
AD. These patients, according to the National Institute of
volume of the macula as provided by the OCT was also
Neurological and Communicative Disorders and Stroke-AD
calculated. The good scan criteria were determined as the
and Related Disorders Association and the Diagnostic and
signal-to-noise ratio >30 and accepted A-scans >95% in fast
Statistical Manual of Mental Disorders IV, had mild cognitive
RNFL scanning. All measurements are given in microns,
impairment according to the Clinical Dementia Rating
according to the calibration provided by the manufacturers
scale. Then ophthalmic medical records of these patients
were reviewed, excluding patients who were previously and the total volume in mm3 .
diagnosed with an ophthalmological pathology (glaucoma
or suspected glaucoma, media opacity, and retinal diseases). 2.3. Statistical Analysis. The data are reported as mean values
After this analysis, 29 patients with AD satisfied all the ± SD. The differences between mild AD and control eyes
requirements to participate in the study (GDS over 4 and were analyzed using the Mann-Whitney test. Data for the
free of ocular disease and systemic disorders affecting vision statistical analysis were introduced and processed in a SPSS
BioMed Research International 3

Photo
12
11 1

10 2

9 3

8 4

7 5
6
(a)
ETDRS

T N

1 5 95 99
(%)
Average thickness (𝜇m) 273,1
Center thickness (𝜇m) 199
Total volume (mm3 ) 7,71

(b)

Figure 1: OCT report of retinal nerve fiber layer (RNFL) thickness analysis. (a) Peripapillary OCT. The thickness for each 12-o’clock hour
position with the 3-o’clock position as nasal, 6-o’clock position as inferior, 9-o’clock position as temporal, and 12-o’clock position as superior
was evaluated. (b) Macular OCT. Diagram showing the concentric rings and quadrants considered for analysis of the macular RNFL thickness
and measurements automatically provided by the analyzer.

19.0 (SPSS Inc©, Inc, Chicago, IL, USA). A 𝑃 value of <0.05 between the study groups. The MMSE scores in mild-AD
was considered statistically significant. patients were significantly decreased in comparison with age-
matched control subjects (Table 1). All mild-AD patients had
MMSE values higher than 17.
3. Results
3.1. Optical Coherence Tomography
Demographic and clinical data for the mild-AD patients and
control group are shown in Table 1. No statistically significant Peripapillary RNFL Segmentation Thickness. Peripapillary
differences in age, gender, or educational level were found RNFL thickness values (Figure 2(a)) showed no statistical
4 BioMed Research International

130.00
300.00
120.00
∗
RNFL thickness (𝜇m)

110.00 250.00

RNFL thickness (𝜇m)

100.00 200.00
90.00
150.00
80.00
70.00 100.00
60.00 50.00
50.00
1 2 3 4 5 6 7 8 9 10 11 12 0.00
Segmented sector RNFL thickness

AD AD
Control Control
(a) (b)
320.00 280.00

310.00 ∗ ∗ 270.00
∗

RNFL thickness (𝜇m)

RNFL thickness (𝜇m)

260.00
300.00
∗ 250.00
290.00 ∗
240.00
280.00
230.00
270.00 220.00

260.00 210.00
Superior Nasal Inferior Temporal Superior Nasal Inferior Temporal

AD AD
Control Control
(c) (d)

Figure 2: Mean data of RNFL thickness against eye quadrants assessed with optical coherence tomography (OCT). (a) Peripapillary
segmentation retinal nerve fiber layer, (b) Central macular ring (1 mm away from the fovea). (c) Inner macular ring (3 mm away from the
fovea). (d) Outer macular ring (6 mm away from the fovea). ∗ 𝑃 value < 0.01.

Table 1: Demographic and clinical data of the study groups. Macular RNFL Thickness and Total Volume. As we reported
in a previous study [33] the analysis of the RNFL revealed
AD Control
𝑃 value that, in patients with mild AD, the values for the central
(𝑛 = 23) (𝑛 = 28)
ring (fovea) (Figure 2(b)) and the four inner quadrants (3 mm
Age§ 79.3 ± 4.6 72.3 ± 5.1 0.274
from the fovea) (Figure 2(c)) were significantly decreased in
Gender
comparison with control subject (𝑃 < 0.05 in both instances;
Male 9 9
0.615 Mann-Whitney 𝑈 test) (Table 2). The RNFL thickness of the
Female 14 19
outer macular quadrants (6 mm from the fovea) (Figure 2(d))
Race Caucasian Caucasian
in patients with mild AD was diminished in comparison
23.3 ± 3.1 28.2 ± 1.9
MMSE§ 0.001∗ with control subjects; however, only the values of the outer
Range (17–29) Range (25–31)
temporal quadrant were significantly lower (𝑃 < 0.05; Mann-
§
Mean value ± SD; ∗ 𝑃 < 0.01 [AD, Alzheimer’s disease; MMSE, Mini-Mental Whitney 𝑈 test) (Table 2).
State Examination; SD, standard deviation].
The total macular volume was significantly reduced in
mild-AD patients in comparison with control subjects (𝑃 <
difference between mild-AD patients and control subjects 0.05; Mann-Whitney 𝑈 test) (Table 2).
(Table 2).
Although the differences were not significant in any of the
sectors, it was shown that peripapillary sectors 2, 3, 4, 8, 9, 11, 4. Discussion
and 12 were thinner in the mild-AD patients than in controls;
in peripapillary sectors 1, 5, 6, 7, and 10 the retina in mild-AD Alzheimer’s dementia syndromes, like all neurodegenerative
patients was thicker with respect to the control (Figure 2(a); diseases, lack objective disease- and stage-specific biomarkers
Table 2). [34]. As a part of the CNS, the retina or neural portion
BioMed Research International 5

Table 2: RNFL thickness and total macular volume.

Retinal area of study AD group§ Control group§ % RNFL decrease 𝑃 value
Sector 1 101.2 ± 24.4 100.0 ± 16.9 1.24 0.790
Sector 2 96.8 ± 20.8 100.4 ± 14.4 −3.62 0.618
Sector 3 80.1 ± 22.9 90.8 ± 16.0 −11.85 0.084
Sector 4 60.5 ± 18.3 65.4 ± 10.3 −7.60 0.464
Sector 5 73.7 ± 12.7 68.3 ± 11.3 7.83 0.173
Sector 6 101.4 ± 23.7 99.3 ± 19.0 2.02 0.790
Peripapillary thickness (𝜇m)
Sector 7 119.4 ± 19.1 116.3 ± 21.9 2.62 0.756
Sector 8 111.7 ± 21.6 114.5 ± 24.5 −2.48 0.564
Sector 9 66.3 ± 18.2 67.5 ± 16.5 −1.80 0.877
Sector 10 60.7 ± 13.8 57.3 ± 9.6 5.99 0.464
Sector 11 77.0 ± 20.6 78.5 ± 14.7 −1.96 0.94
Sector 12 101.3 ± 27.1 110.7 ± 15.8 −8.45 0.335
Foveal thickness (𝜇m) Fovea 221.2 ± 21.6 243.7 ± 24.8 −9.24 0.015∗
Superior area 283.4 ± 11.1 294.9 ± 18.1 −3.91 0.002∗
Inferior area 279.8 ± 18.1 295.8 ± 13.5 −5.40 0.002∗
Inner macular quadrant (𝜇m)
Nasal area 285.6 ± 17.2 300.1 ± 15.1 −4.83 0.007∗
Temporal area 273.1 ± 12.7 285.2 ± 14.6 −4.22 0.002∗
Superior area 245.4 ± 12.5 252.1 ± 13.7 −2.65 0.084
Inferior area 242.8 ± 17.4 245.2 ± 13.9 −0.99 0.531
Outer macular quadrant (𝜇m)
Nasal area 263.7 ± 12.1 267.5 ± 19.1 −1.41 0.110
Temporal area 228.0 ± 18.8 238.5 ± 12.3 −4.43 0.009∗
Total macular volume (mm3 ) 7.1 ± 0.3 7.3 ± 0.3 9.34 0.024∗
§
Mean value ± SD; ∗ 𝑃 < 0.05 [AD, Alzheimer’s disease; RNFL: retinal nerve fiber layer; SD, standard deviation].

of the eye shares many features with the brain, including difference between the peripapillary RNFL segmentation and
embryological origin as well as anatomical and physiological the macula thickness in our mild-AD patients in that only the
characteristics. Its peripheral location provides an accessible macular thickness was significantly decreased in comparison
and noninvasive way of examining brain pathology [35]. OCT with the control group.
is a reliable noninvasive technique that enables quantitative Widespread axonal degeneration in the optic nerve was
cross-sectional imaging of the RNFL [36]. found in a postmortem study of patients with AD [17].
Thinning of the RNFL has been found in several neu- Morphometric analysis of the whole-mount retina has shown
rological diseases, such as Parkinson’s disease [16, 37–39], that Alzheimer’s patients had a predominant loss of the largest
dementia with Lewy Bodies [16], amnestic mild cognitive class of retinal-ganglion cells (M-cells), which could be a
impairment [8, 15], neuromyelitis optica [40], migraine [41], primary process or a consequence of retrograde neurodegen-
and AD [8, 11–17, 20, 27, 36, 37, 42–45]. The loss of RNFL eration occurring in the cortical regions [19]. In vivo studies
thickness in AD is linked to a depletion of retinal-ganglion using different methodologies have confirmed optic-nerve-
cells and optic-nerve axons [13, 14, 32, 46, 47]. It has been fiber damage in AD when compared with controls. Optic
postulated that the defects in RNFL may be the earliest sign disc pallor, pathologic disc cupping, and thinning of the
of AD, even prior to damage to the hippocampal region that neuroretinal rim and the RNFL have been reported in studies
impairs memory [36]. In addition, published data suggest an based on the subjective evaluation of fundus photographs
association between the thinning of RNFL and severity of AD [11, 42] and the optic-nerve analyzer [42].
[8, 11]. There is controversy on the reduction of the peripapillary
In the present work, we compare the peripapillary RNFL RNFL thickness measured by OCT in AD. A reduction in the
segmentation thickness, macular thickness, and the total thickness of all peripapillary RNFL quadrants as measured by
macular volume in mild-AD patients and age-matched con- OCT has been reported [12, 43], and it has been suggested that
trol subjects. One of the relevant issues of the study was this morphologic abnormality is related to retinal dysfunc-
that the sample analyzed here was homogeneous in that (i) tion as revealed by abnormal patterns in electroretinogram
all patients had recently been diagnosed as having mild AD responses [43]. However, some OCT studies on peripapillary
(GDS 4, Reisberg scale [48]) with mean MMSE score values thickness in AD [8, 13, 15, 16, 36, 44, 45, 49] found that the
of 23.7 ± 3.3; (ii) all the individuals were Caucasians; and (iii) RNFL thinning was restricted to the superior quadrant [13,
there were no significant differences in age or educational 44, 50–52] or to the superior and inferior quadrants [15, 37,
level among the groups. The results of our study showed a 45] in comparison with control subjects. Some studies have
6 BioMed Research International

correlated cognitive decline with decreased RNFL thickness neuroinflammation. In the degenerating retina, endogenous
[33, 50, 53]. It has been suggested that the inferior quadrant of signals activate microglial cells, leading to their local prolif-
the RNFL may be a more specific and sensitive area than other eration, migration, enhanced phagocytosis, and secretion of
RNFL quadrants in predicting the deterioration of cognitive cytokines, chemokines, and neurotoxins. These immunolog-
status to reflect retinal abnormality in the early stages of AD ical responses and the loss of limiting control mechanisms
[15, 50]. The reason for the variability of the results among may contribute significantly to retinal tissue damage and
studies could be related to MMSE scores. Thus, Parisi et al. proapoptotic events in retinal neurodegeneration [57–59]. A
[43] and Iseri et al. [12], whose patients had more advanced limitation to be considered in our study, as well as those
AD (ranges of MMSE scores 11 to 19 and 8 to 28, resp.), reported in the literature on RNFL thickness evaluation
showed a reduction in RNFL thickness in all peripapillary by OCT, is the number of patients included. Studies on
quadrants. Kesler et al. [15], whose patients had a mean early-stage Alzheimer’s patients are difficult to perform, one
MMSE score of 23.6, showed a decrease in the superior and reason being that these patients usually come for diagnosis at
inferior peripapillary quadrants. By contrast, both Berisha advanced stages of the disease. Taking this into consideration
et al. [13], whose Alzheimer’s patients had higher MMSE and the homogeneity of the patients included in the present
scores (17 to 30), and Paquet et al. [8], whose patients had work, we consider that our data provide preliminary evidence
a mean MMSE score of 22.6, found a thickness reduction to warrant a more extensive study.
only in the superior peripapillary quadrant, postulating this
finding as being the earliest peripapillary retinal damage
in AD patients. In our patients, with MMSE values similar 5. Conclusions
to those reported by Berisha and Paquet, the reduction of In the present study, the analysis of the OCT values of both
mean peripapillary RNFL thickness did not reach statisti- peripapillary and macular RNFL thickness in patients with
cal significance in comparison to control, but peripapillary mild AD (MMSE = 23.7) showed that only in the macula
RNFL thickness diminished or increased, depending on the was there a significant thickness reduction compared to
segment studied. It should be noted that the thinning sectors aged-matched controls. Our data, taken together with those
of papilla corresponded to 2, 3, 4, 8, 9, and 11, while sectors 1, reported in the literature, move us to propose the hypothesis
5, 6, 7, and 10 showed a thickening. These values differ from that the first affected area of the retina in mild AD is the
those found in the same sectors of the controls but in any case macular area, where, due to the arrangement of the multilayer
reach statistical significance. bodies of the ganglion cells, the decrease is easier to detect.
Most authors, although working in more advanced stages Subsequently, as the neurodegeneration progresses, a sig-
of disease (MMSE > 23.7), agree that the peripapillary nificant decline in peripapillary RNFL thickness will become
RNFL thinning is significant in the superior and inferior apparent. The study of the peripapillary segmentation reveals,
sectors [13, 37, 44]. However, sectors 1, 5, 6, 7, and 10 in in a more accurate way, the changes that occur in RNFL
our patients showed thickening. This dissimilarity could be thickness in relation to the macular-thickness changes. In
explained because of difference in the stage of the disease, this sense, our patients with mild AD differed with respect
which in our case corresponded to a much earlier stage. This to controls, although without reaching statistical significance;
tendency towards greater thickness, although not statistically perhaps due to the early stage of the disease. In addition, the
significant, could be related to the findings of Ascaso et al. in increase in peripapillary thickness in our mild-AD patients
the macula of patients with mild cognitive impairment (MCI) may indicate the existence of an inflammatory process that
and AD. Patients with MCI had greater RNFL thickness would lead to neurodegeneration of the peripapillary fibers.
compared to AD and controls, suggesting that this difference More extensive studies should be conducted to test these
could be caused by inflammation after gliosis neuronal findings.
death [54]. Similarly, the increase in peripapillary thickness
in our mild-AD patients in the sectors 1, 5, 6, 7, and 10,
corresponding to the superior and inferior sectors, may Conflict of Interests
indicate a phase of inflammation and gliosis of neural tissue
prior to the degenerative process. The authors declare that there is no conflict of interests
Reactive astrogliosis in the brain is a well-known feature regarding the publication of this paper.
of AD, but its role in AD is not well understood. Reactive
astrogliosis tends to be focal in AD. Reactive astrocytes are Acknowledgments
intimately associated with amyloid plaques or diffuse amyloid
deposits. Astrocytes surround them with dense layers of This work was supported by the Ophthalmological Network
processes as if forming miniature scars around them, perhaps OFTARED (RD12-0034/0002: Prevención, Detección Precoz
to wall them off and act as neuroprotective barriers [55]. It y Tratamiento de la Patologı́a Ocular Prevalente Degenerativa
is plausible that, in the early stages of the disease, microglial y Crónica), of the Institute of Health of Carlos III of the
activation could help remove amyloid plaques, while in Spanish Ministry of Economy. This work has been funded by
later phases proinflammatory cytokines induced by microglia the PN I+D+i 2008–2011, by the ISCIII-Subdirección General
could contribute to neurodegenerative process [56, 57]. de Redes y Centros de Investigación Cooperativa, and by
In the same way, retinal neurodegenerative diseases the European programme FEDER; grants to Elena Salobrar-
are also associated with chronic microglial activation and Garcia are currently supported by a Predoctoral Fellowship
BioMed Research International 7

(FPU) from the Spanish Ministry of Education, Culture and [16] T. Moreno-Ramos, J. Benito-León, A. Villarejo, and F. Bermejo-
Sport. Pareja, “Retinal nerve fiber layer thinning in dementia associ-
ated with parkinson’s disease, dementia with lewy bodies, and
alzheimer’s disease,” Journal of Alzheimer’s Disease, vol. 34, no.
References 3, pp. 659–664, 2013.
[17] D. R. Hinton, A. A. Sadun, J. C. Blanks, and C. A. Miller, “Optic-
[1] D. Hirtz, D. J. Thurman, K. Gwinn-Hardy, M. Mohamed, A. R.
nerve degeneration in Alzheimer’s disease,” The New England
Chaudhuri, and R. Zalutsky, “How common are the ‘common’
Journal of Medicine, vol. 315, no. 8, pp. 485–487, 1986.
neurologic disorders?” Neurology, vol. 68, no. 5, pp. 326–337,
2007. [18] J. C. Blanks, D. R. Hinton, A. A. Sadun, and C. A. Miller,
“Retinal ganglion cell degeneration in Alzheimer’s disease,”
[2] R. Brookmeyer, E. Johnson, K. Ziegler-Graham, and H. M. Brain Research, vol. 501, no. 2, pp. 364–372, 1989.
Arrighi, “Forecasting the global burden of Alzheimer’s disease,”
[19] A. A. Sadun and C. J. Bassi, “Optic nerve damage in Alzheimer’s
Alzheimer’s and Dementia, vol. 3, no. 3, pp. 186–191, 2007.
disease,” Ophthalmology, vol. 97, no. 1, pp. 9–17, 1990.
[3] J. L. Cummings, H. V. Vinters, G. M. Cole, and Z. S. Khacha-
[20] J. C. Blanks, Y. Torigoe, D. R. Hinton, and R. H. I. Blanks,
turian, “Alzheimer’s disease,” Neurology, vol. 51, no. 1, supple-
“Retinal pathology in Alzheimer’s disease. I. Ganglion cell loss
ment 1, pp. S2–S17, 1998.
in foveal/parafoveal retina,” Neurobiology of Aging, vol. 17, no. 3,
[4] B. J. Small, E. Gagnon, and B. Robinson, “Early identification pp. 377–384, 1996.
of cognitive deficits: preclinical Alzheimer’s disease and mild [21] J. C. Blanks, S. Y. Schmidt, Y. Torigoe, K. V. Porrello, D. R.
cognitive impairment,” Geriatrics, vol. 62, no. 4, pp. 19–23, 2007. Hinton, and R. H. I. Blanks, “Retinal pathology in Alzheimer’s
[5] S. J. McKinnon, “Glaucoma: ocular Alzheimer’s disease?” Fron- disease. II. Regional neuron loss and glial changes in GCL,”
tiers in Bioscience, vol. 8, pp. s1140–s1156, 2003. Neurobiology of Aging, vol. 17, no. 3, pp. 385–395, 1996.
[6] K. Chiu, K. So, and R. C. Chang, “Progressive neurodegener- [22] B. Katz, S. Rimmer, V. Iragui, and R. Katzman, “Abnormal
ation of retina in Alzheimer’s disease—are 𝛽-amyloid peptide pattern electroretinogram in Alzheimer’s disease: evidence for
and Tau new pathological factors in glaucoma?” in Glaucoma— retinal ganglion cell degeneration?” Annals of Neurology, vol. 26,
Basic and Clinical Aspects, S. Rumelt, Ed., pp. 157–177, 2013. no. 2, pp. 221–225, 1989.
[7] R. Varma, S. Bazzaz, and M. Lai, “Optical tomography- [23] C. A. Curcio and D. N. Drucker, “Retinal ganglion cells in
measured retinal nerve fiber layer thickness in normal latinos,” Alzheimer’s disease and aging,” Annals of Neurology, vol. 33, no.
Investigative Ophthalmology & Visual Science, vol. 44, no. 8, pp. 3, pp. 248–257, 1993.
3369–3373, 2003. [24] D. C. Davies, P. McCoubrie, B. McDonald, and K. A. Jobst,
[8] C. Paquet, M. Boissonnot, F. Roger, P. Dighiero, R. Gil, and “Myelinated axon number in the optic nerve is unaffected by
J. Hugon, “Abnormal retinal thickness in patients with mild Alzheimer’s disease,” British Journal of Ophthalmology, vol. 79,
cognitive impairment and Alzheimer’s disease,” Neuroscience no. 6, pp. 596–600, 1995.
Letters, vol. 420, no. 2, pp. 97–99, 2007. [25] L. Justino, M.-J. Kergoat, H. Bergman, H. Chertkow, A. Robil-
[9] M. J. Cohen, E. Kaliner, S. Frenkel, M. Kogan, H. Miron, and E. lard, and H. Kergoat, “Neuroretinal function is normal in early
Z. Blumenthal, “Morphometric analysis of human peripapillary dementia of the Alzheimer type,” Neurobiology of Aging, vol. 22,
retinal nerve fiber layer thickness,” Investigative Ophthalmology no. 4, pp. 691–695, 2001.
& Visual Science, vol. 49, no. 3, pp. 941–944, 2008. [26] H. Kergoat, M.-J. Kergoat, L. Justino, A. Robillard, H. Bergman,
and H. Chertkow, “Normal optic nerve head topography in the
[10] C. Bowd, L. M. Zangwill, E. Z. Blumenthal et al., “Imaging of the
early stages of dementia of the Alzheimer type,” Dementia and
optic disc and retinal nerve fiber layer: the effects of age, optic
Geriatric Cognitive Disorders, vol. 12, no. 6, pp. 359–363, 2001.
disc area, refractive error, and gender,” Journal of the Optical
Society of America A: Optics and Image Science, and Vision, vol. [27] H. Kergoat, M.-J. Kergoat, L. Justino, H. Chertkow, A. Robillard,
19, no. 1, pp. 197–207, 2002. and H. Bergman, “An evaluation of the retinal nerve fiber
layer thickness by scanning laser polarimetry in individuals
[11] T. R. Hedges III, R. P. Galves, D. Speigelman, N. R. Barbas, E.
with dementia of the Alzheimer type,” Acta Ophthalmologica
Peli, and C. J. Yardley, “Retinal nerve fiber layer abnormalities in
Scandinavica, vol. 79, no. 2, pp. 187–191, 2001.
Alzheimer’s disease,” Acta Ophthalmologica Scandinavica, vol.
74, no. 3, pp. 271–275, 1996. [28] H. L. N. Kergoat, M.-J. Kergoat, L. Justino, H. Chertkow, A.
Robillard, and H. Bergman, “Visual retinocortical function in
[12] P. K. Iseri, Ö. Altinaş, T. Tokay, and N. Yüksel, “Relationship dementia of the Alzheimer type,” Gerontology, vol. 48, no. 4, pp.
between cognitive impairment and retinal morphological and 197–203, 2002.
visual functional abnormalities in Alzheimer disease,” Journal
[29] Y. Koronyo, B. C. Salumbides, K. L. Black, and M. Koronyo-
of Neuro-Ophthalmology, vol. 26, no. 1, pp. 18–24, 2006.
Hamaoui, “Alzheimer’s disease in the retina: imaging retinal A𝛽
[13] F. Berisha, G. T. Feke, C. L. Trempe, J. W. McMeel, and C. L. plaques for early diagnosis and therapy assessment,” Neurode-
Schepens, “Retinal abnormalities in early Alzheimer’s disease,” generative Diseases, vol. 10, no. 1–4, pp. 285–293, 2012.
Investigative Ophthalmology & Visual Science, vol. 48, no. 5, pp. [30] R. C. Petersen, J. C. Stevens, M. Ganguli, E. G. Tangalos, J. L.
2285–2289, 2007. Cummings, and S. T. DeKosky, “Practice parameter: early detec-
[14] D. A. Valenti, “Neuroimaging of retinal nerve fiber layer in AD tion of dementia: mild cognitive impairment (an evidence-
using optical coherence tomography,” Neurology, vol. 69, no. 10, based review) Report of the Quality Standards Subcommittee of
article 1060, 2007. the American Academy of Neurology,” Neurology, vol. 56, no. 9,
[15] A. Kesler, V. Vakhapova, A. D. Korczyn, E. Naftaliev, and M. pp. 1133–1142, 2001.
Neudorfer, “Retinal thickness in patients with mild cognitive [31] J. C. Morris, “Mild cognitive impairment and preclinical
impairment and Alzheimer’s disease,” Clinical Neurology and Alzheimer’s disease,” Geriatrics, vol. 60, no. 6, supplement, pp.
Neurosurgery, vol. 113, no. 7, pp. 523–526, 2011. 9–14, 2005.
8 BioMed Research International

[32] M. K. Ikram, C. Y. Cheung, T. Y. Wong, and C. P. L. H. Chen, [48] B. Reisberg, S. H. Ferris, M. J. de Leon, and T. Crook, “The global
“Retinal pathology as biomarker for cognitive impairment and deterioration scale for assessment of primary degenerative
Alzheimer’s disease,” Journal of Neurology, Neurosurgery & dementia,” American Journal of Psychiatry, vol. 139, no. 9, pp.
Psychiatry, vol. 83, no. 9, pp. 917–922, 2012. 1136–1139, 1982.
[33] E. S. Garcia-Martin, B. Rojas, A. I. Ramirez et al., “Macular [49] M. M. Moschos, I. Markopoulos, I. Chatziralli et al., “Structural
thickness as a potential biomarker of mild Alzheimer’s disease,” and functional impairment of the retina and optic nerve in
Ophthalmology, vol. 121, no. 5, pp. 1149–1151, 2014. Alzheimer’s disease,” Current Alzheimer Research, vol. 9, no. 7,
[34] D. S. Geldmacher, “Differential diagnosis of dementia syn- pp. 782–788, 2012.
dromes,” Clinics in Geriatric Medicine, vol. 20, no. 1, pp. 27–43, [50] Y. Shen, Z. Shi, R. Jia et al., “The attenuation of retinal nerve
2004. fiber layer thickness and cognitive deterioration,” Frontiers in
[35] N. Patton, T. Aslam, T. MacGillivray, A. Pattie, I. J. Deary, Cellular Neuroscience, vol. 7, article 142, 2013.
and B. Dhillon, “Retinal vascular image analysis as a potential [51] R. Kromer, N. Serbecic, L. Hausner, F. Aboul-Enein, L. Froelich,
screening tool for cerebrovascular disease: a rationale based and S. Beutelspacher, “Detection of retinal nerve fiber layer
on homology between cerebral and retinal microvasculatures,” defects in Alzheimer’s disease using SD-OCT,” Frontiers in
Journal of Anatomy, vol. 206, no. 4, pp. 319–348, 2005. Psychiatry, vol. 5, p. 22, 2014.
[36] X.-F. He, Y.-T. Liu, C. Peng, F. Zhang, S. Zhuang, and J.-S. [52] S. Kirbas, K. Turkyilmaz, O. Anlar, A. Tufekci, and M. Durmus,
Zhang, “Optical coherence tomography assessed retinal nerve “Retinal nerve fiber layer thickness in patients with Alzheimer
fiber layer thickness in patients with Alzheimer’s disease: a sisease,” Journal of Neuro-Ophthalmology, vol. 33, no. 1, pp. 58–
meta-analysis,” International Journal of Ophthalmology, vol. 5, 61, 2013.
no. 3, pp. 401–405, 2012. [53] E. O. Oktem, E. Derle, S. Kibaroglu, C. Oktem, I. Akkoyun,
[37] M. M. Moschos, G. Tagaris, I. Markopoulos et al., “Morphologic and U. Can, “The relationship between the degree of cognitive
changes and functional retinal impairment in patients with impairment and retinal nerve fiber layer thickness,” Neurologi-
Parkinson disease without visual loss,” European Journal of cal Sciences, 2015.
Ophthalmology, vol. 21, no. 1, pp. 24–29, 2011. [54] F. J. Ascaso, N. Cruz, P. J. Modrego et al., “Retinal alterations in
[38] P. Albrecht, A.-K. Müller, M. Südmeyer et al., “Optical coher- mild cognitive impairment and Alzheimer’s disease: an optical
ence tomography in Parkinsonian syndromes,” PLoS ONE, vol. coherence tomography study,” Journal of Neurology, vol. 261, no.
7, no. 4, Article ID e34891, 2012. 8, pp. 1522–1530, 2014.
[39] C. R. Adam, E. Shrier, Y. Ding, S. Glazman, and I. Bodis- [55] M. V. Sofroniew and H. V. Vinters, “Astrocytes: biology and
Wollner, “Correlation of inner retinal thickness evaluated pathology,” Acta Neuropathologica, vol. 119, no. 1, pp. 7–35, 2010.
by spectral-domain optical coherence tomography and con- [56] W. S. T. Griffin, L. Liu, Y. Li, R. E. Mrak, and S. W. Barger,
trast sensitivity in Parkinson disease,” Journal of Neuro- “Interleukin-1 mediates Alzheimer and Lewy body pathologies,”
Ophthalmology, vol. 33, no. 2, pp. 137–142, 2013. Journal of Neuroinflammation, vol. 3, article 5, 2006.
[40] J. N. Ratchford, M. E. Quigg, A. Conger et al., “Optical [57] N. Cuenca, L. Fernández-Sánchez, L. Campello et al., “Cellular
coherence tomography helps differentiate neuromyelitis optica responses following retinal injuries and therapeutic approaches
and MS optic neuropathies,” Neurology, vol. 73, no. 4, pp. 302– for neurodegenerative diseases,” Progress in Retinal and Eye
308, 2009. Research, vol. 43, pp. 17–75, 2014.
[41] A. Martinez, N. Proupim, and M. Sanchez, “Retinal nerve fibre [58] M. Karlstetter, S. Ebert, and T. Langmann, “Microglia in the
layer thickness measurements using optical coherence tomog- healthy and degenerating retina: insights from novel mouse
raphy in migraine patients,” British Journal of Ophthalmology, models,” Immunobiology, vol. 215, no. 9-10, pp. 685–691, 2010.
vol. 92, no. 8, pp. 1069–1075, 2008. [59] T. Langmann, “Microglia activation in retinal degeneration,”
[42] C. S. Tsai, R. Ritch, B. Schwartz et al., “Optic nerve head and Journal of Leukocyte Biology, vol. 81, no. 6, pp. 1345–1351, 2007.
nerve fiber layer in Alzheimer’s disease,” Archives of Ophthal-
mology, vol. 109, no. 2, pp. 199–204, 1991.
[43] V. Parisi, R. Restuccia, F. Fattapposta, C. Mina, M. G. Bucci, and
F. Pierelli, “Morphological and functional retinal impairment in
Alzheimer’s disease patients,” Clinical Neurophysiology, vol. 112,
no. 10, pp. 1860–1867, 2001.
[44] Y. Chi, Y.-H. Wang, and L. Yang, “The investigation of retinal
nerve fiber loss in Alzheimer’s disease,” Chinese Journal of
Ophthalmology, vol. 46, no. 2, pp. 134–139, 2010.
[45] Y. Lu, Z. Li, X. Zhang et al., “Retinal nerve fiber layer structure
abnormalities in early Alzheimer’s disease: evidence in optical
coherence tomography,” Neuroscience Letters, vol. 480, no. 1, pp.
69–72, 2010.
[46] H. V. Danesh-Meyer, H. Birch, J. Y.-F. Ku, S. Carroll, and
G. Gamble, “Reduction of optic nerve fibers in patients with
Alzheimer disease identified by laser imaging,” Neurology, vol.
67, no. 10, pp. 1852–1854, 2006.
[47] N. Miller and D. A. Drachman, “The optic nerve: a window into
diseases of the brain?” Neurology, vol. 67, no. 10, pp. 1742–1743,
2006.
 MEDIATORS of

INFLAMMATION

The Scientific Gastroenterology Journal of

World Journal
Hindawi Publishing Corporation
Research and Practice
Hindawi Publishing Corporation
Hindawi Publishing Corporation
Diabetes Research
Hindawi Publishing Corporation
Disease Markers
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

Journal of International Journal of

Immunology Research
Hindawi Publishing Corporation
Endocrinology
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

Submit your manuscripts at

http://www.hindawi.com

BioMed
PPAR Research
Hindawi Publishing Corporation
Research International
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

Journal of
Obesity

Evidence-Based
Journal of Stem Cells Complementary and Journal of
Ophthalmology
Hindawi Publishing Corporation
International
Hindawi Publishing Corporation
Alternative Medicine
Hindawi Publishing Corporation Hindawi Publishing Corporation
Oncology
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

Parkinson’s
Disease

Computational and
Mathematical Methods
in Medicine
Behavioural
Neurology
AIDS
Research and Treatment
Oxidative Medicine and
Cellular Longevity
Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014
 Vol. 9, No. 11 | 1 Nov 2018 | BIOMEDICAL OPTICS EXPRESS 5759

Automatic segmentation of OCT retinal

boundaries using recurrent neural networks
and graph search
JASON KUGELMAN, DAVID ALONSO-CANEIRO,* SCOTT A. READ, STEPHEN
J. VINCENT, AND MICHAEL J. COLLINS
Contact Lens and Visual Optics Laboratory, School of Optometry and Vision Science, Queensland
University of Technology, Brisbane, Queensland, Australia
*[email protected]

Abstract: The manual segmentation of individual retinal layers within optical coherence
tomography (OCT) images is a time-consuming task and is prone to errors. The investigation
into automatic segmentation methods that are both efficient and accurate has seen a variety of
methods proposed. In particular, recent machine learning approaches have focused on the use
of convolutional neural networks (CNNs). Traditionally applied to sequential data, recurrent
neural networks (RNNs) have recently demonstrated success in the area of image analysis,
primarily due to their usefulness to extract temporal features from sequences of images or
volumetric data. However, their potential use in OCT retinal layer segmentation has not
previously been reported, and their direct application for extracting spatial features from
individual 2D images has been limited. This paper proposes the use of a recurrent neural
network trained as a patch-based image classifier (retinal boundary classifier) with a graph
search (RNN-GS) to segment seven retinal layer boundaries in OCT images from healthy
children and three retinal layer boundaries in OCT images from patients with age-related
macular degeneration (AMD). The optimal architecture configuration to maximize
classification performance is explored. The results demonstrate that a RNN is a viable
alternative to a CNN for image classification tasks in the case where the images exhibit a
clear sequential structure. Compared to a CNN, the RNN showed a slightly superior average
generalization classification accuracy. Secondly, in terms of segmentation, the RNN-GS
performed competitively against a previously proposed CNN based method (CNN-GS) with
respect to both accuracy and consistency. These findings apply to both normal and AMD
data. Overall, the RNN-GS method yielded superior mean absolute errors in terms of the
boundary position with an average error of 0.53 pixels (normal) and 1.17 pixels (AMD). The
methodology and results described in this paper may assist the future investigation of
techniques within the area of OCT retinal segmentation and highlight the potential of RNN
methods for OCT image analysis.

© 2018 Optical Society of America under the terms of the OSA Open Access Publishing Agreement

1. Introduction
Optical coherence tomography (OCT) is a non-invasive imaging technique that allows high-
resolution cross-sectional imaging of ocular tissues such as the retina [1–3]. Retinal OCT
imaging has been used extensively for characterizing the normal retina and its individual
layers and in the detection and monitoring of ocular diseases such as age-related macular
degeneration (AMD) [4–6], glaucoma [7], and diabetic retinopathy [8] which present changes
in the normal retinal layer topology. However, the manual analysis of OCT images to extract
the boundary position and subsequent structural characteristics of the retinal layers is time
consuming, subjective, and prone to error, hence the need for efficient and automatic
segmentation techniques.

#345109 https://doi.org/10.1364/BOE.9.005759
Journal © 2018 Received 5 Sep 2018; revised 13 Oct 2018; accepted 15 Oct 2018; published 26 Oct 2018
 Vol. 9, No. 11 | 1 Nov 2018 | BIOMEDICAL OPTICS EXPRESS 5760

There has been extensive previous work in the area of automatic OCT image analysis, and
the segmentation of retinal layers has been a significant focus, with a range of approaches
presented including: Markov boundary models [9], sparse higher order potentials [10],
diffusion filtering [11], diffusion maps [12], gradient maps [13], Chan-Vese (C-V) models
[14], kernel regression (KR)-based classification [15], and graph-based methods [13,16,17].
The drawback with a number of these methods is their reliance on specific “ad-hoc” rules,
resulting in poor performance in the presence of noise and/or artifacts. Recent work has
attempted to address these issues by utilizing machine learning techniques for OCT image
segmentation. These recent studies have used a number of approaches including support
vector machines [18], neural networks [19–30] and random forests [31]. Fang et al [20]
utilized a patch-based convolutional neural network and graph search based approach (CNN-
GS) to segment nine retinal layers in OCT images of patients with non-exudative age-related
macular degeneration. Extending upon this, Hamwood et al [21], investigated the effect of
patch size and network complexity on the overall performance and proposed an improved
CNN architecture for analysis of OCT images of healthy eyes. Loo et al [22], used a novel
deep learning approach (DOCTAD) that combined convolutional neural networks and
transfer learning to perform the segmentation of photoreceptor ellipsoid zone defects in OCT
images. ReLayNet, proposed by Roy et al [23], utilized a fully-convolutional neural network
architecture to perform semantic segmentation of retinal layers and intraretinal fluid in
macular OCT images. Fully-convolutional networks were also utilized by Xu et al [24], in a
dual-stage deep learning framework for retinal pigment epithelium detachment segmentation.
The recurrent neural network (RNN) has been used for a variety of problems, particularly
where sequential data is present. Some examples include language modelling [32], machine
translation [33], handwriting recognition/generation [34], sequence generation [35], and
speech recognition [36,37]. The effectiveness of RNNs when applied to image-based tasks
has also been investigated. Visin et al. [38] compared the image classification ability of an
RNN to that of a CNN and presented a novel architecture (ReNet) specifically designed to
handle 2D images as the input. One layer of the ReNet architecture contained four RNNs,
each of which was designed to sweep over the image in a different direction. Performance of
the ReNet architecture for evaluating the MNIST (28x28 grayscale), CIFAR-10 (32x32
RGB), and SVHN (32x32 RGB) image data sets was reported to be similar to some CNNs.
However, their architecture did not outperform state-of-the-art CNNs on any of the data sets.
In a more recent paper [39], the ReNet architecture was explored further and coupled with a
CNN for use in end-to-end image segmentation, reporting state-of-the-art performance on the
Weizmann Horse (variable sized grayscale and RGB), Oxford Flowers 17 (variable sized
RGB), and CamVid (960x720 RGB) data sets. In particular, the authors cited the usefulness
of RNNs to capture global dependencies. Earlier work by Graves et al [40], also explored the
application of RNNs to images, presenting a multi-dimensional RNN (MD-RNN) in which
they replaced the single recurrent connection with one for every dimension in the input.
Likewise, another variant of the RNN, the convolutional LSTM (C-LSTM), was introduced
by Shi et al [41], in their work with short-term precipitation forecasting.
Recently, several papers have examined the use of RNNs for medical image analysis.
Chen et al [42], combined a fully-convolutional network (FCN) approach with a recurrent
neural network for 3D biomedical image segmentation in which the images were comprised
of a sequence of 2D slices. The FCN was used to extract intra-slice contexts into feature
maps, which were then passed to the RNN to extract the inter-slice contexts. In their work on
automatic fetal ultrasound standard plane detection, Chen et al [43], used a knowledge
transferred network that first extracted spatial features using a CNN which were then explored
temporally using a RNN. A similar methodology was employed by Kong et al [44], using a
temporal regression network (TempReg-Net) for end-systole and end-diastole frame
identification in MRI cardiac sequences. Stollenga et al [45], introduced their PyraMid-LSTM
and both highlighted the effectiveness of RNNs for spatiotemporal tasks and reported state-
 Vol. 9, No. 11 | 1 Nov 2018 | BIOMEDICAL OPTICS EXPRESS 5761

of-the-art performance on a data set for brain MRI segmentation. Investigating dynamic
cardiac MRI reconstruction, Qin et al [46], used a novel convolutional recurrent neural
network which outperformed existing techniques in terms of both speed and accuracy and
also required fewer hyper parameters. In another paper, Xie et al [47], attempted to use a
clockwork recurrent neural network (CW-RNN) [48] based architecture to accurately segment
muscle perimysium.
While a RNN approach has proven successful in the analysis of a range of medical
images, to the best of our knowledge, there is no previous work utilizing RNNs as an
approach to segment retinal boundaries in OCT images. There is also little evidence of RNNs
being applied directly to individual medical images to extract spatial features. Instead, they
have been used predominantly to extract temporal features from sequences of feature maps,
with convolutional neural networks (CNNs) preferred to operate spatially on each image as a
prior step. In this work, a novel recurrent neural network combined with a graph search
approach (RNN-GS) is presented. This combines patch-based boundary classification using
RNNs with a subsequent graph search to delineate retinal layer boundaries. This approach is
partly inspired by the work of Fang et al [20], but in the methodology presented here the CNN
is replaced with a RNN. A detailed selection of the RNN architecture and configuration as
well as the evaluation of the optimal RNN model is presented.
The paper is structured as follows. In Section 2, the RNN-GS methodology and approach
is outlined, including details about the data sets used as well as the RNN model and
architecture selection. Section 3 presents experimental classification results for a range of
RNN architectures which were used to inform the empirical selection of a suitable RNN
design. Section 4 presents the segmentation results for the selected RNN design with
performance evaluated against other CNN based methods. Discussion of the method and
results are provided in Section 5. Concluding remarks are included in Section 6.
2. Method
2.1 Data set 1 (normal OCT images)
The first data set (data set 1) used in this work consists of a range of OCT retinal images from
a longitudinal study that has been described in detail in a number of previous publications
[21,49–51], The data comprises OCT retinal scans for 101 children taken at four different
visits over an 18-month period. All subjects had normal vision in both eyes and no history of
ocular pathology. The images were acquired using the Heidelberg Spectralis (Heidelberg
Engineering, Heidelberg, Germany) SD-OCT instrument. At each visit, each subject had two
sets of six foveal centered radial retinal scans taken. The instrument’s Enhanced Depth
Imaging mode was used and automatic real time tracking was also utilized to improve the
signal to noise ratio by averaging 30 frames for each image. The acquired images each
measure 1536x496 pixels (width x height). With a vertical scale of 3.9 µm per pixel and a
horizontal scale of 5.7 µm per pixel, this corresponds to an approximate physical area of size
8.8x1.9 mm (width x height). These images were exported and analyzed using custom
software where an automated graph based method [13,52], was used to segment seven retinal
layer boundaries for each image. This segmented data was then assessed by an expert human
observer who manually corrected any segmentation errors. Throughout this paper, “B-scan”
refers to an individual full-size (1536x496) image while “A-scan” corresponds to a single
column of a B-scan.
The seven layer boundaries within the labelled data include the outer boundary of the
retinal pigment epithelium (RPE), the inner boundary of the inner segment ellipsoid zone
(ISe), the inner boundary of the external limiting membrane (ELM), the boundary between
the outer plexiform layer and inner nuclear layer (OPL/INL), the boundary between the inner
nuclear layer and the inner plexiform layer (INL/IPL), the boundary between the ganglion cell
layer and the nerve fiber layer (GCL/NFL) and the inner boundary of the inner limiting
membrane (ILM).
 Vol. 9, No. 11 | 1 Nov 2018 | BIOMEDICAL OPTICS EXPRESS 5762

For compu utational reaso

ons, only a porrtion of all thee described datta was used thhroughout
this work. Twwo subsets of data
d are defined d here. For neuural network trraining and vallidation, a
set of images (labelled dataa A) was selectted. For traininng, 842 B-scanns were used cconsisting
of all availablle scans from both
b sets from
m 39 randomly selected particcipants’ first tw wo visits.
For validationn, 223 B-scans were used con nsisting of all aavailable scanss from both setss from 20
different rand
domly selected participants’ third
t visit. A ssecond set of immages (labelled data B)
was selected for evaluation of the method d. This set conssisted of 115 iimages comprised of all
available scanns from one seet from 20 rand domly selectedd participants’ fourth visit. TThere was
no overlap off participants or
o visits betweeen the trainingg, validation an
and evaluation sets with
the intention of
o obtaining an n accurate reprresentation of tthe network geeneralizability tto unseen
data.
2.2 Data sett 2 (AMD OCT
T images)
The second datad set (data set 2) [53] consists
c of a range of OCT retinal images from
individuals with
w and withou ut age-related macular degenneration (AMD D). Within thiss, 100 B-
scans each froom 269 AMD participants
p an
nd 115 healthy pparticipants arre provided. Foor each B-
scan, segmenttation of three retinal layer boundaries
b aree provided (from an automateed graph-
search metho od that was manually
m correected by expeert image graaders). The three layer
boundaries within
w the labelled data includde the inner asspect of the innner limiting m
membrane
(ILM), the inner aspect of thet retinal pigm ment epitheliuum drusen com mplex (RPEDC C) and the
outer aspect of
o Bruch’s mem mbrane (BM).
Like data set 1, not all available
a data was
w used in thiis work due to computationall reasons.
Two subsetss of data aree again defin ned here witth ‘labelled data A’ for network
training/validation and ‘lab belled data B’B for evaluattion. Only im mages from thhe AMD
individuals were
w used from this data set. Labelled
L data A comprised of 15 B-scans (scans 40-
54) each fromm 160 participaants (participan nts 1-160) (2,4400 B-scans tottal). From thiss, the first
120 participan nts (participants 1-120) weree used for trainning (1,800 B--scans) and thee final 40
participants (pparticipants 12
21-160) were used
u for validaation (600 B-sscans). Labelleed data B
consisted of 15
1 scans (scan ns 40-54) each h from the nextxt 20 participannts not used inn labelled
data A (particcipants 161-180 0) (300 B-scanns total). Theree was no overlaap between datta used in
labelled data A and B or bettween the train ning and validattion sets.

Fig. 1. Overview of thee RNN-GS method d for segmentationn of retinal layerss, where red boxess
repressent training stepss and blue are ev valuation steps. F
For training (labeelled data A) andd
evaluaation (labelled dataa B) there was no overlap between pparticipants.
 Vol. 9, No. 11 | 1 Nov 2018 | BIOMEDICAL OPTICS EXPRESS 5763

2.3 Overview
w
An overview of the recurren nt neural network and graph search methodd (RNN-GS) uutilized in
this paper is presented
p in Figg. 1. This meth
hod is similar too that used by Fang et al [20]] but here
a RNN is used d in place of th
he CNN.
Further deetails of the reccurrent neural network modeel are providedd in Section 2.44. For the
method, theree are two main phases: trainin ng and evaluatiion. In the trainning phase, thee labelled
OCT data (lab belled data A),, is used to connstruct a dictioonary of imagee patches for eaach class.
The patches are
a centered up pon the layer booundaries of innterest. These ppatches are theen used to
train the RNNN as a classifierr as detailed in
n Section 2.5. Inn the second phase, evaluatioon of new
data (labelledd data B) is performed. The OCT images aare split into ppatches for eveery single
pixel of the immage. The train ned neural nettwork is then uused to predict the class of eaach patch
and generate per class prob bability maps for
f the entire iimage. The layyer boundariess are then
delineated by performing a shortest-path
s graph
g search ass outlined in Seection 2.6. Thee methods
and models used
u for compaarison are desccribed in Sectioon 2.7. Finallyy, the overall eevaluation
of the method d is described inn Section 2.8.
2.4 Recurren
nt neural netw
work model
Recurrent neu ural networks (RNNs)
( differ from other neetwork configuurations, such aas CNNs,
with the addittion of feedbacck from the outtput to the inpuut. This enablees the network to utilize
a past sequen nce of inputs to inform futture outputs, which ultimattely gives the network
memory. RNN Ns may be traained using an extension of tthe backpropaggation algorithhm called
backpropagatiion-through-tim me (BPTT) [54 4–56]. In theirr basic form, RRNNs are com mprised of
recurrent units which pass a concatenation n of the previouus output and ccurrent input thhrough an
activation funnction. Howev ver, these sim
mple recurrent units have issues in practiice. Most
notably are the
t problems of vanishing and explodingg gradients [557–59]. More complex
recurrent unitts have been deesigned and prooposed to deal with this, nammely the long-sshort term
memory (LST TM) [57], and gated
g recurrentt unit (GRU) [660].

Fig. 2.
2 Example of a mo odel with three staacked RNN layers showing how the activation volumee
is man nipulated as it passses through the neetwork. Each greyy volume correspoonds to the volumee
(recep
ptive field width x receptive field heeight x channels) pprocessed at a parrticular step withinn
the firrst sequence operaated on by the RN NN. The direction of this operation iis indicated by thee
solid arrows.
a The dottedd outline volumes belong to each steep of the followingg sequence and thee
dashed arrows indicate the order the sequ uences are processsed in. 1) Horizonntal unidirectionall
RNN with a 2x2 recep ptive field and 8 filters. 2) Horizonntal bidirectional RNN with a 1x11
recepttive field and 16 fiilters (8 / pass). 3) Vertical bidirectioonal RNN with a 22x2 receptive fieldd
and 24 4 filters (12 / pass)).

The layerrs within the ReNet architeecture [38], usse a set of foour RNNs opeerating in
different direcctions to perfo
orm the task off image classiffication. Each RNN layer can an operate
either horizon
ntally or verticaally over sequeences of imagee pixels and thiis can be done with one
pass (unidireectional), or two
t passes (bbidirectional). A number oof parameters are also
 Vol. 9, No. 11 | 1 Nov 2018 | BIOMEDICAL OPTICS EXPRESS 5764

associated with each layer including the receptive field size (width x height in pixels), number
of filters, and type of recurrent unit. Each additional filter can give the network an opportunity
to learn a different pattern. In terms of the recurrent unit types, both the LSTM and GRU are
considered as options within each layer. The receptive field corresponds to the volume that is
processed by the RNN at each step in each sequence. For each RNN, the spatial dimensions
(width and height) of the output volume are equal to the respective sequence lengths in each
direction. These sequence lengths are equal to the spatial dimensions of the input volume
divided by the corresponding receptive field dimensions. Meanwhile, the depth (number of
channels) of the output volume is simply equal to the number of filters.
For example: an 8 filter horizontal unidirectional RNN operating with an input volume
size of 16x16x3 (width x height x channels) and a receptive field size of 2x2, would mean
that a volume of size 2x2x3 is processed by the RNN for each of 8 steps (horizontally), for
each of 8 sequences (vertically), with an overall output volume size of 8x8x8. A visualization
of this example is illustrated in step 1 of Fig. 2. For bidirectional layers, the output volumes of
each pass are concatenated together along the depth dimension as depicted in steps 2 and 3.
To avoid overfitting to the training data, the dropout regularization technique was utilized
[61]. Here, each layer is equipped with a level of dropout which corresponds to the percentage
of units in that layer which are randomly turned off (dropped) in each epoch. Batch
normalization [62], at the input to each layer was also used. This ensures that the mean and
variance of each mini-batch is scaled to 0 and 1 respectively, which can help to improve the
performance and stability of the network during training.
2.5 Training and patch classification
The RNN model is trained as a classifier. This is done by constructing small overlapping
patches from the OCT images and assigning each to a class based on the layer boundary that
they are centered upon. These constitute the “positive” training examples. For patches not
centered upon a layer boundary, Fang et al [20], utilized a single class for “negative” training
examples, also called the “background” class. In this study, two background classes were
used with the intention to better capture the different background features of the OCT image,
particularly some of the features and image artefacts in the retina, vitreous (anterior to the
retina) and in the choroid and scleral region (posterior to the retina). The first background
class consists of patches centered within the retina (between the ILM and RPE for data set 1
and between the ILM and BM for data set 2) as well as in a small region of both the vitreous
and choroid directly above and below the retina respectively. The height of these smaller
regions is set to be equal to the patch height. All patches within the described area that are not
centered on any boundary are considered part of this class. The second background class
consists of patches centered in a region bounded between the bottom of the first background
class region and the bottom of the image. Zero-padding is added to any patches at the edge of
images where required.
Using the ‘labelled data A’ images as described in Section 2.1 and 2.2, patches were
created for both training and validation (for both data set 1 and data set 2). Patches were
created for each class with background examples randomly selected within their
corresponding ranges. To reduce computational burden, the total number of patches was
restricted with patches only created for every eighth column of each image. For data set 1, the
training set was comprised of ~1,450,000 patches with an additional ~380,000 for validation.
This was a nine-class classification problem with equally weighted classes. Similarly, for data
set 2, the training set was comprised of ~980,000 patches with an additional ~320,000 for
validation. This was a five-class classification problem with equally weighted classes. In an
effort to maximize training performance, all patches were normalized (0-1) before they were
input to the network. However, unlike previous studies [13,20,21], intensity normalization
and other image pre-processing steps were not used in this work for any of the data sets.
 Vol. 9, No. 11 | 1 Nov 2018 | BIOMEDICAL OPTICS EXPRESS 5765

The Adam algorithm [63], a stochastic gradient-based optimizer, was used to train the
network by minimizing log loss (cross-entropy) [64]. Empirically, Adam performs well in
practice with little to no parameter adjustment and compares favorably to other stochastic
optimization methods [63]. Due to Adam’s relatively quick convergence to good solutions in
a small number of epochs, no early stopping was used for training. In addition, given the
adaptive per-parameter learning rates that this optimizer possesses, no learning rate scheme
was deemed to be necessary. The network was trained for 50 epochs with a batch size of 1024
and the model that yielded minimum validation loss was selected. This is similar to
approaches described and used elsewhere [65,66]. The number of epochs and the batch size
were chosen empirically, while the algorithm parameters were left at their recommended
default values. The Keras API [67], with Tensorflow [68], backend in Python was the
programming environment of choice.
2.6 Probability maps and graph search
For a single OCT test image, patches are generated for every pixel and passed to the trained
neural network to be classified. From this, per class probability maps can be constructed and a
graph search performed to delineate the layer boundaries. The idea was proposed by Chiu et
al [13], for the segmentation of OCT images which has been adapted in a number of CNN
studies [20,21]. However, in contrast to previous work, in this study the search path was not
limited between the top and bottom layer boundaries. Each probability map can be used to
construct a directed graph where the pixels in the map correspond to vertices in the graph.
Each vertex is connected to its three rightmost neighbors (diagonally above, horizontally,
diagonally below). The weights of these connections are given by the equation:
wsd = 2 − ( Ps + Pd ) +wmin , (1)

where Ps and Pd are the probabilities (0-1) of the source and destination vertices respectively,
and wmin = 1x10−5 is a small positive number added for system stability.
To automate the start and end point initialization, a column of maximum intensity pixels is
appended to both the left and right of the image. As well as being connected to their rightmost
neighbors, vertices in these columns are also connected vertically from top to bottom. This
allows for a graph search algorithm, like Dijkstra’s shortest-path algorithm [69], as used here,
to start at the top-left corner and traverse the graph through to the bottom-right corner without
any manual interaction. In this way, a graph cut is performed and this shortest path is used as
the predicted location of the layer boundary.
2.7 Comparison of methods
A comparison between the RNN-based method and a patch-based CNN method is presented.
The CNN method used is identical except that the RNN is replaced by a CNN. The CNN used
here is the so-called “complex CNN” proposed by Hamwood et al [21]. This is trained
identically to the RNN as described in Section 2.5.
In addition, a comparison between the patch-based method and a full image-based method
is also provided. The method for comparison is a fully-convolutional network and graph
search method (FCN-GS). For this, the patch-based classifier network is replaced with a U-
Net [70], style architecture similar to that used by Venhuizen et al [26]. The FCN used here
consists of four down sampling blocks each with two 3x3 convolutional layers. The network
was trained for 50 epochs with a batch size of three using Adam with default parameters in a
similar way to the patch-based networks in Section 2.5. Cross-entropy loss is used here to
classify each pixel of the image into one of eight area classes. These eight areas are
constructed between adjacent layer boundaries and the top and bottom of the image as
required to create an overall area mask. For A-scans where at least one layer boundary is not
defined, the image is zeroed with the corresponding columns in the area mask set to be
defined as the top-most region. The overall method is similar to that used by Ben-Cohen et al
 Vol. 9, No. 11 | 1 Nov 2018 | BIOMEDICAL OPTICS EXPRESS 5766

[25], where the FCN is used for semantic segmentation on whole OCT images. Instead of
classifying patches to generate probability maps, the Sobel filter is applied to the area
probability maps output from the FCN to extract the boundary probability maps. A shortest-
path graph search is then performed using these boundary probability maps in the same way
as the patch-based method.
2.8 Evaluation
As described in Section 2.1 for data set 1 and Section 2.2 for data set 2, the images contained
in labelled data B were used to evaluate the whole method. By comparing the predicted
boundary positions to the truth (the segmentation from the expert human observer), the mean
error and mean absolute error with their associated standard deviations were calculated for
each layer across the whole test set. For data set 1, the full-width image was used for both
patch creation and performing the graph search. However, due to the lack of consistency of
the layers around the left and right extremities of the image (e.g. presence of optic nerve head
and shadows), the first and last 100 pixels of each side were excluded from the final error
calculations and comparisons. For data set 2, the full-width image was used as input to the
network with a full-width probability map used for the graph search. However, as the true
layer boundaries were not defined in every column, only those columns with all true boundary
locations present were used for error calculations and comparisons.
3. RNN design
In order to design a suitable RNN architecture for patch classification, the impact of various
network parameters on performance was examined. This section presents the results for a
range of experiments including; the impact of patch size and direction of operation (Section
3.1), receptive field size (Section 3.2), number of filters (Section 3.3), stacking and ordering
of layers (Section 3.4), and fully-connected layers (Section 3.5).
Visin et al. [38], used dropout after each layer in their network while Srivastava et al. [61],
show that 50% dropout is a sensible choice for a variety of tasks. As such, 50% dropout is
added after each fully-connected layer in Section 3.5. However, due to the relatively small
number of parameters in the RNN layers used here, this level of dropout was deemed to be
unnecessary and a potential hindrance to performance. Instead, 25% dropout is used after
each RNN layer to ensure sufficient network capacity.
All networks were trained identically as described in Section 2.5. The experiments are
performed only utilising data set 1 to allow for a fair comparison with Hamwood et al. [21],
and their CNN which was also tested and optimized using normal OCT images. The
generalizability of each network is assessed by measuring the classification accuracy on the
validation set (described as part of labelled data A in Section 2.1). This is computed by
dividing the number of correctly classified patches by the total number of patches. Due to
randomness associated with both the network weight initialization and batch ordering leading
to possibly different solutions, each experiment was performed three times and the results
were averaged. These experiments were used to inform the careful selection of the most
suitable final RNN architecture that was employed, which is described in Section 3.6.
3.1 Patch size and direction
For the design of the RNN, it is of interest to investigate the effect of the patch size (height x
width pixels) on the network performance. In their CNN-GS approach, Fang et al [20], used a
33x33 patch size centered on the layer boundaries, while Hamwood et al [21], showed that
increasing the size can improve network performance. With this in mind, 32x32, 64x32,
32x64 and 64x64 patches are compared on a range of RNN architectures. Even-sized
dimensions are chosen to facilitate the network model and to avoid additional zero padding.
Because of the even-size, the patch cannot be truly centered, and therefore each is
 Vol. 9, No. 11 | 1 Nov 2018 | BIOMEDICAL OPTICS EXPRESS 5767

consistently placed with the layer boundary positioned on the pixel above and to the left of
the central point.
Table 1 outlines the results of the experiments undertaken. A small but significant
improvement in classification accuracy was observed when using a vertically oriented 64x32
patch (longer along the A-scan direction) compared to a 32x32 (about 1.1% mean
improvement). However, this level of improvement is much less pronounced when comparing
the 32x32 with the horizontally oriented 32x64 patch size (about 0.4% mean improvement).
Despite possessing twice as many pixels, the 64x64 patch does not exhibit a clear
performance benefit compared to the 64x32 patch (below 0.1% mean improvement). Thus,
the 64x32 patch size appears to yield the best trade-off between accuracy and complexity for
the tested sizes. It should be noted that other patch sizes were not tested for computational
reasons.
Table 1. Effect of patch size and direction on validation classification accuracy (%). The
mean (standard deviation) of the accuracy for three training runs (GRU, 32 filters / pass,
2x2 receptive field, 25% dropout).

Patch Size (height x width pixels)

Architecture 32x32 64x32 32x64 64x64 Mean
Vertical Unidirectional 95.17 (0.05) 96.28 (0.02) 95.44 (0.06) 96.26 (0.07) 95.78
Horizontal Unidirectional 94.48 (0.03) 95.77 (0.04) 95.13 (0.06) 96.10 (0.02) 95.37
Vertical Bidirectional 95.33 (0.06) 96.18 (0.05) 95.51 (0.02) 96.01 (0.12) 95.75
Horizontal Bidirectional 94.55 (0.08) 95.71 (0.09) 95.19 (0.06) 95.95 (0.15) 95.35
Mean 94.88 95.98 95.31 96.08

Within the ReNet layers [38], RNNs were used separately to process the input
horizontally or vertically. To better understand the impact that the direction of operation has
on network performance, these different options were considered. As shown in Table 1, the
direction of operation appears to have a small impact on the classification accuracy, although
it is worth noting that RNNs operating in the vertical direction outperform their horizontal
counterparts by a small percentage. However, operating bi-directionally does not appear to
yield improved performance.
Table 2. Effect of receptive field size on validation classification accuracy (%). The mean
(standard deviation) of the accuracy for three training runs are reported. (GRU, 32
filters, 64x32 patch size, single-layer vertical unidirectional RNN, 25% dropout).

Square Rectangular
Size Acc. (%) Size Acc. (%) Size Acc. (%)
1x1 96.12 (0.05) 1x4 96.10 (0.12) 8x1 96.37 (0.03)
2x2 96.28 (0.02) 4x1 96.28 (0.01) 2x8 96.08 (0.08)
4x4 96.29 (0.06) 2x4 96.24 (0.05) 8x2 96.44 (0.02)
8x8 96.13 (0.10) 4x2 96.35 (0.03) 4x8 96.13 (0.08)
16x16 95.82 (0.06) 1x8 96.09 (0.05) 8x4 96.38 (0.03)

3.2 Receptive field size

The effect of the receptive field size on the network performance was also investigated. Visin
et al [38], used a receptive field size of 2x2 between each of their ReNet layers. Here, a
variety of square and rectangular receptive field sizes were compared on a single-layer
vertical unidirectional RNN with the results outlined in Table 2. Similar to the effect of patch
size described in Section 3.1, the vertical rectangular receptive fields provide a marginal
improvement in performance compared to the equivalent horizontal variants, attributable to
the vertical nature of the layer structure in the image. Overall, most of the tested sizes give
similar performance indicating that the size of the receptive field does not have a significant
impact on the accuracy for the tested data set.
 Vol. 9, No. 11 | 1 Nov 2018 | BIOMEDICAL OPTICS EXPRESS 5768

3.3 Number of filters

Increasing the number of filters gives the neural network more parameters and hence more
opportunity to learn. The change in classification accuracy, as the number of filters in a single
layer vertical unidirectional RNN is varied, was investigated to better estimate the optimal
number of filters and the impact on performance. Table 3 shows that adding more filters
yields a small increase in classification accuracy, albeit with diminishing returns. For this
single layer network, choosing 32 filters gives a good trade-off between accuracy and
complexity.
Table 3. Effect of number of filters on validation classification accuracy (%). The mean
(standard deviation) of the accuracy for three training runs are reported. (GRU, 2x2
receptive field, 64x32 patch size, single-layer vertical unidirectional RNN, 25% dropout).

Filters Network parameters Acc. (%)

8 37,273 95.62 (0.08)
16 74,857 96.05 (0.04)
32 151,177 96.28 (0.02)
64 308,425 96.32 (0.02)

3.4 Stacked layers and order

The ReNet architecture [38], uses several layers of RNNs, each of which first operate on the
input vertically before horizontally. Here, the effect of adding additional layers to the network
as well as the order that these are stacked together was evaluated. The results presented in
Table 4 indicate that stacking layers improves the classification accuracy. Further, stacking
both horizontal and vertical RNN layers yields greater performance than solely using vertical
ones. There is no noticeable performance difference when changing the stacking order. This is
also the case when using bi-directional RNNs, reinforcing the results presented in Section 3.1.
Table 4. Effect of stacked layers and order on validation classification accuracy (%). The
mean (standard deviation) of the accuracy for three training runs are reported. (GRU,
2x2 receptive field, 32 filters / pass, 64x32 patch size, 25% dropout each layer).

Architecture Acc. (%) Architecture Acc. (%)

Vertical Unidirectional 96.28 (0.02) Horizontal Unidirectional + 96.69 (0.02)
Vertical Unidirectional
2 x Vertical Unidirectional 96.59 (0.02) Vertical Unidirectional + 96.69 (0.02)
Horizontal Unidirectional
3 x Vertical Unidirectional 96.67 (0.01) Horizontal Bidirectional + 96.70 (0.01)
Vertical Bidirectional
Vertical Bidirectional + 96.73 (0.02)
Horizontal Bidirectional

3.5 Fully-connected layers

Visin et al. [38], used one or more fully-connected (FC) output layers of size 4096 in their
ReNet architecture. The effect of including a fully-connected layer in our network design was
also evaluated. The results presented in Table 5 show that adding a fully-connected layer has
little benefit given the corresponding drastic increase in network parameters.
Table 5. Effect of fully-connected output layer size on validation classification accuracy
(%). The mean (standard deviation) of the accuracy for three training runs are reported.
(GRU, 2x2 receptive field, 32 filters, 64x32 patch size, single-layer vertical unidirectional
RNN, 25% dropout for RNN layer, 50% dropout for fully-connected layer)

FC layer size Network parameters Acc. (%)

0 151,177 96.28 (0.02)
32 528,329 96.37 (0.04)
64 1,052,937 96.39 (0.04)
128 2,102,153 96.32 (0.03)
 Vol. 9, No. 11 | 1 Nov 2018 | BIOMEDICAL OPTICS EXPRESS 5769

3.6 RNN architecture selection

Based on the experimental findings presented in Sections 3.1-3.5, a RNN architecture was
selected. An overview of this architecture is provided in Table 6. As discussed in Section 3.5,
no fully-connected layers are used due to their seemingly negligible performance benefit.
Two sets of vertical and horizontal bi-directional layers are used, each with a size of 32 filters
(16 per direction) and 25% dropout. Because the classification is based on pixel level
accuracy, the first two layers are equipped with a 1x1 receptive field to enable the network to
initially process the full-sized image on a pixel by pixel basis. The subsequent layers utilize a
2x2 receptive field with the intention of allowing the network to learn context at different
levels. As described in Sections 3.1-3.5, the network operates with gated recurrent units
(GRUs) which were found to perform comparably to LSTMs for this problem.
Table 6. The selected RNN architecture. 4 bidirectional layers are utilized with two
operating vertically and two horizontally. Each layer contains 16 filters per pass for a
total of 32 filters each.

Layer Architecture Receptive Field Filters (/ pass)

Vertical Bidirectional 1x1 16
Horizontal Bidirectional 1x1 16
Vertical Bidirectional 2x2 16
Horizontal Bidirectional 2x2 16

4. Results
4.1 Normal OCT data (data set 1)
Using normal OCT images (data set 1) as described in Section 2.1, the RNN-GS method was
evaluated as described in Section 2 using the RNN architecture selected and trained as
outlined in Section 3. Utilizing a 64x32 patch size, the network yielded a validation
classification accuracy of 96.84% (0.05) taken as the average over three training runs. The
mean accuracy of the seven boundary classes (excluding the background) was 98.25% (0.06)
with the individual per-class accuracies ranging between 96.52% (0.08) (the IPL) and 99.24%
(0.08) (the ILM). With the chosen patch size, the RNN architecture consisted of ~70,000 total
parameters. Using an Nvidia GeForce GTX 1080Ti + Intel Xeon W-2125 the average
evaluation time per B-scan was ~145 seconds. Here, the time to generate the probability maps
was ~105 seconds on average with an average of ~40 seconds to perform the graph search for
all seven boundaries.
The segmentation results for each layer boundary are presented below in terms of the
mean error and the mean absolute error as well as their standard deviations. The patch-based
approach was also evaluated using the Complex CNN architecture as described in Section 2.7
using the same set of 64x32 patches. To support the patch dimensionality, a 13x5 fully-
connected output layer was used. Averaged over three training runs, the CNN provided a
validation classification accuracy of 96.36% (0.04), 0.48% lower than the RNN. The per-class
accuracies ranged between 95.65% (0.85) (the IPL) and 99.17% (0.11) (the ILM) with a mean
accuracy for the seven boundary classes of 97.94% (0.08), 0.58% lower than the RNN. This
CNN architecture consisted of ~1,200,000 total parameters, approximately 17 times as many
as the RNN. Using the same hardware, the average evaluation time per B-scan was
approximately 65 seconds, about 2.2 times faster than the RNN. Given the same time for the
graph search (~40 seconds), this corresponds to ~25 seconds on average to generate the
probability maps which is about 4.2 times faster than the RNN.
The segmentation errors in terms of boundary positions (in pixels) are presented in Table
7. The mean errors (and mean absolute errors) between methods are of similar magnitude,
which suggests that the two networks give a similar level of performance with the RNN based
approach performing marginally better on each boundary with a 0.02 to 0.05 pixels mean
absolute error improvement with the exception of the GCL/NFL (0.12 pixels improvement in
 Vol. 9, No. 11 | 1 Nov 2018 | BIOMEDICAL OPTICS EXPRESS 5770

mean absolute error). This corresponds to an average improvement of 0.05 pixels (mean
absolute error) with the RNN-GS yielding an average of 0.53 pixels mean absolute error on
each boundary compared to the CNN-GS with 0.58. Both RNN-GS and CNN-GS performed
the best on the ISe boundary with 0.33 and 0.35 pixels mean absolute error respectively,
whereas both performed the poorest on the GCL/NFL with respective mean absolute errors of
0.84 and 0.96 pixels. The standard deviations of the errors are also consistently smaller for the
RNN-GS method for each of the considered layers, indicating a greater level of consistency in
the segmentation compared to the CNN-GS approach.
Table 7. Data set 1 (normal OCT images) position error (in pixels) of each layer
boundary for each of the tested methods. The results are reported in mean values and
(per A-scan standard deviation). The best results (smallest error) for each boundary are
highlighted in bold text.

RNN-GS CNN-GS FCN-GS

Layer Mean error Mean abs. Mean error Mean abs. Mean error Mean abs.
Boundary error error error
ILM 0.01 (0.97) 0.46 (0.85) −0.02 (1.01) 0.48 (0.89) −0.16 (0.87) 0.53 (0.71)
GCL/NFL 0.11 (2.22) 0.84 (2.06) -0.06 (2.90) 0.96 (2.74) −0.10 (1.54) 0.79 (1.33)
INL/IPL -0.13 (1.10) 0.56 (0.95) −0.16 (1.17) 0.60 (1.02) −0.26 (1.07) 0.61 (0.92)
OPL/INL -0.10 (1.31) 0.69 (1.12) −0.11 (1.41) 0.73 (1.21) −0.16 (0.96) 0.63 (0.75)
ELM 0.02 (0.75) 0.35 (0.67) 0.09 (0.91) 0.38 (0.83) −0.20 (0.88) 0.42 (0.80)
ISe 0.02 (1.02) 0.33 (0.96) 0.06 (1.03) 0.35 (0.98) −0.05 (0.78) 0.38 (0.68)
RPE −0.13 (1.10) 0.48 (1.00) −0.16 (1.15) 0.53 (1.03) -0.12 (0.90) 0.47 (0.78)

The error profiles in Fig. 3 also demonstrate consistently small errors across the central 6
mm of the B-scan for each layer, and also shows a high level of similarity between the two
considered methods, with the exception of the GCL/NFL where RNN-GS performed
noticeably better across the entire boundary. Observing these profiles also shows that both
networks exhibit a noticeable central error spike for the OPL/INL boundary attributable to the
merging of the layer boundaries at the fovea. Also, all the layer boundaries showed a spike in
error on the far right side of the profile, which corresponds to the location of the optic nerve
head for a number of scans, where retinal boundaries disappear in this region. Some example
segmentation plots for data set 1 using the RNN-GS method are displayed in Fig. 4.
The patch-based method employed here is also compared with a fully-convolutional based
approach (FCN-GS) as described in Section 2.7. In terms of the boundary position error
(Table 7), the FCN-GS method is comparable in accuracy to RNN-GS and CNN-GS with an
average mean absolute error of 0.55 pixels compared to 0.53 and 0.58 for RNN-GS and CNN-
GS respectively. However, FCN-GS shows a greater level of consistency for the
segmentations with smaller standard deviations on all boundaries with the exception of the
ELM. Similar to the two patch-based methods, FCN-GS showed the lowest error on the ISe
and highest error on the GCL/NFL. In addition, Fig. 3 shows the error profiles of FCN-GS to
be somewhat similar to the two patch-based methods. The FCN contained ~490,000
parameters, approximately 7 times more than the RNN. However, the FCN was much faster
in general with a per-image probability map creation requiring about one second,
approximately 100 times faster than the RNN. For per-image evaluation overall, FCN-GS was
~3.5 times faster than RNN-GS when taking the graph search into consideration.
 Vol. 9, No. 11 | 1 Nov 2018 | BIOMEDICAL OPTICS EXPRESS 5771

Fig. 3. Data set 1 (norm

mal OCT images) mean
m absolute erroor profiles of each boundary for eachh
of the tested methods.

Fig. 4.
4 Example RNN--GS segmentation n plots of two diffferent participantts from data set 1
(norm
mal OCT images) with
w the locationss of the true (solidd lines) and prediccted (dotted lines))
bounddaries marked show
wing the close leveel of agreement beetween them.

4.2 AMD OC
CT data (data set 2)
Given the immportance of ap pplying autom matic image seegmentation too pathological data, the
RNN-GS, CN NN-GS and FCN N-GS methodss were also evaaluated using oonly AMD OC CT images
(data set 2), as described ini Section 2.2. Again, 64x3 2 size patchess are employeed for the
patch-based methods.
m Tablee 8 lists the reesults for the bboundary posittion errors for the three
layer boundarries present in this data. For RNN-GS,
R the IILM was relatiively simple too segment
with a mean absolute errorr of 0.38 pixells and standardd deviation off 0.92 pixels shhowing a
similar level of performancce to that for the normal O OCT images inn data set 1 (Table 7).
However, CN NN-GS and FCN-GS
F perfoormed significcantly worse tthan RNN-GS S on this
boundary in terms
t of both accuracy
a and consistency,
c atttributable to a small number of major
failure cases not evident for RNN-GS. For the RP PEDC boundar ary, the three methods
performed co omparably in terms
t of both accuracy andd consistency, while FCN-G GS clearly
provided the best
b overall peerformance on the BM bounddary. Overall, tthe RNN-GS aand FCN-
GS performed d comparably overall in term ms of mean abbsolute error w with an averagge of 1.17
and 1.07 pixeels respectively
y while CNN-G GS was lower w with an averagge of 1.53 pixells. Figure
5 shows somee example segm mentation plotss from data set 2 using the RN NN-GS methodd.
 Vol. 9, No. 11 | 1 Nov 2018 | BIOMEDICAL OPTICS EXPRESS 5772

Tablle 8. Data set 2 (A

AMD OCT imagess) position error ((in pixels) of each
h layer boundary
for each
e d in mean values and (per A-scan
of the tested methods. The ressults are reported
standdard deviation). The
T best results (smallest error) foor each boundaryy are highlighted
in
i bold text.

RN
NN-GS CNN--GS S
FCN-GS

Layer Boundary Mean error Mean abs. Mean error Mean abs. M
Mean error M
Mean abs.
error error error
ILM -0.03 (0.99) 0.38 (0.92) -0.68 (7.26) 1.10 (7.21) 0.14 (4.29) 0.65 (4.24)
RPEDC −0.52 (3.05
5) 1.05 (2.91) −0.64 (3.30) 1.17 (3.15) --0.43 (3.12) 1.03 (2.97)
BM 0.91 (4.69)) 2.07 (4.31) 1.19 (5.01) 2.31 (4.60) 0.46 (3.79) 1.53 (3.50)

Fig. 5.
5 Example RNN--GS segmentation n plots of two diffferent participantts from data set 2
(AMD D OCT images) withw the locations of the true (solidd lines) and prediccted (dotted lines))
bound daries marked. Thee top image showws close agreementt between the preddictions and truthss
while the bottom imag ge shows an exam mple of a failure case for the BM M boundary with a
vely high level of disagreement
relativ d betw
ween the true and ppredicted boundariies.

5. Discussio
on
This study haas proposed th he novel use ofo recurrent neeural networks to segment thhe retinal
layers in macular OCT imag ges of healthy and pathologiical data sets. T The RNN-GS approach
presented com mbines a recurrrent neural neetwork operatinng as a patch--based classifier with a
subsequent graph search over o the corressponding probbability maps. This work w was partly
inspired by thhe previously proposed
p CNN-GS method foor retinal imagge segmentationn [20], as
well as the RNNN-based ReN Net architecturee [38], proposeed as an alternaative to CNNs ffor image
classification.. Despite the extensive
e work k in the area oof OCT retinall layer segmenntation, to
the best of our knowledge, RNNs
R have nott previously beeen applied to O OCT image annalysis.
The carefu
ful selection off the final RNN N architecturee was informedd using insighhts gained
from a rangee of experimeents. Of the tested t patch ssizes, the verrtically orienteed 64x32
rectangular paatch (longer allong the A-scaan) yielded thee best trade-offf between accuuracy and
complexity. InI particular, its
i superior peerformance coompared to the horizontallyy oriented
32x64 varian nt, which is co onsistent withh the vertical gradient channge observed aalong the
transition betw
ween adjacent layers. A similar result was oobtained whenn observing thee effect of
the RNN operrating direction n, with verticall RNNs outperrforming horizoontal ones. On the other
hand, varying g the size of the
t receptive field
f had negliigible effects on performancce. For a
single layer RNN,
R diminishhing returns were observed w with respect too the number of filters.
This informed d an appropriaate trade-off beetween accuraccy and compleexity of 32 filteers which
was used for each
e layer in th
he final RNN architecture.
a W
With regards to the operating ddirection,
stacking bothh horizontal and d vertical layeers outperformeed unidirectionnal stacked layyers. This
informed the decision to utilize
u layers of
o both directi ons within thee final networrk. Fully-
connected outtput layers did d not provide a sufficient acccuracy/compleexity trade-off and were
 Vol. 9, No. 11 | 1 Nov 2018 | BIOMEDICAL OPTICS EXPRESS 5773

therefore not included in the chosen RNN design. Like Visin et al [38], there was no observed
performance difference between the LSTM and GRU recurrent units.
Using a data set comprising normal OCT images (data set 1), the segmentation results
showed the RNN based approach performed competitively in comparison to a CNN approach
using the same patch-size. RNN-GS showed marginally smaller mean absolute errors for all
seven layer boundaries and a greater consistency (i.e. smaller standard deviations) in the
segmentation than CNN-GS. Overall, mean absolute errors of less than one pixel for all seven
layer boundaries were observed, with less than half a pixel for four of those boundaries,
indicating close agreement to the truth. Despite possessing 17 times fewer parameters, the
evaluation time of RNN-GS was longer than that of CNN-GS. This can be attributed to the
relatively high number of operations required to process an image sequentially as is the case
with the RNN.
To gauge the performance on pathological data, RNN-GS and CNN-GS were also
evaluated using a data set comprising OCT images from patients exhibiting age-related
macular degeneration (data set 2). The RNN showed competitive performance with smaller
mean absolute errors for all three layer boundaries corresponding to a mean improvement of
0.36 pixels. In particular, CNN-GS exhibited a small number of major failure cases for the
ILM boundary. These failure cases were the result of the relatively high level of noise within
some B-scans where the ILM was less well defined. These failures were not evident for RNN-
GS possibly indicating a greater robustness of the method in the presence of noise.
Segmentation of the RPEDC and BM boundaries proved more challenging in the presence of
pathological features. However, for both of these boundaries, RNN-GS exhibited marginally
superior mean absolute errors and standard deviations compared to CNN-GS continuing the
trend evident within the results from the normal OCT images.
It should be noted that this work did not focus on the performance of the method with
regards to different ocular pathologies with only one type of pathological data (AMD)
investigated here. In addition, the RNN network architecture here was optimized using data
from normal OCT images. Future work should attempt to further explore the application of
this method to pathological data by extending the types of pathologies present within the data
as well as investigating an optimal network architecture for such.
In the past, RNNs have proven useful for tasks involving sequential data whereas CNNs
have had considerable success when applied to image data. Consequently, RNNs have
received less attention for image classification problems. Here, the ability of an RNN to
perform competitively against a CNN on such a task was investigated. RNNs are suited to
sequential data, so the good performance relative to the CNN may be attributed to the
sequential nature of the retinal layer structure and features. In all, RNNs provide a viable
alternative to CNNs for this particular problem even with retinal pathology and poor image
quality conditions (AMD data).
It should be noted that the 64x32 patch size used here is not necessarily the most optimal,
with a number of alternative patch sizes not tested for computational reasons. Nonetheless,
these are promising results and the performance here is encouraging. Future work may
investigate other patch sizes and, in particular, larger vertically-oriented rectangular patches
as these appear to give the best tradeoff between performance and speed.
The patch-based approach presented here (RNN-GS) was compared to a full image-based
approach utilizing a fully-convolutional network (FCN-GS). For normal OCT images (data
set 1), the accuracy was comparable with RNN-GS. However, FCN-GS was more consistent
in the segmentation with lower standard deviations for most boundaries. FCN-GS was also
much faster in terms of evaluation highlighting this as a possible drawback of the patch-based
method especially when time is critical (e.g. for many clinical applications where rapid
segmentation performance is required). However, it should be noted that optimizing the speed
of the RNN was not a focus here and should be investigated in future work.
 Vol. 9, No. 11 | 1 Nov 2018 | BIOMEDICAL OPTICS EXPRESS 5774

For AMD OCT images (data set 2), the overall accuracy between RNN-GS and FCN-GS
was again comparable. Like CNN-GS, FCN-GS exhibited a number of major failure cases,
which were responsible for the relatively high mean absolute error and standard deviation on
the ILM. On the other hand, FCN-GS was more accurate and consistent on the BM boundary.
It is possible that the superior performance is a result of the greater amount of context
available to the FCN while processing the whole image at once. Future work in the area may
further investigate the relative performance of the patch-based method compared to full-
image based methods.
6. Conclusions
In this paper, the RNN-GS method exhibited promising results for the segmentation of retinal
layers in healthy individuals and AMD patients. In addition, RNNs have been identified as a
sensible alternative to CNNs for tasks with images involving a sequence as is the case with
the layer structure observed in the OCT retinal images used in this work. The results and the
RNN-GS methodology presented here may assist future work in the domain of OCT retinal
segmentation and highlight the potential of RNN-based methods for OCT image analysis.
Funding
Rebecca L. Cooper 2018 Project Grant (DAC); Telethon – Perth Children’s Hospital
Research Fund (DAC).
Acknowledgement
The Titan Xp used for this research was donated by the NVIDIA Corporation.
Disclosures
The authors declare that there are no conflicts of interest related to this article.
References
1. D. Huang, E. A. Swanson, C. P. Lin, J. S. Schuman, W. G. Stinson, W. Chang, M. R. Hee, T. Flotte, K. Gregory,
C. A. Puliafito, and J. G. Fujimoto, “Optical coherence tomography,” Science 254(5035), 1178–1181 (1991).
2. J. F. de Boer, R. Leitgeb, and M. Wojtkowski, “Twenty-five years of optical coherence tomography: the
paradigm shift in sensitivity and speed provided by Fourier domain OCT,” Biomed. Opt. Express 8(7), 3248–
3280 (2017).
3. M. Adhi and J. S. Duker, “Optical coherence tomography - current and future applications,” Curr. Opin.
Ophthalmol. 24(3), 213–221 (2013).
4. X. Chen, M. Niemeijer, L. Zhang, K. Lee, M. D. Abràmoff, and M. Sonka, “3D Segmentation of Fluid-
Associated Abnormalities in Retinal OCT: Probability Constrained Graph-Search–Graph-Cut,” IEEE Trans.
Med. Imaging 31(8), 1521–1531 (2012).
5. P. A. Keane, S. Liakopoulos, R. V. Jivrajka, K. T. Chang, T. Alasil, A. C. Walsh, and S. R. Sadda, “Evaluation
of Optical Coherence Tomography Retinal Thickness Parameters for Use in Clinical Trials for Neovascular Age-
Related Macular Degeneration,” Invest. Ophthalmol. Vis. Sci. 50(7), 3378–3385 (2009).
6. P. Malamos, C. Ahlers, G. Mylonas, C. Schütze, G. Deak, M. Ritter, S. Sacu, and U. Schmidt-Erfurth,
“Evaluation of segmentation procedures using spectral domain optical coherence tomography in exudative age-
related macular degeneration,” Retina 31(3), 453–463 (2011).
7. C. A. Puliafito, M. R. Hee, C. P. Lin, E. Reichel, J. S. Schuman, J. S. Duker, J. A. Izatt, E. A. Swanson, and J. G.
Fujimoto, “Imaging of macular diseases with optical coherence tomography,” Ophthalmology 102(2), 217–229
(1995).
8. J. C. Bavinger, G. E. Dunbar, M. S. Stem, T. S. Blachley, L. Kwark, S. Farsiu, G. R. Jackson, and T. W.
Gardner, “The Effects of Diabetic Retinopathy and Pan-Retinal Photocoagulation on Photoreceptor Cell
Function as Assessed by Dark Adaptometry,” Invest. Ophthalmol. Vis. Sci. 57(1), 208–217 (2016).
9. D. Koozekanani, K. Boyer, and C. Roberts, “Retinal thickness measurements from optical coherence
tomography using a Markov boundary model,” IEEE Trans. Med. Imaging 20(9), 900–916 (2001).
10. J. Oliveira, S. Pereira, L. Gonçalves, M. Ferreira, and C. A. Silva, “Multi-surface segmentation of OCT images
with AMD using sparse high order potentials,” Biomed. Opt. Express 8(1), 281–297 (2017).
11. D. Cabrera Fernández, H. M. Salinas, and C. A. Puliafito, “Automated detection of retinal layer structures on
optical coherence tomography images,” Opt. Express 13(25), 10200–10216 (2005).
12. R. Kafieh, H. Rabbani, M. D. Abramoff, and M. Sonka, “Intra-retinal layer segmentation of 3D optical
coherence tomography using coarse grained diffusion map,” Med. Image Anal. 17(8), 907–928 (2013).
 Vol. 9, No. 11 | 1 Nov 2018 | BIOMEDICAL OPTICS EXPRESS 5775

13. S. J. Chiu, X. T. Li, P. Nicholas, C. A. Toth, J. A. Izatt, and S. Farsiu, “Automatic segmentation of seven retinal
layers in SDOCT images congruent with expert manual segmentation,” Opt. Express 18(18), 19413–19428
(2010).
14. S. Niu, L. de Sisternes, Q. Chen, T. Leng, and D. L. Rubin, “Automated geographic atrophy segmentation for
SD-OCT images using region-based C-V model via local similarity factor,” Biomed. Opt. Express 7(2), 581–600
(2016).
15. S. J. Chiu, M. J. Allingham, P. S. Mettu, S. W. Cousins, J. A. Izatt, and S. Farsiu, “Kernel regression based
segmentation of optical coherence tomography images with diabetic macular edema,” Biomed. Opt. Express
6(4), 1172–1194 (2015).
16. K. Li, X. Wu, D. Z. Chen, and M. Sonka, “Optimal Surface Segmentation in Volumetric Images - a Graph-
Theoretic Approach,” IEEE Trans. Pattern Anal. Mach. Intell. 28(1), 119–134 (2006).
17. J. Tian, B. Varga, G. M. Somfai, W.-H. Lee, W. E. Smiddy, and D. C. DeBuc, “Real-time automatic
segmentation of optical coherence tomography volume data of the macular region,” PLoS One 10(8), e0133908
(2015).
18. P. P. Srinivasan, S. J. Heflin, J. A. Izatt, V. Y. Arshavsky, and S. Farsiu, “Automatic segmentation of up to ten
layer boundaries in SD-OCT images of the mouse retina with and without missing layers due to pathology,”
Biomed. Opt. Express 5(2), 348–365 (2014).
19. K. McDonough, I. Kolmanovsky, and I. V. Glybina, “A neural network approach to retinal layer boundary
indentification from optical coherence tomography images,” in Proceedings of 2015 IEEE conference on
Computational Intelligence in Bioinformatics and Computational Biology (IEEE, 2015), 1–8.
20. L. Fang, D. Cunefare, C. Wang, R. H. Guymer, S. Li, and S. Farsiu, “Automatic segmentation of nine retinal
layer boundaries in OCT images of non-exudative AMD patients using deep learning and graph search,”
Biomed. Opt. Express 8(5), 2732–2744 (2017).
21. J. Hamwood, D. Alonso-Caneiro, S. A. Read, S. J. Vincent, and M. J. Collins, “Effect of patch size and network
architecture on a convolutional neural network approach for automatic segmentation of OCT retinal layers,”
Biomed. Opt. Express 9(7), 3049–3066 (2018).
22. J. Loo, L. Fang, D. Cunefare, G. J. Jaffe, and S. Farsiu, “Deep longitudinal transfer learning-based automatic
segmentation of photoreceptor ellipsoid zone defects on optical coherence tomography images of macular
telangiectasia type 2,” Biomed. Opt. Express 9(6), 2681–2698 (2018).
23. A. G. Roy, S. Conjeti, S. P. K. Karri, D. Sheet, A. Katouzian, C. Wachinger, and N. Navab, “ReLayNet: Retinal
Layer and Fluid Segmentation of Macular Optical Coherence Tomography using Fully Convolutional
Networks,” Biomed. Opt. Express 8(8), 3627–3642 (2017).
24. Y. Xu, K. Yan, J. Kim, X. Wang, C. Li, L. Su, S. Yu, X. Xu, and D. D. Feng, “Dual-stage deep learning
framework for pigment epithelium detachment segmentation in polypoidal choroidal vasculopathy,” Biomed.
Opt. Express 8(9), 4061–4076 (2017).
25. A. Ben-Cohen, D. Mark, I. Kovler, D. Zur, A. Barak, M. Iglicki, and R. Soferman, “Retinal layers segmentation
using fully convolutional network in OCT images,” RSIP Vision 2017.
26. F. G. Venhuizen, B. van Ginneken, B. Liefers, M. J. J. P. van Grinsven, S. Fauser, C. Hoyng, T. Theelen, and C.
I. Sánchez, “Robust total retina thickness segmentation in optical coherence tomography images using
convolutional neural networks,” Biomed. Opt. Express 8(7), 3292–3316 (2017).
27. M. Chen, J. Wang, I. Oguz, B. L. VanderBeek, and J. C. Gee, “Automated segmentation of the choroid in edi-oct
images with retinal pathology using convolution neural networks,” in Fetal, Infant and Ophthalmic Medical
Image Analysis (Springer, 2017), 177–184.
28. X. Sui, Y. Zheng, B. Wei, H. Bi, J. Wu, X. Pan, Y. Yin, and S. Zhang, “Choroid segmentation from optical
coherence tomography with graph edge weights learned from deep convolutional neural networks,” Journal of
Neurocomputing 237, 332–341 (2017).
29. A. Shah, M. D. Abramoff, and X. Wu, “Simultaneous multiple surface segmentation using deep learning,” in
Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support (Springer,
2017), 3–11.
30. O. Cicek, A. Abdulkadir, S. S. Lienkamp, T. Brox, and O. Ronneberger, “3D U-net: learning dense volumetric
segmentation from sparse annotation,” in International Conference on Medical Image Computing and Computer-
Assisted Intervention (Springer, 2016), 424–432.
31. S. P. Karri, D. Chakraborthi, and J. Chatterjee, “Learning layer-specific edges for segmenting retinal layers with
large deformations,” Biomed. Opt. Express 7(7), 2888–2901 (2016).
32. T. Mikolov, M. Karafiát, L. Burget, J. Cernocky, and S. Khudanpur, “Recurrent neural network based language
model,” in Proceedings of Interspeech (ISCA, 2011), 1045–1048.
33. B. Zhang, D. Xiong, and J. Su, “Recurrent Neural Machine Translation,” arXiv preprint arXiv:1607.08725
(2016).
34. A. Graves, M. Liwicki, S. Fernández, R. Bertolami, H. Bunke, and J. Schmidhuber, “A Novel Connectionist
system for Unconstrained Handwriting Recognition,” IEEE Trans. Pattern Anal. Mach. Intell. 31(5), 855–868
(2009).
35. A. Graves, “Generating Sequence With Recurrent Neural Networks,” arXiv preprint arXiv:1308.0850 (2013).
36. A. Graves, A. Mohamed, and G. Hinton, “Speech recognition with deep recurrent neural networks,” in
Proceedings of 2013 IEEE International Conference on Acoustics, Speech and Signal Processing (IEEE, 2013),
6645–6649.
 Vol. 9, No. 11 | 1 Nov 2018 | BIOMEDICAL OPTICS EXPRESS 5776

37. A. Graves and N. Jaitly, “Towards end-to-end speech recognition with recurrent neural networks,” in
Proceedings of the 31st International Conference on Machine Learning , Volume 32 (JMLR.org, 2014), 1764–
1772.
38. F. Visin, K. Kastner, K. Cho, M. Matteucci, A. C. Courville, and Y. Bengio, “ReNet: A recurrent neural network
based alternative to convolutional networks,” arXiv preprint arXiv:1505.00393 (2015).
39. F. Visin, M. Ciccone, A. Romero, K. Kastner, K. Cho, Y. Bengio, M. Matteucci, and A. Courville, “ReSeg: A
Recurrent Neural Network-based Model for Semantic Segmentation,” arXiv preprint arXiv:1511.07053 (2016).
40. A. Graves, S. Fernandez, and J. Schmidhuber, “Multi-Dimensional Recurrent Neural Networks,” in Artifical
Neural Networks – ICANN 2007, J. M. de Sá, L. A. Alexandre, W. Duch, and D. P. Mandic (Eds.) (Springer,
2007), 549–558.
41. X. Shi, Z. Chen, H. Wang, D. Yeung, W. Wong, and W. Woo, “Convolutional LSTM Network: A Machine
Learning Approach for Precipitation Nowcasting,” in Proceedings of the 28th International Conference on
Neural Information Processing Systems (MIT Press, 2015), 802–810.
42. J. Chen, L. Yang, Y. Zhang, M. Alber, and D. Z. Chen, “Combining Fully Convolutional and Recurrent Neural
Networks for 3D Biomedical Image Segmentation,” in Proceedings of the 30th International Conference on
Neural Information Processing Systems (Curran Associates Inc., 2016), 3044–3052.
43. H. Chen, Q. Dou, D. Ni, J. Cheng, J. Qin, S. Li, and P. Heng, “Automatic Fetal Ultrasound Standard Plane
Detection Using Knowledge Transferred Recurrent Neural Networks,” in Proceedings of the 18th International
Conference on Medical Image Computing and Computer-Assisted Intervention (Springer, 2015), 507–514.
44. B. Kong, Y. Zhan, M. Shin, T. Denny, and S. Zhang, “Recognizing End-Diastole and End-Systole Frames via
Deep Temporal Regression Network,” in Proceedings of International Conference on Medical Image Computing
and Computer-Assisted Intervention (Springer, 2016), 264–272.
45. M. F. Stollenga, W. Byeon, M. Liwicki, and J. Schmidhuber, “Parallel Multi-Dimensional LSTM, With
Application to Fast Biomedical Volumetric Image Segmentation,” in Proceedings of the 28th International
Conference on Neural Information Processing Systems, Volume 2 (MIT Press, 2015), 2998–3006.
46. C. Qin, J. V. Hajnal, D. Rueckert, J. Schlemper, J. Caballero, and A. N. Price, “Convolutional Recurrent Neural
Networks for Dynamic MR Image Reconstruction,” IEEE Trans. Med. Imaging. in press.
47. Y. Xie, Z. Zhang, M. Sapkota, and L. Yang, “Spatial Clockwork Recurrent Neural Network for Muscle
Perimysium Segmentation,” in Proceedings of Medical Image Computing and Computer-Assisted Intervention
(Springer, 2016), 185–193.
48. J. Koutník, K. Greff, F. Gomez, and J. Schmidhuber, “A Clockwork RNN,” in Proceedings of the 31st
International Conference on Machine Learning, Volume 32 (JMLR.org, 2014), 1863–1871.
49. S. A. Read, D. Alonso-Caneiro, and S. J. Vincent, “Longitudinal changes in macular retinal layer thickness in
pediatric populations: Myopic vs non-myopic eyes,” PLoS One 12(6), e0180462 (2017).
50. S. A. Read, D. Alonso-Caneiro, S. J. Vincent, and M. J. Collins, “Longitudinal changes in choroidal thickness
and eye growth in childhood,” Invest. Ophthalmol. Vis. Sci. 56(5), 3103–3112 (2015).
51. S. A. Read, M. J. Collins, and S. J. Vincent, “Light exposure and eye growth in childhood,” Invest. Ophthalmol.
Vis. Sci. 56(11), 6779–6787 (2015).
52. S. A. Read, M. J. Collins, S. J. Vincent, and D. Alonso-Caneiro, “Macular retinal layer thickness in childhood,”
Retina 35(6), 1223–1233 (2015).
53. S. Farsiu, S. J. Chiu, R. V. O’Connell, F. A. Folgar, E. Yuan, J. A. Izatt, and C. A. Toth, “Quantitative
Classification of Eyes with and without Intermediate Age-related Macular Degeneration Using Optical
Coherence Tomography,” Ophthalmology 121(1), 162–172 (2014).
54. M. C. Mozer, “A focused backpropagation algorithm for temporal pattern recognition,” Complex Syst. 3(4),
349–381 (1989).
55. P. J. Werbos, “Generalization of backpropagation with application to a recurrent gas market model,” Neural
Netw. 1(4), 339–356 (1988).
56. P. J. Werbos, “Backpropagation through time: what it does and how to do it,” Proc. IEEE 78(10), 1550–1560
(1990).
57. S. Hochreiter and J. Schmidhuber, “Long Short-term Memory,” Neural Comput. 9(8), 1735–1780 (1997).
58. H. Salehinejad, S. Sankar, J. Barfett, E. Colak, and S. Valaee, “Recent Advances in Recurrent Neural Networks,”
arXiv preprint arXiv:1801.01078 (2018).
59. R. Pascanu, T. Mikolov, and Y. Bengio, “On the difficulty of training Recurrent Neural Networks,” in
Proceedings of the 30th International Conference on Machine Learning, Volume 28 (JMLR.org, 2013), 1310–
1318.
60. K. Cho, B. van Merrienboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y. Bengio, “Learning
Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation,” in Proceedings of the
2014 Conference on Empirical Methods in Natural Language Processing (Association for Computational
Linguistics, 2014), 1724–1734.
61. N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, “Dropout: a simple way to prevent
neural networks from overfitting,” J. Mach. Learn. Res. 15(1), 1929–1958 (2014).
62. S. Ioffe and C. Szegedy, “Batch Normalization: Accelerating Deep Network Training by Reducing Internal
Covariate Shift,” in Proceedings of the 32nd International Conference on Machine Learning – Volume 37
(JMLR.org, 2015), 448–456.
 Vol. 9, No. 11 | 1 Nov 2018 | BIOMEDICAL OPTICS EXPRESS 5777

63. D. P. Kingma and J. Ba, “Adam: A Method for Stochastic Optimization,” arXiv preprint arXiv:1412.6980
(2017).
64. K. Janocha and W. M. Czarnecki, “On Loss Functions for Deep Neural Networks in Classification,” arXiv
preprint arXiv:1702.05659 (2017).
65. L. Prechelt, “Early Stopping - But When?” in Neural Networks: Tricks of the Trade, G. B. Orr, O. R. Muller
(Eds.) (Springer-Verlag, 1998).
66. R. Caruana, S. Lawrence, and L. Giles, “Overfitting in neural nets: backpropagation, conjugate gradient, and
early stopping,” in Proceedings of the 13th International Conference on Neural Information Processing Systems,
T. K. Leen, T. G. Dietterich, and V. Tresp, eds. (MIT Press, 2000), 381–387.
67. F. Chollet, “Keras” https://github.com/fchollet/keras.
68. Tensorflow white paper, “Tensorflow: Large-scale machine learning on heterogeneous systems”
https://tensorflow.org
69. E. W. Dijkstra, “A note on two problems in connexion with graphs,” Numer. Math. 1(1), 269–271 (1959).
70. O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional Networks for Biomedical Image
Segmentation,” arXiv preprint arXiv:1505.04597 (2015).
 Progress in Retinal and Eye Research xxx (2016) 1e19

Contents lists available at ScienceDirect

Progress in Retinal and Eye Research

journal homepage: www.elsevier.com/locate/prer

Imaging retina to study dementia and stroke

Carol Yim-lui Cheung a, *, 1, 2, M. Kamran Ikram c, d, 1, 2, Christopher Chen e, f, 2,
Tien Yin Wong b, c, 2
a
Department of Ophthalmology and Visual Sciences, The Chinese University of Hong Kong, Hong Kong
b
Singapore Eye Research Institute, Singapore National Eye Centre, Singapore
c
Duke-NUS Graduate Medical School, National University of Singapore, Singapore
d
Departments of Neurology & Epidemiology, Erasmus University Medical Center, Rotterdam, The Netherlands
e
Memory Aging and Cognition Centre, National University Health System, Singapore
f
Department of Pharmacology, National University of Singapore, Singapore

a r t i c l e i n f o a b s t r a c t

Article history: With increase in life expectancy, the number of persons suffering from common age-related brain dis-
Received 26 August 2016 eases, including neurodegenerative (e.g., dementia) and cerebrovascular (e.g., stroke) disease is expected
Received in revised form to rise substantially. As current neuro-imaging modalities such as magnetic resonance imaging may not
7 November 2016
be able to detect subtle subclinical changes (resolution <100e500 mm) in dementia and stroke, there is
Accepted 1 January 2017
Available online xxx
an urgent need for other complementary techniques to probe the pathophysiology of these diseases. The
retina - due to its anatomical, embryological and physiological similarities with the brain - offers a
unique and accessible “window” to study correlates and consequences of subclinical pathology in the
Keywords:
Retinal imaging
brain. Retinal components such as the microvasculature and retinal ganglion cell axons can now be
Retinal microvasculature visualized non-invasively using different retinal imaging techniques e.g., ocular fundus photography and
Retinal ganglion cell optical coherence tomography. Advances in retinal imaging may provide new and potentially important
Dementia insights into cerebrovascular neurodegenerative processes in addition to what is currently possible with
Alzheimer's disease neuro-imaging. In this review, we present an overview of the current literature on the application of
Stroke retinal imaging in the study of dementia and stroke. We discuss clinical implications of these studies,
Cerebrovascular disease novel state-of-the-art retinal imaging techniques and future directions aimed at evaluating whether
retinal imaging can be an additional investigation tool in the study of dementia and stroke.
© 2017 Elsevier Ltd. All rights reserved.

Contents

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
2. The retina as a model to study cerebral microvascular and neuronal damage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3. Advances in retinal imaging techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.1. Ocular fundus or retinal photography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
3.2. Optical coherence tomography (OCT) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
4. Retinal changes and dementia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
4.1. Retinal vascular changes and dementia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
4.2. Retinal neuronal and axonal loss and dementia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
5. Retinal changes and stroke . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
6. Retinal changes and subclinical cerebral small vessel disease . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
7. Potential clinical and research use . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00

* Corresponding author. CUHK Eye Centre, Hong Kong Eye Hospital, 147K Argyle
Street, Kln, Hong Kong.
E-mail address: [email protected] (C.Y.-l. Cheung).
1
These authors are joint first authors and contributed equally.
2
Percentage of work contributed by each author in the production of the
manuscript is as follows: Cheung 50%, Ikram 20%, Chen 15%, Wong 15%.

http://dx.doi.org/10.1016/j.preteyeres.2017.01.001
1350-9462/© 2017 Elsevier Ltd. All rights reserved.

Please cite this article in press as: Cheung, C.Y.-l., et al., Imaging retina to study dementia and stroke, Progress in Retinal and Eye Research (2016),
http://dx.doi.org/10.1016/j.preteyeres.2017.01.001
2 C.Y.-l. Cheung et al. / Progress in Retinal and Eye Research xxx (2016) 1e19

8. Limitations of current studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00

9. Recent new advances in retinal imaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
9.1. Automated retinal vascular imaging software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
9.2. New optical coherence tomography instruments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
9.3. Retinal fundus autofluorescence imaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
9.4. Dynamic retinal vessel analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
9.5. Adaptive optics imaging of the retina . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
10. Future directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
11. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
Conflict of interest statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
Review criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 00

1. Introduction (CNS), can also be measured and quantified using optical coherence
tomography (OCT). Fig. 1 illustrates the retina as a model to study
With increase in life expectancy, the number of persons cerebral microvascular and neuronal damage using retinal imaging.
suffering from common age-related brain diseases, including These retinal imaging tools offer unique information on the status
neurodegenerative (e.g., dementia) and cerebrovascular (e.g., of the cerebral microvasculature and neuronal structure that is
stroke) disease is expected to rise substantially (Cotter, 2007; distinct from brain imaging measures, suggesting that retinal im-
Larson and Langa, 2008). In Western countries, one in four per- aging provides a complementary approach to study the patho-
sons above the age of 55 years will develop dementia, and one in physiology of dementia and stroke, and may support a given
five will suffer a stroke (Donnan et al., 2008; Fox et al., 2001; diagnosis or prognosis (Cheung et al., 2014b, 2015a; Heringa et al.,
Rimmer et al., 2005). Globally, these brain diseases constitute a 2013; Ikram et al., 2012; MacGillivray et al., 2014; Patton et al.,
major health and societal burden for patients and their families. As 2005, 2006).
therapeutic options are limited, effective preventive strategies and In this review, we summarize recent findings on the application
methods for early diagnosis are needed. While clinical symptoms of retinal imaging as a tool to study the pathophysiology of both
for both dementia and stroke manifest late in the disease course, neurodegenerative (dementia) and cerebrovascular (stroke) dis-
the underlying subclinical pathological processes (e.g., cerebral eases. We will finally discuss the clinical implications of these
atrophy, white matter lesions and microvascular diseases) occur findings and future research directions.
much earlier and in fact are highly prevalent in the older population
years before the onset of clinical disease (Ferri et al., 2005;
Sahadevan et al., 2008; Venketasubramanian et al., 2005). 2. The retina as a model to study cerebral microvascular and
Non-invasive neuroimaging tools such as magnetic resonance neuronal damage
imaging (MRI), are the most widely used modality to study sub-
clinical pathology in dementia and stroke. Studies show subclinical During embryonic development, the retina and optic nerve
structural brain changes seen on MRI (e.g. periventricular and develop as a direct extension of the diencephalon. In terms of the
subcortical white matter lesions, lacunar infarcts, microbleeds, at- microvasculature, the retinal arterioles and venules, measuring
rophy) are related to future risk of dementia and stroke (Kantarci, 100e300 mm in diameter, share similar features with cerebral small
2005; Smith et al., 2003; Vermeer et al., 2003). While MRI mea- blood vessels including end arterioles without anastomoses, barrier
sures are considered as good candidates, they may not be useful at function, auto-regulation and relatively low-flow and high-oxygen-
the early stages of the pathological process and still cannot deter- extraction systems, which allow a unique perspective on the ce-
mine many of the mechanisms that underlie the subclinical brain rebral microvasculature, in particular new insights into the
changes due to limitations in spatial resolution to detect subtle microvascular etiology (versus macrovascular etiology) of diseases
degenerative and microvascular changes of less than 500 mm (Cheung et al., 2014b, 2015a; Ikram et al., 2012; London et al., 2013;
(Pantoni and Garcia, 1997). Furthermore, although other bio- Patton et al., 2005, 2006). Moreover, the retina is composed of
markers, including the detection of increased amyloid and tau layers of specialized neurons that are interconnected through
deposition by positron emission tomography (PET) imaging and synapses. RGCs - displaying typical properties of CNS neurons and
assays for amyloid and tau in cerebrospinal fluid, have been comprising a cell body, dendrites and an axon - are neurons located
developed in dementia research in recent years, the availability and in the ganglion cell layer in the retina that receive visual informa-
acceptability of these examinations are limited to highly special- tion from photoreceptors via various intermediate neuronal cells in
ized clinics with cost-intensive equipment, and are unlikely to be the retina, including the bipolar, amacrine, and horizontal cells
widely available. Additional tools to study subclinical pathology (Wassle, 2004). Visual information is passed along the axons of
and biomarkers of dementia and stroke are therefore needed. RGCs, which are myelinated as they leave the eyes, forming the
The retina shares similar embryological origin, anatomical fea- optic nerve. The optic nerves converge at the optic chiasm before
tures and physiological properties with the brain and hence offers a partial decussation and continuing as the optic tracts to terminate
unique and accessible “window” to study the correlates and con- at the lateral geniculate nucleus, where visual information is
sequences of subclinical pathology in both dementia and stroke relayed to the visual cortex in the occipital lobe, providing the
(Cheung et al., 2014b; London et al., 2013; Patton et al., 2005). The connection between the eye and the CNS (Dowling, 2012; Wassle,
retinal microvasculature can now be visualized, quantified and 2004). Given this strong connection, it has been suggested that
monitored non-invasively using ocular fundus photography and neurodegenerative processes in the brain may also lead to similar
computer software image analysis. In addition, retinal ganglion cell changes in RGCs and optic nerve (Ho et al., 2012; London et al.,
(RGC) axons, the ocular extension of the central nervous system 2013).

Please cite this article in press as: Cheung, C.Y.-l., et al., Imaging retina to study dementia and stroke, Progress in Retinal and Eye Research (2016),
http://dx.doi.org/10.1016/j.preteyeres.2017.01.001
 C.Y.-l. Cheung et al. / Progress in Retinal and Eye Research xxx (2016) 1e19 3

Fig. 1. A schematic diagram illustrating the retina as a model to study cerebral microvascular and neuronal damage using retinal vascular imaging and optical coherence to-
mography. Magnetic resonance imaging (MRI) and positron emission tomography (PET) are current neuro-imaging tools to study subclinical pathology in dementia and stroke.

3. Advances in retinal imaging techniques documentation of generalized retinal vessel narrowing or

widening, and has been applied in particular in large epidemio-
3.1. Ocular fundus or retinal photography logical studies (Hubbard et al., 1999; Wong et al., 2004a). Retinal
arteriolar narrowing has been correlated to systemic peripheral
Ocular fundus photography is the classic imaging technique to vasoconstriction and hypertension, while retinal venular widening
capture photography of the retina. The optical design of fundus has been correlated to endothelial dysfunction, inflammation, and
cameras is based on the principle of monocular indirect ophthal- microvascular hypoxia (Nguyen and Wong, 2006; Sun et al., 2009).
moscopy. The retina can be imaged directly as the pupil is used as The branching pattern of vascular networks is structurally
both an entrance and exit for the fundus camera's illuminating and developed to minimize the energy required to maintain blood flow,
imaging light rays (Curtin, 2003). Ocular fundus photography can according to Murray's principle (Murray, 1926). Murray's principle
capture three types of retinal vascular signs: (1) qualitative reti- is the first law for the optimal design of blood vessels and the
nopathy and retinal arteriolar signs, (2) changes in retinal vascular optimal configuration corresponds to the blood flow rate being
caliber and (3) changes in global geometrical patterns of the retina proportional to the cube power of the radius of the blood vessel.
(Cheung et al., 2012a). New classes of retinal vascular network parameters (e.g. tortuosity,
Retinopathy signs (e.g. microaneurysms, hemorrhages, cotton fractal dimension, and branching) can now be measured using
wool spots, hard exudates) and retinal arteriolar wall signs advanced image processing techniques (Fig. 2) (Cheung et al.,
(generalized and focal arteriolar narrowing, arteriovenous nicking 2011b, 2012b, 2011a; Grinton et al., 2012; Hughes et al., 2006,
or nipping, and enhanced arteriolar light reflex) are commonly seen 2009; Lau et al., 2014; Thomas et al., 2014). These new retinal
in patients with systemic diseases such as diabetes or systemic vascular network parameters may capture the “optimality”, and
hypertension. Retinal arteriolar wall signs are more common in “efficiency” of blood distribution in the retinal network, which in
hypertension. These signs can be assessed from retinal photographs term reflects the integrity of the cerebral microcirculation (Cheung
according to different classification systems (e.g. modified Airlie et al., 2015a). In line with these observations, histo-pathological
House classification system for Diabetic Retinopathy Study (1981)), studies have suggested that processes such as atherosclerosis, lip-
Wong-Mitchell classification system for hypertensive retinopathy ohyalinosis and arteriosclerosis lead to elongated and tortuous
(Wong and Mitchell, 2004) by ophthalmologists, optometrists, vessels (Benderro and Lamanna, 2011; Lammie, 2002). These
trained readers and increasingly, by computer-assisted retinopathy changes affect the ability of cerebral arterioles to maintain control
detection programs (Abramoff et al., 2016; Faust et al., 2012). of local blood flow, predisposing areas served by these dysfunc-
Quantitative measurement of retinal vascular caliber (or retinal tional vessels to ischemic damage (Tomita et al., 2005). In the
vessel diameter) using computer software has been proposed for retina, reduced vascular fractal dimension is indicative of vessel

Please cite this article in press as: Cheung, C.Y.-l., et al., Imaging retina to study dementia and stroke, Progress in Retinal and Eye Research (2016),
http://dx.doi.org/10.1016/j.preteyeres.2017.01.001
4 C.Y.-l. Cheung et al. / Progress in Retinal and Eye Research xxx (2016) 1e19

Fig. 2. Measurement of retinal vascular caliber and network parameters (tortuosity, fractal dimension, and branching) from a retinal photograph using the Singapore I Vessel
Assessment (SIVA, National University of Singapore, Singapore). SIVA automatically identifies the optic disc, places a grid with reference to the center of optic disc, identifies vessel
type and calculates retinal parameters (Lau et al., 2014). The grid used in SIVA is a modified grid used in the Atherosclerosis Risk in Communities (ARIC) study (Hubbard et al., 1999)
with 4 concentric circles which demarcate an average optic disc (Cheung et al., 2010a). The measured areas for caliber and network parameters are standardized as the regions from
0.5 to 1.0 and 0.5 to 2.0 optic disc diameters from the disc margin according to ARIC study and SIVA grading protocols, respectively. All the vessels within the measured areas are
traced and an unsupervised approach is used to classify vessels into arterioles or venules. Measurement of retinal parameters is performed by trained graders who are responsible
for the visual evaluation of SIVA automated measurement and perform manual intervention if necessary. The SIVA manual options include adjusting the placement of the grid,
correcting any wrongly identified vessel type, identifying connecting or disconnecting vessel segments (if the measurement segment was considered unreliable), and modifying
widths of vessel segments, according to the SIVA grading protocol.

rarefaction and collapse resulting from hypoxia, whereas increased should be taken into consideration (e.g. factoring in refractive error
vessel tortuosity is indicative of vessel wall dysfunction and blood- and magnification) in the interpreting of study findings of retinal
retinal barrier damage (Hammes et al., 2011). vascular caliber and other parameters.
Measurement of the retinal vascular caliber and its branching Although the assessment of retinal vascular signs from fundus
pattern with computer software and standardized photographic photography has now been used in numerous large population-
protocols is increasingly used as a tool in microvascular disease based epidemiological studies around the world, such assessment
research (Cheung et al., 2012a; Liew et al., 2008). Nonetheless, there in these studies were actually based on a relatively narrow view of
are many methodological and technical challenges in quantitative the retina area (one or two 45
or 60
fundus photographs). Ultra-
measurement of retinal vascular caliber and other structural pa- wide field retinal imaging technology, based on principle of
rameters from retinal photographs. For example, most retinal confocal laser scanning microscopy combined with a concave
vessel imaging software are only semi-automated, and require elliptical mirror, has the ability to capture up to 200
of the retina in
manual adjustment and assessment usually by trained technicians a single image without pupil dilation for assessment of peripheral
(Cheung et al., 2011b; Sherry et al., 2002). Manual input, even lesions (Kernt et al., 2012). Recent studies have demonstrated that
following a standardized protocol, will introduce additional vari- identification rates of any diabetic retinopathy and vision-
ability in retinal vessel measurements. Fully automated software to threatening diabetic retinopathy are higher, and ungradable im-
measure retinal vascular caliber and other abnormalities is still age rate of diabetic retinopathy is lower, compared with traditional
under development and is not yet ready for routine clinical or fundus cameras (Silva et al., 2014, 2016). It is possible that the
research use. Additionally, there are other sources of variability. assessment of more peripheral vessels (i.e. including smaller ves-
Variations in refractive error and axial length may affect the sels) may provide an even better representation of the overall
magnification and apparent dimensions of retinal vascular caliber retinal vascular structure (Cheung et al., 2010a). However, quanti-
and structures (Wong et al., 2004b). Differences in retinal tative measurement of retinal vasculature from ultra-wide field
pigmentation, pupil dilation, presence of cataract and other media retinal images has not been clinically tested and is currently under
opacities, photographic technique, camera type (e.g. mydriatic, development (Sagong et al., 2015).
non-mydriatic, hand-held), and image quality (brightness, focus,
and contrast) may introduce additional sources of variation and
affect measurements (Li et al., 2010; Maberley et al., 2004; 3.2. Optical coherence tomography (OCT)
Rochtchina et al., 2008; Wainwright et al., 2010). These issues
The OCT scan, a noninvasive in vivo optical biopsy of the retina,

Please cite this article in press as: Cheung, C.Y.-l., et al., Imaging retina to study dementia and stroke, Progress in Retinal and Eye Research (2016),
http://dx.doi.org/10.1016/j.preteyeres.2017.01.001
 C.Y.-l. Cheung et al. / Progress in Retinal and Eye Research xxx (2016) 1e19 5

providing high-resolution imaging based on the principle of low- neuropsychological tests (Cheung et al., 2014a; Ding et al., 2011;
coherence interferometry, has enabled remarkable advances in Gatto et al., 2012; Kim et al., 2011a; Liew et al., 2009; Ong et al.,
assessing RGC axons by quantifying peripapillary retinal nerve fiber 2014; Patton et al., 2007; Taylor et al., 2015).
layer (RNFL) damage in the last two decades, particularly in the Alzheimer's disease is the most common subtype of dementia
field of glaucoma (Drexler and Fujimoto, 2008; Huang et al., 1991). characterized by gradual decline in cognitive abilities, ultimately
More recently, OCT has evolved from older generations of time- impairing daily functioning (Alzheimer's, 2016). Vascular dementia
domain OCT to spectral-domain OCT (SD-OCT). SD-OCT allows is the second most common subtype of dementia, which is caused
imaging of the retina in 3-D volume at a higher scan speed, higher by cerebrovascular burden leading to stepwise worsening in
axial resolution (~5 vs. 10 mm) and lower measurement variability cognitive functioning (O'Brien and Thomas, 2015). Few studies have
than the time-domain OCT (Leung et al., 2009). Importantly, with explored whether retinal vascular changes are linked to prevalent
improved image processing technology and normative databases, (present cases of a disease in a cross-sectional study) and incident
detailed ganglion cell analysis at the macular region is now possible (new cases detected during follow-up in a longitudinal study) de-
(Fig. 3). Measurement of the ganglion cell layer (GCL); which is mentia and its subtypes as shown in Table 1. The Cardiovascular
composed of cell bodies; and the inner plexiform layer (IPL); which Health Study showed that retinopathy and retinal arteriolar wall
contains the RGC dendrites, is now available and is being used in signs were associated with dementia only in persons with hyper-
assessing optic neuropathy (e.g. glaucoma) at the macular region. tension and in persons without diabetes (Baker et al., 2007). In the
Since the macular region contains more than 50% of total RGCs and AGES-Reykjavik Study, retinopathy was only associated with
as the RGC cell body size is 10e20 times the diameter of their axons, vascular dementia, but not with Alzheimer's disease (Qiu et al.,
changes in the macular ganglion cell-inner plexiform layer (GC-IPL) 2010). In the Rotterdam Study, retinopathy was associated with
and macular ganglion cell complex (GCC, includes RNFL, GCL and prevalent dementia and prevalent Alzheimer's disease (Schrijvers
IPL) are more sensitive than peripapillary RNFL axonal reduction to et al., 2012), whereas longitudinal data from the same study
neurodegenerative processes (Mwanza et al., 2012; Tan et al., showed no association with incident dementia (Schrijvers et al.,
2009). Furthermore, the photoreceptor layer, composed of neu- 2012). In contrast, persons with retinal venular widening were
rons that are directly sensitive to light, can also be visualized using more likely to develop incident dementia, particularly incident
SD-OCT and can be assessed using OCT segmentation algorithms vascular dementia (de Jong et al., 2011).
(Schuman et al., 2009). It is noteworthy that there is age-related Three recent case-control studies have examined the new
reduction in RGCs and RGC axons in the normal retina (Leung retinal geometric branching parameters such as fractal dimensions,
et al., 2012, 2013). tortuosity and branching in patients with Alzheimer's disease
(Cheung et al., 2014c; Frost et al., 2013; Williams et al., 2015).
4. Retinal changes and dementia Findings from these studies were consistent with a previous report
that retinal venular narrowing is associated with Alzheimer's dis-
4.1. Retinal vascular changes and dementia ease, hypothesized to be related to an increased venous wall
thickness due to collagen deposition in cerebral veins (Berisha et al.,
A growing body of evidence therefore now suggests that 2007b). These studies have also consistently shown that a sparser
vascular disease mechanisms play an important role in the risk for retinal vascular network (decreased retinal fractal dimension) are
aging-related cognitive decline and disorders (Chen et al., 2016; associated with Alzheimer's disease, suggesting loss of vessel
Gorelick et al., 2011; Kling et al., 2013; O'Brien and Thomas, 2015; density in the retina may reflect similar pathophysiological pro-
Rosenberg et al., 2016; Sachdev et al., 2014). In particular, micro- cesses in the cerebral microvasculature in patients with Alz-
vascular or small vessel disease is now thought to be a major heimer's disease (de la Torre, 2004). In the Australian Imaging,
contributor to dementia (Brown and Thore, 2011; De Silva and Biomarkers and Lifestyle study of ageing, it has also been reported
Faraci, 2016; van Veluw et al., 2015; Wardlaw et al., 2013). A that healthy persons with neocortical brain amyloid plaque burden
number of population-based studies have demonstrated that as assessed by Pittsburgh Compound-B PET (PiB-PET) imaging have
qualitative retinal vascular signs and quantitative retinal vascular altered retinal vascular network (larger venular branching asym-
measures, including retinal arteriolar narrowing, retinal venular metry factor and arteriolar length-to-diameter ratio) compared
widening and suboptimal retinal vascular network, are associated with healthy persons without amyloid plaque (Frost et al., 2013).
with poorer cognitive performance, as assessed by various However, causal inferences from these studies on the role of

Fig. 3. Detailed retinal layers of a cross-sectional spectral-domain optical coherence topography image centered at the macula from a human subject.

Please cite this article in press as: Cheung, C.Y.-l., et al., Imaging retina to study dementia and stroke, Progress in Retinal and Eye Research (2016),
http://dx.doi.org/10.1016/j.preteyeres.2017.01.001
6 C.Y.-l. Cheung et al. / Progress in Retinal and Eye Research xxx (2016) 1e19

Table 1
Recent clinical studies of dementia and retinal measures, as defined from ocular fundus photography.

Authors, year Sample Retinal measure Summary of results

Baker et al., 2007 2211 subjects from the Retinopathy, arteriolar wall Retinopathy (OR 2.10 [1.04, 4.24]) and FAN (OR 3.02 [1.51, 6.02]) associated with
cardiovascular health study signs, CRAE, CRVE dementia in persons with hypertension, FAN (OR 3.57 [1.31, 9.75]) associated with
mixed dementia.
Qiu et al., 2010 3906 subjects from the Retinopathy Retinopathy associated with VaD (OR 1.95 [1.04, 3.62]).
AGES-Reykjavik study
de Jong et al., 2011 5553 subjects from CRAE, CRVE Increased CRVE associated with incident dementia (HR 1.11 [1.00, 1.22]) and vascular
Rotterdam study dementia (HR 1.44 [1.10, 1.89]).
Schrijvers et al., 6978 subjects from the Retinopathy Retinopathy associated with prevalent dementia (OR 1.92 [1.24, 2.98]), AD (OR 1.89
2012 Rotterdam study [1.15, 3.10]) and VaD (OR 3.01 [1.26, 7.21], age-gender adjusted model only), but not
associated with incident dementia, AD and VaD.
Frost et al., 2013 25 AD, and 123 age- CRAE, CRVE, fractal AD subjects had decreased CRAE (p ¼ 0.01), CRVE (p < 0.001), arteriolar fractal
matched control subjects dimension, tortuosity, dimension (p ¼ 0.008), venular fractal dimension (p < 0.001), venular tortuosity
bifurcation (p ¼ 0.024), venular branching coefficient (p ¼ 0.019) and arteriolar length-diameter-
ratio (p ¼ 0.048), compared with controls.
Cheung et al., 2014c 136 AD, and 290 age- CRAE, CRVE, fractal Decreased CRVE (OR 2.01 [1.27, 3.19]), arteriolar (OR 1.35 [1.08, 1.68]) and venular (OR
matched control subjects dimension, tortuosity, 1.47 [1.17, 1.84]) fractal dimensions, and increased arteriolar (OR 1.80 [1.40, 2.31]) and
bifurcation venular (OR 1.94 [1.48, 2.53]) tortuosity associated with AD.
Williams et al., 213 AD, and 294 age- CRAE, CRVE, fractal Decreased total fractal dimension (OR 0.77 [0.62, 0.97]) and arteriolar tortuosity (OR
2015 matched control subjects. dimension, tortuosity, 0.78 [0.63, 0.97]) associated with AD.
bifurcation

AD ¼ Alzheimer's disease; CRAE ¼ central retinal artery equivalent; CRVE ¼ central retinal vein equivalent; FAN ¼ focal arteriolar narrowing; HR ¼ hazard ratio; OR ¼ odds
ratio; VaD ¼ vascular dementia.

microvascular pathology as reflected by retinal vascular measures the thickness of RGC bodies and dendrites at macular region, was
in dementia and cognitive decline are limited because most studies more strongly related to patients with mild cognitive impairment,
have been either cross-sectional, retrospective, or had a short compared with periperpillary RNFL axonal loss (Cheung et al.,
duration of follow-up. 2015c).
Furthermore, a reduction of RNFL thickness was associated with
4.2. Retinal neuronal and axonal loss and dementia other types of dementia, such as dementia with Lewy bodies or
dementia associated with Parkinson's disease, suggesting that
Previous histological studies have demonstrated that patients retinal RGC axonal loss as measured by OCT is also present in non-
with Alzheimer's disease have loss of RGCs and optic nerve axons Alzheimer dementias (Garcia-Martin et al., 2012; Moreno-Ramos
(Blanks et al., 1989; Hinton et al., 1986). A recent postmortem study et al., 2013; Moschos et al., 2011; Ratchford et al., 2013). Besides
further suggested that melanopsin RGCs, photoreceptors driving dementia, other reports have also used SD-OCT to assess RGC
circadian photoentrainment, may be specifically affected in Alz- neuronal abnormalities in other neurodegenerative diseases, for
heimer's disease (La Morgia et al., 2016). These observations have example RNFL thinning and optic nerve neurodegeneration in
been confirmed by clinical studies using time-domain OCT (Ascaso stroke (Kalesnykas et al., 2008), and RGC degeneration and RNFL
et al., 2014; Berisha et al., 2007a, b; Garcia-Martin et al., 2012; thinning in multiple sclerosis (Ghezzi et al., 1999; Ratchford et al.,
Garcia-Martin et al., 2014; Iseri et al., 2006; Kesler et al., 2011; Lu 2013; Soderstrom, 2001). A recent study has also linked cerebral
et al., 2010; Moschos et al., 2012; Paquet et al., 2007b, a; Parisi et al., atrophy with macular GC-IPL and RNFL thicknesses as measured by
2001) and other ocular imaging modalities (confocal scanning laser SD-OCT (Ong et al., 2015). It found that reduction in grey matter
ophthalmoscopy (Danesh-Meyer et al., 2006) and scanning laser volume of the occipital and temporal lobes was associated with
polarimetry (Kergoat et al., 2001)) in Alzheimer's disease and mild reduction in macular GC-IPL thickness in a community-based
cognitive impairment, showing a significant reduction in RNFL cohort, providing further evidence that retinal OCT can be used
thickness particularly in superior and inferior quadrants compared for assessing brain neurodegeneration.
to age-matched cognitively normal controls. In a recent meta-
analysis including 702 eyes in patients with Alzheimer's disease 5. Retinal changes and stroke
and 214 eyes in patients with mild cognitive impairment, there was
a significant reduction in RNFL thickness in both Alzheimer's dis- Early work examining the relationship between retinal vascular
ease (weighted mean difference 12.4 mm) and mild cognitive changes and cerebral pathology was mainly focused on clinical
impairment (weighted mean difference 8.23 mm) as compared stroke. Histopathological studies performed in the late 1940s and
with healthy controls (Thomson et al., 2015). early 1970s suggested that in subjects with cerebrovascular dis-
Table 2 summarizes the recent clinical studies of dementia and eases, such as those with cerebral infarcts and hemorrhages, similar
retinal measures, as defined from OCT. Several recent studies have pathological changes such as fibrinoid degeneration and fibro-
used the improved spatial resolution SD-OCT to examine the as- hyalinoid thickening were observed both in retinal arterioles and
sociations of RNFL thickness and specific retinal layers at the their counterparts in several regions of the brain including puta-
macular region in Alzheimer's disease (Fig. 4) (Bambo et al., 2014; men, thalamus and pons (Alpers et al., 1948; Goto et al., 1975).
Bayhan et al., 2015; Cheung et al., 2015c; Gao et al., 2015; However, inferences from these initial studies were hampered due
Gharbiya et al., 2014; Kirbas et al., 2013; Kromer et al., 2014; to small sample sizes, highly-selected patient groups and inade-
Larrosa et al., 2014; Marziani et al., 2013; Polo et al., 2014). Mar- quate control subjects.
ziani et al. and Bayhan et al. reported significant reductions in Since then numerous large population-based studies have re-
macular GCC in Alzheimer's disease patients compared with ported a strong link between retinal vascular changes and clinical
healthy controls (Bayhan et al., 2015; Marziani et al., 2013). Another stroke, both in cross-sectional and prospective studies (Table 3).
study reported that thinner macular GC-IPL, reflecting reduction in Data from the Atherosclerosis Risk in Communities Study firstly

Please cite this article in press as: Cheung, C.Y.-l., et al., Imaging retina to study dementia and stroke, Progress in Retinal and Eye Research (2016),
http://dx.doi.org/10.1016/j.preteyeres.2017.01.001
 C.Y.-l. Cheung et al. / Progress in Retinal and Eye Research xxx (2016) 1e19 7

Table 2
Recent clinical studies of dementia and retinal measures, as defined from optical coherence tomography.

Authors, year Sample OCT measure Summary of results

Imaging method: Time-domain optical coherence tomography

Iseri et al., 2006 14 AD, and 15 age-matched Parapapillary RNFL Reduced parapapillary RNFL thicknesses in superior (p ¼ 0.002), inferior (p < 0.001) and
control subjects thicknesses, macular nasal (p < 0.001) quadrants, macular thicknesses (all p ¼ 0.045, except temporal [p ¼ 0.056]
thicknesses and volume and superior [p ¼ 0.060] inner macula, and superior outer macular [p ¼ 0.460]), and macular
volume (all p ¼ 0.027, except measurement at superior [p ¼ 0.0851] and nasal [p ¼ 0.068]
inner macula, temporal [p ¼ 0.272] and superior [p ¼ 0.264] outer macula) in AD.
Berisha et al., 9 AD, and 8 age-matched Parapapillary RNFL Reduced parapapillary RNFL thickness in superior quadrant in AD (p ¼ 0.02).
2007a control subjects thickness
Paquet et al., 23 MCI, 14 mild AD, 12 Parapapillary RNFL Reduced parapapillary average RNFL thickness in MCI, mild AD and moderate to severe AD,
2007b moderate to severe AD, and thickness compared with controls (all p < 0.01).
15 control subjects Reduced average parapapillary RNFL thickness in moderate to severe AD, compared with
MCI (p < 0.01). No difference in parapapillary average RNFL between MCI and mild AD.
Lu et al., 2010 22 AD, and 22 age-matched Parapapillary RNFL Reduced parapapillary RNFL thicknesses in superior and inferior quadrants (all p < 0.05) in
control subjects thickness AD.
Kesler et al., 24 MCI, 30 AD, and 24 age- Parapapillary RNFL Reduced parapapillary RNFL thicknesses in inferior quadrant (p < 0.005) in MCI and in
2011 matched control subjects thickness superior and inferior quadrants (all p < 0.05) in AD, compared with controls. No difference in
parapapillary RNFL between MCI and AD.
Moschos et al., 30 AD, and 30 age-gender- Parapapillary and macular Reduced parapapillary RNFL in superior (p < 0.001) and inferior (p < 0.001) quadrants, and
2012 matched control subjects thicknesses foveal thicknesses (p ¼ 0.001) in AD.
Garcia-Martin 20 mild AD, 28 age- Parapapillary RNFL and Reduced macular thickness (all p ¼ 0.015) and total macular volume (p ¼ 0.024) in AD. No
et al., 2014 matched control subjects macular thicknesses difference in parapapillary RNFL thickness between mild AD and control.
Ascaso et al., 18 AD, 21 amnestic MCI, Parapapillary RNFL and Reduced parapapillary RNFL thickness in AD and MCI, compared with controls (all
2014 and 41 healthy controls macular thickness p < 0.001). Thickened macular thickness in MCI, compared with AD and controls (all
p < 0.05).
La Morgia et al., 21 AD, and 74 age-matched Parapapillary RNFL Reduced parapapillary RNFL thickness in superior quadrant (p ¼ 0.006) in AD.
2016 control subjects thickness
Imaging method: spectral-domain optical coherence tomography
Marziani et al., 21 AD, and 32 healthy RNFL, and RNFL þ GCL Reduced RNFL (all p ¼ 0.014), and combined RNFL and GCL (RNFL þ GCL) thicknesses (all
2013 subjects thicknesses at macular p ¼ 0.008) at macular region in AD.
region
Kirbas et al., 40 AD, and 40 healthy Parapapillary RNFL Reduced parapapillary RNFL thickness in superior quadrant (p ¼ 0.001) in AD.
2013 controls thickness
Moreno-Ramos 10 AD, 10 dementia with Parapapillary RNFL Reduced parapapillary RNFL thickness in AD, dementia with Lewy bodies, and dementia
et al., 2013 Lewy bodies, 10 dementia thickness associated with Parkinson's disease, compared with controls (all p < 0.001).
associated with Parkinson's
disease and 10 age-
matched controls
Gharbiya et al., 21 mild to moderate AD Choroidal, parapapillary Reduced choroidal thicknesses in AD (all p ¼ 0.036). No difference in parapapillary RNFL (all
2014 and 21 age-matched RNFL and central subfield p  0.3) and central subfield retinal (p ¼ 0.9) thicknesses.
control subjects retinal thicknesses
Kromer et al., 22 AD, and 22 age-gender- Parapapillary RNFL Reduced parapapillary RNFL thickness in superonasal sector in AD (p < 0.0001).
2014 matched controls thickness
Larrosa et al., 151 AD, and 61age- A selective combination of Reduced parapapillary RNFL thickness using linear discriminant function in AD (all
2014 matched healthy subjects parapapillary RNFL and p < 0.001).
macular thicknesses using
linear discriminant function
Bambo et al., 57 AD, and 57 healthy Parapapillary RNFL Reduced parapapillary RNFL thickness in superior (p ¼ 0.001) and inferior (p ¼ 0.003)
2014 controls thickness quadrants in AD.
Polo et al., 2014 75 AD, and 75 age-matched Parapapillary RNFL and Reduced parapapillary RNFL in superior (p ¼ 0.006) and inferior (p ¼ 0.018) quadrants, and
healthy subjects macular thickness macular thicknesses (all p ¼ 0.009, except measurement at fovea [p ¼ 0.115]) in AD.
Cheung et al., 41 MCI, 100 AD, and 123 Macular GC-IPL and Reduced macular GC-IPL thickness in MCI (all p ¼ 0.049, except measurement at superior
2015c normal control subjects parapapillary RNFL [p ¼ 0.064], inferonasal [p ¼ 0.051] and supertemporal [p ¼ 0.359] sectors) and AD (all
thicknesses p ¼ 0.031), compared with controls. Reduced parapapillary RNFL thickness in superior
quadrant in AD (p ¼ 0.039). No difference in macular GC-IPL and parapapillary RNFL
thicknesses between MCI and AD.
Gao et al., 2015 25 AD, 26 MCI and 21 Parapapillary RNFL Reduced parapapillary RNFL thicknesses in AD in superior (p ¼ 0.015), inferior (p ¼ 0.004)
healthy controls thickness, macular and temporal (p ¼ 0.002) quadrants, and MCI in temporal quadrant (p ¼ 0.013), compared
thickness with controls. Reduced macular volume in AD (p ¼ 0.038), and MCI (p ¼ 0.018), compared
with controls.
Bayhan et al., 31 AD and 30 age-matched Choroidal thickness, Reduced choroidal (all p ¼ 0.036, except measurement at 3.0 mm temporal to the fovea
2015 control subjects macular ganglion cell [p ¼ 0.067]) and macular ganglion cell complex (all p  0.033) thicknesses in AD.
complex, and outer retinal
thicknesses

AD ¼ Alzheimer's disease; GC-IPL ¼ ganglion cell-inner plexiform layer; GCL ¼ ganglion cell layer; MCI ¼ mild cognitive impairment; RNFL ¼ retinal nerve fiber layer; SD-
OCT ¼ spectral domain optical coherence tomography; TD-OCT ¼ time domain optical coherence tomography.

showed that the presence of retinopathy at baseline were inde- studies such as the population-based Blue Mountains Eye Study,
pendently associated with incident clinical stroke (Wong et al., the Multi-Ethnic Study of Atherosclerosis and the Singapore Malay
2001), and subsequently demonstrated that persons who had Eye Study have reported similar findings subsequently (Cheung
both cerebral white matter lesions (WML) on MRI and retinopathy et al., 2013; Kawasaki et al., 2012; Mitchell et al., 2005; Wang
signs were at substantially higher risk of incident clinical stroke et al., 2011). With respect to retinal vascular caliber, several
than those without either abnormality (Wong et al., 2002). Other studies consistently reported that retinal venular widening was

Please cite this article in press as: Cheung, C.Y.-l., et al., Imaging retina to study dementia and stroke, Progress in Retinal and Eye Research (2016),
http://dx.doi.org/10.1016/j.preteyeres.2017.01.001
8 C.Y.-l. Cheung et al. / Progress in Retinal and Eye Research xxx (2016) 1e19

Fig. 4. An output of the Cirrus HD-OCT ganglion cell analysis (GCA) algorithm (Carl Zeiss Meditec, Dublin, CA) in a human subject with Alzheimer's disease. The GC-IPL thickness
map uses a false color coding with warm colors represent high and cool colors represent low GC-IPL thickness values. The GC-IPL thickness map has a central circular black area
expressing the lack of RGCs in the fovea. The software further compares the measured thickness to the device's internal normative age-matched database, and generates a deviation
map. Significant reduction in GC-IPL is observed.

related to an increased risk of stroke and stroke mortality (Cheung vascular changes reflect specific cerebral microvasculopathy and
et al., 2013; Ikram et al., 2006a; Kawasaki et al., 2012; Wang et al., may allow more accurate sub-typing of stroke (Baker et al., 2010a,
2007; Wieberdink et al., 2010; Wong et al., 2006; Yatsuya et al., 2010b, 2010c; Cheung et al., 2010b; Doubal et al., 2010a; Doubal
2010), which was confirmed by a meta-analysis including data et al., 2010b; Lindley et al., 2009), or even differentiate stroke
from 20,798 participants (McGeechan et al., 2009). Alternations in from other causes of focal neurologic deficits (Vuong et al., 2015).
the new retinal geometric branching parameters (e.g. decreased Interestingly, a recent study showed that retinal vascular changes
fractal dimension and increased tortuosity) have also been linked to were similarly associated with both large-artery stroke and lacunar
stroke and stroke mortality in recent studies (Cheung et al., 2010b, infarcts, suggesting that structural changes in retinal vessels not
2013; Doubal et al., 2010b; Hughes et al., 2016; Kawasaki et al., only reflect small vessel pathology, but may also result from
2011; Ong et al., 2013; Witt et al., 2006). Nevertheless, whilst the downstream effects of large artery pathology in the retinal and
addition of retinal vascular imaging to existing models of stroke can cerebral circulations (Ong et al., 2013).
improve stroke risk stratification, the overall improvement is only In addition to retinal vascular changes, it was also observed that
about 10% beyond that of established risk factors (Cheung et al., presence of localized RNFL defects as assessed by SD-OCT is asso-
2013; McGeechan et al., 2009). ciated with acute ischemic stroke after adjustment for systemic and
Further studies have demonstrated that retinal vascular changes ocular factors (Wang et al., 2014).
vary according to different stroke subtypes, suggesting that retinal

Please cite this article in press as: Cheung, C.Y.-l., et al., Imaging retina to study dementia and stroke, Progress in Retinal and Eye Research (2016),
http://dx.doi.org/10.1016/j.preteyeres.2017.01.001
 C.Y.-l. Cheung et al. / Progress in Retinal and Eye Research xxx (2016) 1e19 9

Table 3
Recent clinical studies of stroke and retinal measures, as defined from ocular fundus photography.

Authors, Sample Retinal measure Summary of results

year

Cross-sectional studies
Lindley 1321 ischemic stroke Retinopathy, arteriolar FAN (OR 2.36 [1.14, 4.48]), AV-nicking (2.45 [1.34, 4.46), enhanced
et al., wall signs, CRAE, CRVE, arteriolar light reflex (OR 2.61 [1.45, 4.68]), decreased CRAE (OR 2.11
2009 AVR [1.07, 4.17]), decreased AVR (OR 1.87 [1.03, 3.41]), and increased CRVE
(OR 2.08 [1.10, 3.95]) associated with lacunar stroke in patients without
diabetes, compared with non-lacunar stroke.
Cheung 392 ischemic stroke Fractal dimension Increased fractal dimension associated with lacunar stroke (OR 4.27
et al., [1.49, 12.17]), compared with non-lacunar stroke.
2010b
Doubal 86 lacunar stroke, and 80 with cortical stroke Fractal dimension Decreased monofractal dimension (b ¼ 0.01, p < 0.001) and multifractal
et al., dimension (b ¼ 0.011, p ¼ 0.002) associated with lacunar stroke,
2010b compared with cortical stroke.
Doubal 104 lacunar stroke, and 101 cortical stroke Branching angle and Increased arteriolar branching coefficient associated with periventricular
et al., coefficient white matter hyperintensities (b ¼ 0.072, p ¼ 0.006), and decreased
2010a arteriolar branching coefficient associated with deep white matter
hyperintensities (b ¼ 0.076, p ¼ 0.003), but not with stroke subtype.
Baker et al., 51 deep intracerebral hemorrhage, 93 lacunar infarction, Retinopathy, arteriolar FAN (OR 3.7 [1.8,7.6]), AV-nicking (OR 2.6 [1.4, 5.1]), decreased CRAE
2010a and 486 nonlacunar infarction wall signs, CRAE, CRVE (p < 0.001) and increased CRVE (p ¼ 0.03) associated with deep
intracerebral hemorrhage, compared with nonlacunar infarction.
Baker et al., 25 lobar intracerebral hemorrhage, 51 deep Retinopathy, arteriolar Retinopathy associated with lobar intracerebral hemorrhage, compared
2010b intracerebral hemorrhage, 93 lacunar infarction, and wall signs with lacunar (OR 2.8 [1.0,7.6]) or nonlacunar cerebral infarction (OR 3.0
486 nonlacunar cerebral infarction [1.3,6.9]).
Baker et al., 1360 acute stroke (203 cortical atrophy, 144 subcortical Retinopathy, arteriolar Retinopathy (OR 1.9 [1.2, 3.0]) and enhanced arteriolar light reflex (OR
2010c atrophy, 424 both atrophy subtypes) wall signs, CRAE, CRVE 2.0 [1.2, 3.2]) associated with subcortical atrophy.
retinopathy
Wang et al., 693 TIA or acute stroke, and 3384 control subjects Retinopathy, arteriolar Retinopathy (non-diabetic retinopathy: OR 2.6 [1.4,4.6]; diabetic
2011 wall signs retinopathy: OR 3.4 [1.2, 10.0]), FAN (OR 8.9 [4.5, 17.6]), AV-nicking (OR
1.9 [1.1, 3.7]), enhanced arteriolar light reflex (OR 4.1 [1.9, 9.1])
associated with TIA, compared with controls. Similar findings observed in
patients with stroke. Patients with TIA and ischemic stroke had similar
prevalence of retinal microvascular signs.
Ong et al., 557 ischemic stroke, and 557 control subjects CRAE, CRVE, tortuosity, Decreased arteriolar (OR 2.28 [1.80, 2.87]) and venular (OR 1.80 [1.46,
2013 fractal dimension 2.23]) fractal dimensions, increased arteriolar (OR 1.56 [1.25, 1.95]) and
venular (OR 1.49 [1.27, 1.76]) tortuosity, decreased CRAE (OR 2.79 [2.21,
3.53]) and increased CRVE (OR 1.57 [1.27, 1.95]) associated with ischemic
stroke, compared with controls.
Similar findings among stroke subtypes.
Wang et al., 152 ischemic stroke, and 2890 control subjects Localized RNFL defects Presence of localized RNFL defects associated with acute (OR 6.23 [4.17,
2014 9.30]) or previous (OR 1.48 [1.02, 2.16]) stroke.
Vuong et al., 257 subjects with focal neurologic deficits Retinopathy, arteriolar Focal and general arteriolar narrowing associated with cerebrovascular
2015 wall signs disease (OR 8.11 [1.82, 36.09], OR 2.39 [1.01, 5.68], respectively) and TIA
(OR 5.19 [1.27, 21.22], OR 2.37 [1.02, 5.54], respectively).
Hughes 1185 subjects from a UK-based triethnic population- Retinopathy, diameter, A higher retinopathy grade (OR 2.79 [1.30, 5.99]), decreased arteriolar
et al., based study bifurcation, tortuosity diameter (OR 0.97 [0.96, 0.99]), fewer arteriolar bifurcations (OR 0.82
2016 [0.73, 0.93]), increased venular tortuosity (OR 1.16 [1.05, 1.29]) and
decreased AVR (OR 0.23 [0.06, 0.85]) associated with strokes/infarcts.
Prospective studies
Wong et al., 10358 subjects from the Atherosclerosis Risk in Retinopathy, arteriolar Retinopathy (RR 2.58 [2.59, 4.20]), AVN (RR 1.60 [1.03, 2.47]) and
2001 Communities study wall signs, AVR decreased AVR (test of trend, p < 0.03) associated with incident stroke.
Wong et al., 1684 subjects from the Atherosclerosis Risk in Retinopathy, arteriolar Retinopathy (RR 4.9 [2.0, 11.9]) associated with incident stroke. Persons
2002 Communities study wall signs, AVR with both WMLs and retinopathy had higher cumulative incidence of
stroke than those without either WMLs or retinopathy (RR 18.1
[5.9,55.4]).
Mitchell 3047 subjects from the Blue Mountains Eye Study Retinopathy, arteriolar Retinopathy associated with stroke event (RR 1.7 [1.0, 2.8]) in persons
et al., wall signs without diabetes.
2005
Wong et al., 1992 subjects from the cardiovascular health study CRAE, CRVE Increased CRVE associated with incident stroke (RR 2.2 [1.1, 4.3]).
2006
Ikram et al., 5540 subjects from the Rotterdam study CRAE, CRVE Increased CRVE associated with incident stroke (HR 1.12 [1.02, 1.24]) and
2006a cerebral infarction (HR 1.15 [1.02, 1.29]).
Witt et al., 132 subjects who dies from stroke or ischemic heart CRAE, CRVE, tortuosity Increased arteriolar length-to-diameter ratio associated with increased
2006 disease and 528 control subjects from the Beaver Dam and bifurcation risk of stroke mortality (b coefficient ¼ 0.075, p ¼ 0.02).
Eye Study
Wang et al., 7497 subjects from the Beaver Dam Eye Study and the CRAE, CRVE Decreased CRAE (HR 1.64 [1.00, 2.67]) and increased CRVE (HR 1.53 [0.94,
2007 Blue Mountains Eye Study 2.47] associated with increased risk of stroke mortality aged 43e69.
Yatsuya 10496 subjects from the Atherosclerosis Risk in Retinopathy, arteriolar Decreased CRAE (HR 1.67 [1.23, 2.26]), increased CRVE (HR 1.44 [1.09,
et al., Communities Study wall signs, CRAE, CRVE 1.91]), FAN (HR 2.22 [1.11, 4.48]), and AV-nicking (HR 2.38 [1.20e4.71])
2010 associated with incident lacunar stroke. Retinopathy associated with
nonlacunar thrombotic (HR 2.41 [1.47, 3.95]) and cardioembolic (HR 2.25
[1.09,4.65]) stroke incidence.
5518 subjects from the Rotterdam study CRAE, CRVE
(continued on next page)

Please cite this article in press as: Cheung, C.Y.-l., et al., Imaging retina to study dementia and stroke, Progress in Retinal and Eye Research (2016),
http://dx.doi.org/10.1016/j.preteyeres.2017.01.001
10 C.Y.-l. Cheung et al. / Progress in Retinal and Eye Research xxx (2016) 1e19

Table 3 (continued )

Authors, Sample Retinal measure Summary of results

year

Wieberdink Increased CRVE associated with incident stroke (HR 1.20 [1.09, 1.33]),
et al., cerebral infarction (HR 1.28 [1.13, 1.46]) and intracerebral hemorrhage
2010 (HR 1.53 [1.09, 2.15]).
Kawasaki 104 stroke cases and 208controls from the Blue Fractal dimension Decreased fractal dimension associated with incident stroke (OR 2.43
et al., Mountains Eye Study [1.04, 5.68]).
2011
Kawasaki 4849 subjects from the Multi-Ethnic Study of Retinopathy, CRAE, Retinopathy (RR 2.96 [1.50, 5.84]) and decreased CRAE (RR 2.83 [1.34,
et al., Atherosclerosis CRVE 5.95]) associated with incident stroke.
2012
Cheung 3189 subjects from Singapore Malay Eye Study Retinopathy, CRAE, Retinopathy (HR 1.94 [1.01, 3.72]) and increased CRVE (HR 3.28 [1.30,
et al., CRVE, tortuosity, 8.26]) associated with incident stroke.
2013 branching angle,
fractal dimension

AV-nicking ¼ arteriovenous nicking; AVR ¼ arterio-venous ratio; CRAE ¼ central retinal artery equivalent; CRVE ¼ central retinal vein equivalent; RNFL ¼ retinal nerve fiver
layer; TIA ¼ transient ischemic attack; WML ¼ white matter lesion.

6. Retinal changes and subclinical cerebral small vessel complementary approach to these brain imaging techniques, and
disease has substantial potential in clinical and research settings, including:
First, retinal imaging may be employed as a risk stratification
Cerebral small vessel disease, a collective term used to denote a tool. While classical risk factors (e.g. elevated serum cholesterol
heterogeneous disease group of cerebral lesions such as WML, level and blood pressure) have been used to predict the risk of both
lacunar infarcts and microbleeds, is associated with increased risk dementia and stroke, it is clear these risk factors do not fully explain
of stroke, cognitive decline and dementia (Pantoni, 2010). Several a person's risk of these neurological diseases. Studies suggest the
clinic-based and population-based studies have shown associations addition of retinal measures improves the prediction of stroke
between retinal vascular changes and markers of cerebral small (overall improvement of about 10% beyond that of established risk
vessel disease (Table 4). Lacunar infarcts, have been linked to factors) (Cheung et al., 2013; McGeechan et al., 2009). Although this
different retinal vascular changes including retinopathy, retinal is only a modest improvement in prediction, these findings suggest
arteriolar wall signs, retinal arteriolar narrowing, retinal venular the addition of combinations of several retinal features and/or
widening and increased retinal vascular fractal dimension, both in functional retinal parameter (i.e. “multiple markers approach”)
population-based studies (Cheung et al., 2010c; Cooper et al., 2006; may lead to further improvement in the prediction of dementia and
Ikram et al., 2006b; Kwa et al., 2002) and in patients with stroke stroke, or to identify a more specific subgroup of patients who
(Cheung et al., 2010b; Liew et al., 2014; Lindley et al., 2009). could benefit from more intensive and expensive investigations
Numerous cross-sectional and prospective studies have also shown (e.g., cranial MRI).
that WML and WML progression were associated with retinopathy, Second, retinal imaging may be of value as a surrogate
retinal arteriolar wall signs, changes in retinal vascular caliber biomarker to evaluate the effectiveness of treatments and monitor
(retinal arteriolar narrowing and venular widening), and subopti- treatment outcomes. Retinal vascular imaging has been used to
mal retinal bifurcation (Baker et al., 2010c; Cheung et al., 2010c; examine treatment effects of anti-hypertensive treatment, showing
Doubal et al., 2010a; Doubal et al., 2010b; Haan et al., 2012; that reduction in blood pressure results in regression of retinal
Hughes et al., 2016; Ikram et al., 2006b; Longstreth et al., 2007; vascular signs (Hughes et al., 2008; Thom et al., 2009). While there
Qiu et al., 2009; Tirsi et al., 2009; Wong et al., 2002). Cerebral have been no major interventional studies using retinal imaging
microbleeds have been suggested as another manifestation or changes as surrogate outcome measures in dementia and stroke,
marker of cerebral small vessel disease. A study has reported that this approach has substantial potential. For example, a randomized
decreased retinal arteriolar fractal dimension and increased arte- trial of aspirin and cognitive decline is currently using retinal im-
riolar tortuosity were associated with cerebral microbleeds, inde- aging as a surrogate marker of response (Reid et al., 2012). Addi-
pendent of vascular risk factors and other cerebral markers such as tionally, retinal imaging may also be valuable as a research or
WML and lacunes, suggesting that these retinal vascular measures clinical tool to investigate other major cerebral and neurological
may be an early manifestation of cerebral small vessel disease (Hilal diseases, such as multiple sclerosis (Bhaduri et al., 2016; Dhillon
et al., 2014). Fig. 5 demonstrates an example of retinal vascular and Dhillon, 2008; Gugleta et al., 2009; Petzold et al., 2010;
changes in a patient with cerebral small vessel disease on MRI. In Saidha et al., 2012), depression (Meier et al., 2014; Yildiz et al.,
addition to retinal vascular changes, presence of RNFL defects from 2016) and schizophrenia (Celik et al., 2016; Meier et al., 2013). For
fundus photographs were associated with cerebral small vessel example, it is shown that measurement from OCT may be a treat-
disease, particularly with WML, suggesting that RNFL defects may ment outcome measure for multiple sclerosis (Petzold et al., 2010).
share a common presumably ischemic pathophysiology with ce- Finally, retinal imaging may also be a screening tool to identify
rebral small vessel disease (Kim et al., 2011b). asymptomatic subjects for recruitment into clinical trials testing
new investigational treatments for dementia or stroke (e.g. anti-
7. Potential clinical and research use amyloid therapy to delay cognitive decline) (Sperling et al., 2014).

As this review has shown, retinal imaging techniques offer 8. Limitations of current studies
unique information on the status of the microvasculature and
neuronal structure distinct from current neuro-imaging (e.g., brain There remain, nevertheless, major challenges to translate cur-
MRI) and systemic (e.g. blood pressure, intima media thickness) rent research and there are limitations that the current studies have
markers. While retinal imaging cannot fully replace MRI or PET not addressed. First, there are few prospective studies available to
imaging for disease diagnosis, retinal imaging offers a demonstrate consistent predictive ability of retinal imaging

Please cite this article in press as: Cheung, C.Y.-l., et al., Imaging retina to study dementia and stroke, Progress in Retinal and Eye Research (2016),
http://dx.doi.org/10.1016/j.preteyeres.2017.01.001
 C.Y.-l. Cheung et al. / Progress in Retinal and Eye Research xxx (2016) 1e19 11

Table 4
Recent clinical studies of small vessel diseases and retinal measures, as defined from ocular fundus photography.

Authors, Sample Retinal measure Summary of results

year

Kwa et al., 179 subjects with symptomatic atherosclerotic Retinopathy, arteriolar Retinal exudates associated with lacunar infarcts (b ¼ 0.15, p ¼ 0.04). Retinal
2002 disease wall signs arteriolar narrowing (Beta-coefficient ¼ 0.16, p ¼ 0.03) and sclerosis (b ¼ 0.28,
p ¼ 0.001) associated with WML.
Wong et al., 1684 subjects from the Atherosclerosis Risk in Retinopathy, arteriolar Retinopathy (OR 2.5 [1.5,4.0]), AV-nicking (OR 2.1 [1.4, 3.2]) and FAN (OR 2.1 [1.4,
2002 Communities study wall signs, CRAE, CRVE 3.1]) associated with presence of WML.
Cooper 1684 subjects from the Atherosclerosis Risk in Retinopathy, arteriolar AV-nicking (OR 2.47 [1.43,4.29]), FAN (OR 1.78 [1.00,3.19]) and decreased AVR (OR
et al., Communities study wall signs, AVR 4.76 [1.68,13.47]) associated with infarct, particular in persons with hypertension.
2006
Ikram et al., 490 subjects from the Rotterdam Scan Study CRAE, CRVE Increased CRVE associated with marked periventricular (OR 1.74 [1.02, 2.95]) and
2006b subcortical WML progression (OR 2.50 [1.30, 4.81]) and incident lacunar infarct
(OR 1.24 [0.72, 2.12]).
Longstreth 1717 subjects from the Cardiovascular Health Retinopathy, arteriolar Decreased AVR associated with prevalent (OR 1.18 [1.05, 1.34]) and incident
et al., Study wall signs, CRAE, CRVE, infarct (OR 1.26 [1.09, 1.46]). Decreased CRAE associated with worsening WML
2007 AVR grade (OR 1.17 [1.03, 1.34]). AV-nicking associated with prevalent (OR 1.84 [1.23,
2.76]) and incident infarct (OR 1.84 [1.15, 2.94]).
Qiu et al., 4176 subjects from the AGES-Reykjavik Study Retinopathy, arteriolar FAN associated with subcortical (OR 1.40 [1.09, 1.79]) and periventricular WML
2009 wall signs (OR 1.40 [1.12, 1.74]). AV-nicking associated with subcortical (OR 1.64 [1.36, 1.97])
and periventricular WML (OR 1.54 [1.31, 1.81]).
Tirsi et al., 46 non-DR subjects AVR Decreased AVR associated with higher white matter hyperintensity score
2009 (b ¼ 0.370, p ¼ 0.016).
Lindley 1321 acute stroke subjects from the Multi- Retinopathy, arteriolar FAN (OR 2.26 [1.25, 4.08]), AV-nicking (OR 1.98 [1.23, 3.18]), enhanced arteriolar
et al., Center Retina and Stroke Study wall signs, CRAE, CRVE light reflex (OR 2.07 [1.31, 3.27]) decreased CRAE (OR 2.40 [1.00, 5.76]) associated
2009 with lacunar stroke. Increased CRVE associated with lacunar stroke in persons
without diabetes (OR 2.08 [1.10, 3.95]).
Cheung 810 subjects from the Atherosclerosis Risk in Retinopathy, arteriolar Retinopathy (OR 3.19 [1.56, 6.50]) and AV-nicking (OR 2.48 [1.39, 4.40]) associated
et al., Communities Study wall signs with incident lacunar infarct. AV-nicking associated with incident WML (OR 2.12
2010c [1.18, 3.81]) and WML progression (OR 2.22 [1.00, 5.88]).
Doubal 205 stroke subjects Bifurcation Increased retinal arteriolar branching coefficient associated with increased
et al., periventricular white matter hyperintensities in patients with mild stroke
2010a (b ¼ 0.072, p ¼ 0.006).
Cheung 392 acute stroke subjects from the Multi- Fractal dimension Increased retinal vascular fractal dimension associated with lacunar stroke (OR
et al., Center Retinal Stroke study 1.85 [1.20, 2.84]).
2010b
Wieberdink 5518 subjects from the Rotterdam Study CRAE, CRVE Increased CRVE associated with incident stroke (HR 1.20 [1.09, 1.33]), cerebral
et al., infarction (HR 1.28 [1.13, 1.46]) and intracerebral hemorrhage (HR 1.53 [1.09,
2010 2.15]).
Doubal 166 stroke subjects Fractal dimension Decreased retinal vascular fractal dimension associated with lacunar stroke
et al., (b ¼ 0.011, p ¼ 0.002).
2010b
Liew et al., 1211 acute stroke subjects from the Multi- Retinopathy, arteriolar Retinopathy associated with WML (OR 1.53 [1.10, 2.13]). Enhanced arteriolar light
2014 Center Retinal Stroke study wall signs, CRAE, CRVE reflex associated with WML (OR 1.61 [1.11, 2.33]) and lacunar infarcts (OR 1.55
[1.01, 2.37]). Increased CRAE associated with lacunar infarcts (OR 1.61 [1.01, 2.57]).
Haan et al., 511 subjects from the Women's Health Retinopathy Retinopathy associated with increased ischemic lesion volumes in the total brain
2012 Initiative Memory Study and the Sight (47% larger, p ¼ 0.04).
Examination Study
Hilal et al., 261 subjects from the Epidemiology of CRAE, CRVE, fractal Decreased CRAE (OR 2.10 [1.06, 4.15]), increased CRVE (OR 2.29 [1.19, 4.40]) and
2014 Dementia in Singapore Study dimension, tortuosity decreased retinal arteriolar fractal dimension (OR 1.89 [1.27, 2.82]) associated with
presence of multiple cerebral microbleeds.
Hughes 1185 subjects from a UK-based triethnic Retinopathy, diameter, A higher retinopathy grade (OR 2.79 [1.30, 5.99]), associated with strokes/infarcts.
et al., population-based study bifurcation, tortuosity Decreased arteriolar diameter (OR 0.97 [0.97, 0.99], 0.98 [0.97, 0.99]), fewer
2016 arteriolar bifurcations (OR 0.82 [0.73, 0.93], 0.88 [0.80, 0.96]), increased venular
tortuosity (OR 1.16 [1.05, 1.29], 1.11 [1.02, 1.21]) and decreased AVR (OR 0.23
[0.06, 0.85], 0.31 [0.12, 0.84]) associated with strokes/infarcts and white matter
hyperintensities, respectively.

AV-nicking ¼ arteriovenous nicking; AVR ¼ arterio-venous ratio; b ¼ beta coefficient; CRAE ¼ central retinal artery equivalent; CRVE ¼ central retinal vein equivalent;
FAN ¼ focal arteriolar narrowing; HR ¼ hazard ratio; OR ¼ odds ratio; RR ¼ relative risk; WML ¼ white matter lesion.

measures with risk or progression of dementia and stroke. The parameters (e.g., fractal dimension) and new SD-OCT measures
strength of associations seen in the few longitudinal studies to date have not been examined in prospective studies, with most studies
between retinal vascular changes and risk of dementia or stroke has reporting only cross-sectional associations in small clinical series.
also been relatively modest (relative risks or hazard ratios ranging Publication bias in the literature may also be present (i.e., only
from 1.1 to 3.3, as shown in Tables 1 and 3). There are also very few positive association studies have been reported), since there is lack
studies which have specifically evaluated and demonstrate incre- of pre-study publications of protocols for the studies and trials
mental benefit and clinical utility of adding retinal imaging to discussed (Chalmers and Nylenna, 2014).
existing methods or models to predict dementia or stroke (Cheung Third, co-morbidities such as hypertension and diabetes are
et al., 2013; McGeechan et al., 2009). Further prospective studies, important confounders since most of the studies have been per-
including clinical trials and cost-effectiveness studies are clearly formed on elderly subjects with a high proportion of these
needed to address this gap (McGeechan et al., 2008; Wang, 2011). comorbidities (Feigin et al., 2016; Gorelick et al., 2011). In the vast
Second, several of the newer retinal vascular network majority of these studies; however, the associations between

Please cite this article in press as: Cheung, C.Y.-l., et al., Imaging retina to study dementia and stroke, Progress in Retinal and Eye Research (2016),
http://dx.doi.org/10.1016/j.preteyeres.2017.01.001
12 C.Y.-l. Cheung et al. / Progress in Retinal and Eye Research xxx (2016) 1e19

Fig. 5. An example of retinal vascular changes in a patient with cerebral small vessel disease on magnetic resonance imaging (MRI). (A) Retinal photograph showing narrowed
retinal arterioles (grey arrows), widened venules (black arrows) and tortuous venules (white arrows). (B) MRI images showing white matter lesions (white arrows), lacunes (yellow
arrow) and microbleeds (red arrows). FLAIR ¼ fluid attenuated inversion recovery; SWI ¼ susceptibility weighted image.

retinal vascular changes and dementia or stroke have remained the outer retina including retinal pigmented epithelium and pho-
significant while adjusting for these systemic risk factors, sug- toreceptors, and also to the prelaminar portion of the optic nerve.
gesting that the associations between retinal vascular changes and Enhanced depth imaging with SD-OCT, by positioning the choroid-
dementia/stroke are independent of these risk factors. Neverthe- scleral interface adjacent to the zero delay during imaging, provides
less, retinal vascular changes are common in elderly persons and visualization of the choroidal anatomy, and improve the under-
these may be related to concurrent co-morbidities such as diabetes standing of the contribution of choroidal morphology to eye dis-
(e.g., diabetic retinopathy) and hypertension (e.g., retinal vascular eases (Mrejen and Spaide, 2013). Two recent case-control studies
occlusion), and may not be specific markers of dementia or stroke. have observed reduction in choroidal thickness using enhanced
Finally, a number of logistic and technical issues also need to be depth imaging with SD-OCT in patients with Alzheimer's disease
considered. While retinal imaging is a non-invasive imaging mo- without eye diseases (e.g. age-related macular degeneration,
dality which is easily performed in the elderly and is already widely glaucoma, except cataract) (Bayhan et al., 2015; Gharbiya et al.,
used clinically to screen and diagnose common eye diseases, sub- 2014). These studies suggest that choroidal thinning, probably
jects do need to be able to sit upright to undergo these examina- indicating an abnormal choroidal blood supply associated with
tions and understand instructions. Hence, subjects with major hypoperfusion or atrophic changes, may be related to a series of
disabilities, such as hemiplegia or severe dementia, and are unable pathologic events (e.g. inflammatory cascade) triggered by amy-
to sit upright and/or understand instructions, cannot undergo loid-b deposition in the brain (Gharbiya et al., 2014). In addition, it
retinal imaging examinations. Other ocular factors such as poor has been hypothesized that Alzheimer's disease and age-related
pupil dilation, cataract and other media opacities may also interfere macular degeneration share common pathogenic pathways, as
with the quality of images obtained with retinal imaging (Hardin amyloid-b has also been found in drusen deposits that are seen in
et al., 2015; Kim et al., 2012; Li et al., 2010; Maberley et al., 2004; age-related macular degeneration (Anderson et al., 2004). Recent
Rochtchina et al., 2008; Wainwright et al., 2010). studies have reported that eyes with age-related macular degen-
eration have reduced choroidal thickness (Yiu et al., 2015). Imaging
9. Recent new advances in retinal imaging of the choroid thus opens another area to study the eye-brain link.
The latest generation of OCT, swept-source OCT, which uses a
9.1. Automated retinal vascular imaging software continuous and repetitively tunable (or “swept”) light source over a
broad wavelength range, allows a faster scanning speed (100,000
Currently available software programs for measuring retinal A-scans/sec) and better penetration (1,050 nm wave length) for
vascular changes require extensive training, and additional input by visualization of the choroid than even SD-OCT. Detailed choroidal
technicians and hence are not fully automated. With respect to vasculature can also be imaged easily using en face imaging with
quantitative features such as calibers, tortuosity, network swept-source OCT (Ferrara et al., 2015) that offers a novel
complexity, newer software programs are being developed, that perspective to study its relationship to dementia and stroke. The
can automatically measure changes in these retinal features over rapid scan acquisition speed of swept-source OCT also allows better
time by comparing two retinal images taken at different time patient cooperation from patients with dementia.
points (Cavallari et al., 2015; Cheung and Wong, 2012; Joshi et al.,
2014). Fully automated detection of retinopathy signs will also 9.3. Retinal fundus autofluorescence imaging
facilitate a more efficient retinopathy assessment (Abramoff et al.,
2016; Nguyen et al., 2013; Walton et al., 2016). Fundus autofluorescence imaging is an in vivo imaging method
for metabolic mapping of naturally or pathologically occurring
9.2. New optical coherence tomography instruments fluorophores in the ocular fundus for diagnosing various retinal
disorders. There is growing evidence to show that ex vivo retinal
The choroid, a vascular layer of the eye accounting for seventy amyloid-b plaques are present and appear hyper-autofluorescent
percent of the blood flow from the eye, contributes blood supply to under specific wavelengths in eyes from patients with

Please cite this article in press as: Cheung, C.Y.-l., et al., Imaging retina to study dementia and stroke, Progress in Retinal and Eye Research (2016),
http://dx.doi.org/10.1016/j.preteyeres.2017.01.001
 C.Y.-l. Cheung et al. / Progress in Retinal and Eye Research xxx (2016) 1e19 13

Alzheimer's disease (Koronyo-Hamaoui et al., 2011) and from 10. Future directions
transgenic mouse model of Alzheimer's disease (Hwang et al.,
2011). In addition, it was recently demonstrated that curcumin- This review has highlighted several gaps which are areas for
labeled amyloid-b plaques can be imaged in vivo in retina of future research. First, the biological mechanisms which underlie
transgenic Alzheimer's disease mice (Koronyo-Hamaoui et al., the associations of retinal changes with dementia and stroke are
2011). In a case report, hyper-autofluorescence deposits was not completely understood. Experimental research using animal
observed in transthyretin-related familial amyloid polyneuropathy models is thus essential to investigate the specific underlying
using retinal fundus autofluorescence imaging, demonstrating that pathophysiological mechanisms to further understanding of the
retinal fundus autofluorescence imaging may be useful in detecting etiology, which in turn will support how retinal imaging can be
ocular amyloidosis (Veronese et al., 2013). It has also been shown translated into a useful tool in clinical practice and facilitate the
that parameters from retinal fluorescence imaging correlated with development of novel therapies. Second, future studies with lon-
Alzheimer-specific marker (p-tau181-protein concentration in ce- gitudinal data and systematic review (Chalmers and Nylenna,
rebrospinal fluid) in a pilot study (Jentsch et al., 2015). Therefore, 2014), will be important to clarify the role of retinal imaging in
these evolving retinal imaging techniques might potentially offer a reflecting pathological microvascular dysfunction and neuronal
new means to detect Alzheimer-specific pathology in vivo. loss in dementia and stroke, and to develop risk prediction algo-
rithms. For example, given the complexity of dementia (i.e. few
patients with “pure” Alzheimer's disease, many with mixed etiol-
9.4. Dynamic retinal vessel analysis ogy such as small vessel disease and Alzheimer's disease), baseline
retinal vascular imaging (to identify those with small vessel dis-
It has been proposed that endothelial dysfunction is one of the ease) and OCT findings (to identify those with neuronal/axonal
earliest events in the development of vascular diseases, including atrophy) could help to determine likely underlying pathologies
cerebral small vessel disease and stroke (Hassan et al., 2004). In the contributing to dementia development from future longitudinal
retina, the endothelium produces vasoactive factors, especially ni- studies. Additional studies are also required to further elucidate the
tric oxide, in response to increased activity of the neural retina due clinical incremental usefulness (e.g. how it may influence clinical
to a stimulus such as flickering light. This resulting vasodilatory decision-making, estimation the cost-effectiveness and accept-
response of retinal vessels has been suggested as a measure of ability to patients). Third, a growing body of evidence shows that
endothelial dysfunction. This impaired flicker light-induced vaso- neurons, glial cells, and the cerebral microvascular endothelium
dilation may reflect either direct damage to the microcirculation or operate as an integrated unit, playing a crucial role in this process,
even neurodegeneration as retinal blood flow is coupled with and disruption to the “neurovascular unit” has been implicated in
neuronal activity (Lim et al., 2013). The flicker-induced response conditions such as ischemic stroke and Alzheimer's disease
can now be assessed in real-time using the Dynamic Vessel (Girouard and Iadecola, 2006; Iadecola, 2004). The retina may be a
Analyzer (IMEDOS, Jena, Germany). Using this technique, it has potential site to study interactions between neuronal, glial and
been reported that flicker light-induced vasodilation was reduced vascular compartments in the brain (neurovascular unit). Fourth,
in persons with diabetes and with increasing severity of diabetic there are new retinal imaging technologies now being studied to
retinopathy (Cheung et al., 2015b). Extending the use of the DVA to further measure and analyze the detailed structure and functions of
the study of cerebral small vessels, changes in vasomotor reactivity the retina, including non-dye-based mapping of retinal capillary
of the retinal vessels may potentially reflect similar changes in the network by OCT-angiography (Fig. 6) (Jia et al., 2015), oxygen
cerebral microvasculature. saturation by retinal oximetry (Einarsdottir et al., 2015; Yip et al.,
2014), and blood flow by Doppler OCT (Leitgeb et al., 2014), offer-
ing a new means of studying dementia and stroke. Finally, studies
9.5. Adaptive optics imaging of the retina should determine how possible ocular factors and confounders
should be controlled for in the further development of retinal im-
Adaptive optics, derived from astronomical research tools, is a aging for clinical use.
technology used to improve the performance of optical systems by
reducing the effects of optical aberrations. Currently systems such 11. Conclusions
as retinal fundus camera (Liang et al., 1997; Rha et al., 2006), OCT
(Cense et al., 2009; Kocaoglu et al., 2011) and scanning laser Retinal imaging is an exciting and highly promising method to
ophthalmoscope (Dubra and Sulai, 2011; Roorda et al., 2002) study dementia and stroke. Studies to date have now shown a
employing adaptive optics can achieve resolutions in the order of pattern of associations between retinal changes with these dis-
2 mm in the human retina resulting in the direct visualization of eases, suggesting that retinal imaging can provide new and
capillaries and nerve fiber bundles (Lombardo et al., 2013). The potentially important insights into neurodegenerative and cere-
application of adaptive optics in retinal imaging promises in vivo brovascular processes. Further research will be needed to clarify the
detection of degenerative and vascular processes at a cellular level underlying pathophysiological mechanisms of these retinal find-
(Chui et al., 2013). Parameters that can be measured are wall-to- ings, and the utility of retinal imaging before it can be fully trans-
lumen ratios of retinal microvessels, which is a hallmark of lated as a useful tool in clinical practice.
microangiopathy and is predictive of end-organ damage (Koch
et al., 2014). Currently only a few adaptive optics systems exist Conflict of interest statement
worldwide, but with the recent appearance of commercial in-
struments there will be more widespread testing of this technology We declare no competing interests.
in the pursuit of identifying subclinical changes that may become
early biomarkers. Besides the application of adaptive optics in the Review criteria
study of ocular diseases such as glaucoma (Hasegawa et al., 2016),
age-related macular degeneration (Zayit-Soudry et al., 2013) and We searched Medline and PubMed for reports published in
diabetic retinopathy (Soliman et al., 2016), there is potential for English, with no limitations on the date of publication. The search
adaptive optics techniques in the study of dementia and stroke. terms “stroke”, “dementia”, “retinal photographs”, “retinal

Please cite this article in press as: Cheung, C.Y.-l., et al., Imaging retina to study dementia and stroke, Progress in Retinal and Eye Research (2016),
http://dx.doi.org/10.1016/j.preteyeres.2017.01.001
14 C.Y.-l. Cheung et al. / Progress in Retinal and Eye Research xxx (2016) 1e19

Fig. 6. (A) Detailed retinal (superficial capillary plexus, deep capillary plexus and outer retina) and choroidal (choriocapillaris) vasculature at the macula from a human subject
imaged by optical coherence tomography-angiography. (B) New image processing and quantification methods for assessment of capillary network are currently under development.

Please cite this article in press as: Cheung, C.Y.-l., et al., Imaging retina to study dementia and stroke, Progress in Retinal and Eye Research (2016),
http://dx.doi.org/10.1016/j.preteyeres.2017.01.001
 C.Y.-l. Cheung et al. / Progress in Retinal and Eye Research xxx (2016) 1e19 15

microvasculature”, “retinopathy”, “optical coherence tomography”, over an extended area. Microcirculation 17, 495e503.
Cheung, C.Y., Zheng, Y., Hsu, W., Lee, M.L., Lau, Q.P., Mitchell, P., Wang, J.J., Klein, R.,
“retinal ganglion cells”, “retinal nerve fiber layer” and “optic nerve”
Wong, T.Y., May 2011a. Retinal vascular tortuosity, blood pressure, and car-
were used. The final reference list was generated on the basis of diovascular risk factors. Ophthalmology 118 (5), 812e818.
relevance to the topics covered in this Review. Cheung, N., Liew, G., Lindley, R.I., Liu, E.Y., Wang, J.J., Hand, P., Baker, M., Mitchell, P.,
Wong, T.Y., 2010b. Retinal fractals and acute lacunar stroke. Ann. Neurol. 68,
107e111.
References Cheung, N., Mosley, T., Islam, A., Kawasaki, R., Sharrett, A.R., Klein, R., Coker, L.H.,
Knopman, D.S., Shibata, D.K., Catellier, D., Wong, T.Y., 2010c. Retinal microvas-
Abramoff, M.D., Lou, Y., Erginay, A., Clarida, W., Amelon, R., Folk, J.C., Niemeijer, M., cular abnormalities and subclinical magnetic resonance imaging brain infarct: a
2016. Improved automated detection of diabetic retinopathy on a publicly prospective study. Brain 133, 1987e1993.
available dataset through integration of deep learning. Invest Ophthalmol. Vis. Cheung, C.Y., Tay, W.T., Mitchell, P., Wang, J.J., Hsu, W., Lee, M.L., Lau, Q.P., Zhu, A.,
Sci. 57, 5200e5206. Klein, R., Saw, S.M., Wong, T.Y., Jul 2011b. Quantitative and qualitative retinal
Alpers, B.J., Forster, F.M., Herbut, P.A., 1948. Retinal, cerebral and systemic arterio- microvascular characteristics and blood pressure. J. Hypertens. 29 (7),
sclerosis; a histopathologic study. Arch. Neurol. Psychiatry 60, 440e456. 1380e1391.
Alzheimer's, A., 2016. 2016 Alzheimer's disease facts and figures. Alzheimers Cheung, C.Y., Ikram, M.K., Sabanayagam, C., Wong, T.Y., 2012a. Retinal microvascu-
Dement. 12, 459e509. lature as a model to study the manifestations of hypertension. Hypertension 60,
Anderson, D.H., Talaga, K.C., Rivest, A.J., Barron, E., Hageman, G.S., Johnson, L.V., 1094e1103.
2004. Characterization of beta amyloid assemblies in drusen: the deposits Cheung, C.Y., Thomas, G.N., Tay, W., Ikram, M.K., Hsu, W., Lee, M.L., Lau, Q.P.,
associated with aging and age-related macular degeneration. Exp. Eye Res. 78, Wong, T.Y., 2012b. Retinal vascular fractal dimension and its relationship with
243e256. cardiovascular and ocular risk factors. Am. J. Ophthalmol. 154 (663e674), e661.
Ascaso, F.J., Cruz, N., Modrego, P.J., Lopez-Anton, R., Santabarbara, J., Pascual, L.F., Cheung, C.Y., Tay, W.T., Ikram, M.K., Ong, Y.T., De Silva, D.A., Chow, K.Y., Wong, T.Y.,
Lobo, A., Cristobal, J.A., 2014. Retinal alterations in mild cognitive impairment 2013. Retinal microvascular changes and risk of stroke: the Singapore Malay Eye
and Alzheimer's disease: an optical coherence tomography study. J. Neurol. 261, Study. Stroke 44, 2402e2408.
1522e1530. Cheung, C.Y., Ong, S., Ikram, M.K., Ong, Y.T., Chen, C.P., Venketasubramanian, N.,
Baker, M.L., Marino Larsen, E.K., Kuller, L.H., Klein, R., Klein, B.E., Siscovick, D.S., Wong, T.Y., 2014a. Retinal vascular fractal dimension is associated with cogni-
Bernick, C., Manolio, T.A., Wong, T.Y., 2007. Retinal microvascular signs, cogni- tive dysfunction. J. Stroke Cerebrovasc. Dis. 23, 43e50.
tive function, and dementia in older persons: the Cardiovascular Health Study. Cheung, C.Y., Ong, Y.T., Ikram, M.K., Chen, C., Wong, T.Y., 2014b. Retinal microvas-
Stroke 38, 2041e2047. culature in Alzheimer's disease. J. Alzheimers Dis. 42 (Suppl. 4), S339eS352.
Baker, M.L., Hand, P.J., Liew, G., Wong, T.Y., Rochtchina, E., Mitchell, P., Lindley, R.I., Cheung, C.Y., Ong, Y.T., Ikram, M.K., Ong, S.Y., Li, X., Hilal, S., Catindig, J.A.,
Hankey, G.J., Wang, J.J., 2010a. Retinal microvascular signs may provide clues to Venketasubramanian, N., Yap, P., Seow, D., Chen, C.P., Wong, T.Y., 2014c.
the underlying vasculopathy in patients with deep intracerebral hemorrhage. Microvascular network alterations in the retina of patients with Alzheimer's
Stroke 41, 618e623. disease. Alzheimer's Dement. 10, 135e142.
Baker, M.L., Hand, P.J., Wong, T.Y., Liew, G., Rochtchina, E., Mitchell, P., Lindley, R.I., Cheung, C.Y., Chen, C., Wong, T.Y., 2015a. Ocular fundus photography as a tool to
Hankey, G.J., Wang, J.J., 2010b. Retinopathy and lobar intracerebral hemorrhage: study stroke and dementia. Semin. Neurol. 35, 481e490.
insights into pathogenesis. Arch. Neurol. 67, 1224e1230. Cheung, C.Y., Ikram, M.K., Klein, R., Wong, T.Y., 2015b. The clinical implications of
Baker, M.L., Wang, J.J., Liew, G., Hand, P.J., De Silva, D.A., Lindley, R.I., Mitchell, P., recent studies on the structure and function of the retinal microvasculature in
Wong, M.C., Rochtchina, E., Wong, T.Y., Wardlaw, J.M., Hankey, G.J., 2010c. Dif- diabetes. Diabetologia 58, 871e885.
ferential associations of cortical and subcortical cerebral atrophy with retinal Cheung, C.Y., Ong, Y.T., Hilal, S., Ikram, M.K., Low, S., Ong, Y.L.,
vascular signs in patients with acute stroke. Stroke 41, 2143e2150. Venketasubramanian, N., Yap, P., Seow, D., Chen, C.L., Wong, T.Y., 2015c. Retinal
Bambo, M.P., Garcia-Martin, E., Pinilla, J., Herrero, R., Satue, M., Otin, S., Fuertes, I., ganglion cell analysis using high-definition optical coherence tomography in
Marques, M.L., Pablo, L.E., 2014. Detection of retinal nerve fiber layer degener- patients with mild cognitive impairment and Alzheimer's disease. J. Alzheimers
ation in patients with Alzheimer's disease using optical coherence tomography: Dis. 45, 45e56.
searching new biomarkers. Acta Ophthalmol. 92, e581e582. Chui, T.Y., Gast, T.J., Burns, S.A., 2013. Imaging of vascular wall fine structure in the
Bayhan, H.A., Aslan Bayhan, S., Celikbilek, A., Tanik, N., Gurdal, C., 2015. Evaluation human retina using adaptive optics scanning laser ophthalmoscopy. Invest
of the chorioretinal thickness changes in Alzheimer's disease using spectral- Ophthalmol. Vis. Sci. 54, 7115e7124.
domain optical coherence tomography. Clin. Exp. Ophthalmol. 43, 145e151. Cooper, L.S., Wong, T.Y., Klein, R., Sharrett, A.R., Bryan, R.N., Hubbard, L.D.,
Benderro, G.F., Lamanna, J.C., 2011. Hypoxia-induced angiogenesis is delayed in Couper, D.J., Heiss, G., Sorlie, P.D., 2006. Retinal microvascular abnormalities and
aging mouse brain. Brain Res. 1389, 50e60. MRI-defined subclinical cerebral infarction: the Atherosclerosis Risk in Com-
Berisha, F., Feke, G.T., Trempe, C.L., McMeel, J.W., Schepens, C.L., 2007a. Retinal munities Study. Stroke 37, 82e86.
abnormalities in early Alzheimer's disease. Invest Ophthalmol. Vis. Sci. 48, Cotter, V.T., 2007. The burden of dementia. Am. J. Manag. Care 13 (Suppl. 8),
2285e2289. S193eS197.
Berisha, F., Feke, G.T., Trempe, C.L., McMeel, J.W., Schepens, C.L., 2007b. Retinal Curtin, R., 2003. Ophthalmic Photography: Retinal Photography, Angiography, and
abnormalities in early Alzheimer's disease. Invest Ophthalmol. Vis.Sci. 48, Electronic Imaging, , second ed.vol. 34. Ophthalmic Surg Lasers Imaging,
2285e2289. pp. 79e80.
Bhaduri, B., Nolan, R.M., Shelton, R.L., Pilutti, L.A., Motl, R.W., Moss, H.E., Pula, J.H., Danesh-Meyer, H.V., Birch, H., Ku, J.Y., Carroll, S., Gamble, G., 2006. Reduction of
Boppart, S.A., 2016. Detection of retinal blood vessel changes in multiple scle- optic nerve fibers in patients with Alzheimer disease identified by laser imag-
rosis with optical coherence tomography. Biomed. Opt. Express 7, 2321e2330. ing. Neurology 67, 1852e1854.
Blanks, J.C., Hinton, D.R., Sadun, A.A., Miller, C.A., 1989. Retinal ganglion cell de Jong, F.J., Schrijvers, E.M., Ikram, M.K., Koudstaal, P.J., de Jong, P.T., Hofman, A.,
degeneration in Alzheimer's disease. Brain Res. 501, 364e372. Vingerling, J.R., Breteler, M.M., 2011. Retinal vascular caliber and risk of de-
Brown, W.R., Thore, C.R., 2011. Review: cerebral microvascular pathology in ageing mentia: the Rotterdam study. Neurology 76, 816e821.
and neurodegeneration. Neuropathol. Appl. Neurobiol. 37, 56e74. de la Torre, J.C., 2004. Is Alzheimer's disease a neurodegenerative or a vascular
Cavallari, M., Stamile, C., Umeton, R., Calimeri, F., Orzi, F., 2015. Novel method for disorder? Data, dogma, and dialectics. Lancet Neurol. 3, 184e190.
automated analysis of retinal images: results in subjects with hypertensive De Silva, T.M., Faraci, F.M., 2016. Microvascular dysfunction and cognitive impair-
retinopathy and CADASIL. Biomed. Res. Int. 2015, 752957. ment. Cell Mol. Neurobiol. 36, 241e258.
Celik, M., Kalenderoglu, A., Sevgi Karadag, A., Bekir Egilmez, O., Han-Almis, B., Dhillon, B., Dhillon, N., 2008. The retina as a window to the brain. Arch. Neurol. 65,
Simsek, A., 2016. Decreases in ganglion cell layer and inner plexiform layer 1547e1548 author reply 1548.
volumes correlate better with disease severity in schizophrenia patients than Diabetic Retinopathy Study. 1981. Report Number 6. Design, methods, and baseline
retinal nerve fiber layer thickness: findings from spectral optic coherence to- results. Report Number 7. A modification of the Airlie House classification of
mography. Eur. Psychiatry 32, 9e15. diabetic retinopathy. Prepared by the Diabetic Retinopathy. Invest Ophthalmol
Cense, B., Gao, W., Brown, J.M., Jones, S.M., Jonnal, R.S., Mujat, M., Park, B.H., de Vis Sci 21, 1e226.
Boer, J.F., Miller, D.T., 2009. Retinal imaging with polarization-sensitive optical Ding, J., Strachan, M.W., Fowkes, F.G., Wong, T.Y., Macgillivray, T.J., Patton, N.,
coherence tomography and adaptive optics. Opt. Express 17, 21634e21651. Gardiner, T.A., Deary, I.J., Price, J.F., 2011. Association of retinal arteriolar dila-
Chalmers, I., Nylenna, M., 2014. A new network to promote evidence-based tation with lower verbal memory: the Edinburgh Type 2 Diabetes Study. Dia-
research. Lancet 384, 1903e1904. betologia 54, 1653e1662.
Chen, C., Homma, A., Mok, V.C., Krishnamoorthy, E., Alladi, S., Meguro, K., Abe, K., Donnan, G.A., Fisher, M., Macleod, M., Davis, S.M., 2008. Stroke. Lancet 371,
Dominguez, J., Marasigan, S., Kandiah, N., Kim, S.Y., Lee, D.Y., De Silva, H.A., 1612e1623.
Yang, Y.H., Pai, M.C., Senanarong, V., Dash, A., 2016 Oct. Alzheimer's disease with Doubal, F.N., de Haan, R., MacGillivray, T.J., Cohn-Hokke, P.E., Dhillon, B.,
cerebrovascular disease: current status in the Asia-Pacific region. J. Intern Med. Dennis, M.S., Wardlaw, J.M., 2010a. Retinal arteriolar geometry is associated
280 (4), 359e374. with cerebral white matter hyperintensities on magnetic resonance imaging.
Cheung, N., Wong, T.Y., 2012. Predicting risk of diabetic retinopathy from retinal Int. J. Stroke 5, 434e439.
vessel analysis: personalized medicine in transition. Arch. Ophthalmol. 130, Doubal, F.N., MacGillivray, T.J., Patton, N., Dhillon, B., Dennis, M.S., Wardlaw, J.M.,
783e784. 2010b. Fractal analysis of retinal vessels suggests that a distinct vasculopathy
Cheung, C.Y., Hsu, W., Lee, M.L., Wang, J.J., Mitchell, P., Lau, Q.P., Hamzah, H., Ho, M., causes lacunar stroke. Neurology 74, 1102e1107.
Wong, T.Y., 2010a. A new method to measure peripheral retinal vascular caliber Dowling, J.E., 2012. The Retina an Approachable Part of the Brain, Revised Edition.

Please cite this article in press as: Cheung, C.Y.-l., et al., Imaging retina to study dementia and stroke, Progress in Retinal and Eye Research (2016),
http://dx.doi.org/10.1016/j.preteyeres.2017.01.001
16 C.Y.-l. Cheung et al. / Progress in Retinal and Eye Research xxx (2016) 1e19

Harvard University Press. integrity in glaucoma: an adaptive-optics scanning laser ophthalmoscopy study.
Drexler, W., Fujimoto, J.G., 2008. State-of-the-art retinal optical coherence tomog- Am. J. Ophthalmol. 171, 53e66.
raphy. Prog. Retin Eye Res. 27, 45e88. Hassan, A., Gormley, K., O'Sullivan, M., Knight, J., Sham, P., Vallance, P., Bamford, J.,
Dubra, A., Sulai, Y., 2011. Reflective afocal broadband adaptive optics scanning Markus, H., 2004. Endothelial nitric oxide gene haplotypes and risk of cerebral
ophthalmoscope. Biomed. Opt. Express 2, 1757e1768. small-vessel disease. Stroke 35, 654e659.
Einarsdottir, A.B., Hardarson, S.H., Kristjansdottir, J.V., Bragason, D.T., Snaedal, J., Heringa, S.M., Bouvy, W.H., van den Berg, E., Moll, A.C., Kappelle, L.J., Biessels, G.J.,
Stefansson, E., 2015. Retinal oximetry imaging in Alzheimer's disease. 2013. Associations between retinal microvascular changes and dementia,
J. Alzheimers Dis. 49, 79e83. cognitive functioning, and brain imaging abnormalities: a systematic review.
Faust, O., Acharya, U.R., Ng, E.Y., Ng, K.H., Suri, J.S., 2012. Algorithms for the auto- J. Cereb. Blood Flow. Metab. 33, 983e995.
mated detection of diabetic retinopathy using digital fundus images: a review. Hilal, S., Ong, Y.T., Cheung, C.Y., Tan, C.S., Venketasubramanian, N., Niessen, W.J.,
J. Med. Syst. 36, 145e157. Vrooman, H., Anuar, A.R., Chew, M., Chen, C., Wong, T.Y., Ikram, M.K., 2014.
Feigin, V.L., Roth, G.A., Naghavi, M., Parmar, P., Krishnamurthi, R., Chugh, S., Microvascular network alterations in retina of subjects with cerebral small
Mensah, G.A., Norrving, B., Shiue, I., Ng, M., Estep, K., Cercy, K., Murray, C.J., vessel disease. Neurosci. Lett. 577, 95e100.
Forouzanfar, M.H., Global Burden of Diseases, I., Risk Factors, S., Stroke Experts Hinton, D.R., Sadun, A.A., Blanks, J.C., Miller, C.A., 1986. Optic-nerve degeneration in
Writing, G., 2016. Global burden of stroke and risk factors in 188 countries, Alzheimer's disease. N. Engl. J. Med. 315, 485e487.
during 1990-2013: a systematic analysis for the Global Burden of Disease Study Ho, W.L., Leung, Y., Tsang, A.W., So, K.F., Chiu, K., Chang, R.C., 2012. Review: tau-
2013. Lancet Neurol. 15, 913e924. opathy in the retina and optic nerve: does it shadow pathological changes in
Ferrara, D., Waheed, N.K., Duker, J.S., 2016 May. Investigating the choriocapillaris the brain? Mol. Vis. 18, 2700e2710.
and choroidal vasculature with new optical coherence tomography technolo- Huang, D., Swanson, E.A., Lin, C.P., Schuman, J.S., Stinson, W.G., Chang, W., Hee, M.R.,
gies. Prog. Retin Eye Res. 52, 130e155. Flotte, T., Gregory, K., Puliafito, C.A., et al., 1991. Optical coherence tomography.
Ferri, C.P., Prince, M., Brayne, C., Brodaty, H., Fratiglioni, L., Ganguli, M., Hall, K., Science 254, 1178e1181.
Hasegawa, K., Hendrie, H., Huang, Y., Jorm, A., Mathers, C., Menezes, P.R., Hubbard, L.D., Brothers, R.J., King, W.N., Clegg, L.X., Klein, R., Cooper, L.S.,
Rimmer, E., Scazufca, M., 2005. Global prevalence of dementia: a Delphi Sharrett, A.R., Davis, M.D., Cai, J., 1999. Methods for evaluation of retinal
consensus study. Lancet 366, 2112e2117. microvascular abnormalities associated with hypertension/sclerosis in the
Fox, P.J., Kohatsu, N., Max, W., Arnsberger, P., 2001. Estimating the costs of caring for Atherosclerosis Risk in Communities Study. Ophthalmology 106, 2269e2280.
people with Alzheimer disease in California: 2000-2040. J. Public Health Policy Hughes, A.D., Martinez-Perez, E., Jabbar, A.S., Hassan, A., Witt, N.W., Mistry, P.D.,
22, 88e97. Chapman, N., Stanton, A.V., Beevers, G., Pedrinelli, R., Parker, K.H., Thom, S.A.,
Frost, S., Kanagasingam, Y., Sohrabi, H., Vignarajan, J., Bourgeat, P., Salvado, O., 2006. Quantification of topological changes in retinal vascular architecture in
Villemagne, V., Rowe, C.C., Macaulay, S.L., Szoeke, C., Ellis, K.A., Ames, D., essential and malignant hypertension. J. Hypertens. 24, 889e894.
Masters, C.L., Rainey-Smith, S., Martins, R.N., Group, A.R., 2013. Retinal vascular Hughes, A.D., Stanton, A.V., Jabbar, A.S., Chapman, N., Martinez-Perez, M.E., McG
biomarkers for early detection and monitoring of Alzheimer's disease. Transl. Thom, S.A., 2008. Effect of antihypertensive treatment on retinal microvascular
Psychiatry 3, e233. changes in hypertension. J. Hypertens. 26, 1703e1707.
Gao, L., Liu, Y., Li, X., Bai, Q., Liu, P., 2015. Abnormal retinal nerve fiber layer thickness Hughes, A.D., Wong, T.Y., Witt, N., Evans, R., Thom, S.A., Klein, B.E., Chaturvedi, N.,
and macula lutea in patients with mild cognitive impairment and Alzheimer's Klein, R., 2009. Determinants of retinal microvascular architecture in normal
disease. Arch. Gerontol. Geriatr. 60, 162e167. subjects. Microcirculation 16, 159e166.
Garcia-Martin, E., Satue, M., Fuertes, I., Otin, S., Alarcia, R., Herrero, R., Bambo, M.P., Hughes, A.D., Falaschetti, E., Witt, N., Wijetunge, S., Thom, S.A., Tillin, T.,
Fernandez, J., Pablo, L.E., 2012. Ability and reproducibility of Fourier-domain Aldington, S.J., Chaturvedi, N., 2016. Association of retinopathy and retinal
optical coherence tomography to detect retinal nerve fiber layer atrophy in microvascular abnormalities with stroke and cerebrovascular disease. Stroke 47,
Parkinson's disease. Ophthalmology 119, 2161e2167. 2862e2864.
Garcia-Martin, E.S., Rojas, B., Ramirez, A.I., de Hoz, R., Salazar, J.J., Yubero, R., Gil, P., Hwang, J.Y., Wachsmann-Hogiu, S., Ramanujan, V.K., Nowatzyk, A.G., Koronyo, Y.,
Trivino, A., Ramirez, J.M., 2014. Macular thickness as a potential biomarker of Medina-Kauwe, L.K., Gross, Z., Gray, H.B., Farkas, D.L., 2011. Multimodal wide-
mild Alzheimer's disease. Ophthalmology 121 (1149e1151), e1143. field two-photon excitation imaging: characterization of the technique for
Gatto, N.M., Varma, R., Torres, M., Wong, T.Y., Johnson, P.L., Segal-Gidan, F., in vivo applications. Biomed. Opt. Express 2, 356e364.
Mack, W.J., 2012. Retinal microvascular abnormalities and cognitive function in Iadecola, C., 2004. Neurovascular regulation in the normal brain and in Alzheimer's
Latino adults in Los Angeles. Ophthalmic Epidemiol. 19, 127e136. disease. Nat. Rev. Neurosci. 5, 347e360.
Gharbiya, M., Trebbastoni, A., Parisi, F., Manganiello, S., Cruciani, F., D'Antonio, F., De Ikram, M.K., de Jong, F.J., Bos, M.J., Vingerling, J.R., Hofman, A., Koudstaal, P.J., de
Vico, U., Imbriano, L., Campanelli, A., De Lena, C., 2014. Choroidal thinning as a Jong, P.T., Breteler, M.M., 2006a. Retinal vessel diameters and risk of stroke: the
new finding in Alzheimer's disease: evidence from enhanced depth imaging Rotterdam study. Neurology 66, 1339e1343.
spectral domain optical coherence tomography. J. Alzheimers Dis. 40, 907e917. Ikram, M.K., de Jong, F.J., Van Dijk, E.J., Prins, N.D., Hofman, A., Breteler, M.M., de
Ghezzi, A., Martinelli, V., Torri, V., Zaffaroni, M., Rodegher, M., Comi, G., Zibetti, A., Jong, P.T., 2006b. Retinal vessel diameters and cerebral small vessel disease: the
Canal, N., 1999. Long-term follow-up of isolated optic neuritis: the risk of Rotterdam Scan Study. Brain 129, 182e188.
developing multiple sclerosis, its outcome, and the prognostic role of para- Ikram, M.K., Cheung, C.Y., Wong, T.Y., Chen, C.P., 2012. Retinal pathology as
clinical tests. J. Neurol. 246, 770e775. biomarker for cognitive impairment and Alzheimer's disease. J. Neurol. Neu-
Girouard, H., Iadecola, C., 2006. Neurovascular coupling in the normal brain and in rosurg. Psychiatry 83, 917e922.
hypertension, stroke, and Alzheimer disease. J. Appl. Physiol. 100 (1985), Iseri, P.K., Altinas, O., Tokay, T., Yuksel, N., 2006. Relationship between cognitive
328e335. impairment and retinal morphological and visual functional abnormalities in
Gorelick, P.B., Scuteri, A., Black, S.E., Decarli, C., Greenberg, S.M., Iadecola, C., Alzheimer disease. J. Neuroophthalmol. 26, 18e24.
Launer, L.J., Laurent, S., Lopez, O.L., Nyenhuis, D., Petersen, R.C., Schneider, J.A., Jentsch, S., Schweitzer, D., Schmidtke, K.U., Peters, S., Dawczynski, J., Bar, K.J.,
Tzourio, C., Arnett, D.K., Bennett, D.A., Chui, H.C., Higashida, R.T., Lindquist, R., Hammer, M., 2015. Retinal fluorescence lifetime imaging ophthalmoscopy
Nilsson, P.M., Roman, G.C., Sellke, F.W., Seshadri, S., 2011. Vascular contributions measures depend on the severity of Alzheimer's disease. Acta Ophthalmol. 93,
to cognitive impairment and dementia: a statement for healthcare pro- e241e247.
fessionals from the American heart association/American stroke association. Jia, Y., Bailey, S.T., Hwang, T.S., McClintic, S.M., Gao, S.S., Pennesi, M.E., Flaxel, C.J.,
Stroke 42, 2672e2713. Lauer, A.K., Wilson, D.J., Hornegger, J., Fujimoto, J.G., Huang, D., 2015. Quanti-
Goto, I., Katsuki, S., Ikui, H., Kimoto, K., Mimatsu, T., 1975. Pathological studies on tative optical coherence tomography angiography of vascular abnormalities in
the intracerebral and retinal arteries in cerebrovascular and noncerebrovascular the living human eye. Proc. Natl. Acad. Sci. U. S. A. 112, E2395eE2402.
diseases. Stroke 6, 263e269. Joshi, V., Agurto, C., VanNess, R., Nemeth, S., Soliz, P., Barriga, S., 2014. Compre-
Grinton, M.E., Laude, A., MacGillivray, T., Henderson, R., Starr, J.M., Deary, I.J., hensive automatic assessment of retinal vascular abnormalities for computer-
Aspinall, P., Dhillon, B., 2012. The association between retinal vessel assisted retinopathy grading. Conf. Proc. IEEE Eng. Med. Biol. Soc. 2014,
morphology and retinal nerve fiber layer thickness in an elderly population. 6320e6323.
Ophthalmic Surg. Lasers Imaging 43, S61eS66. Kalesnykas, G., Tuulos, T., Uusitalo, H., Jolkkonen, J., 2008. Neurodegeneration and
Gugleta, K., Kochkorov, A., Kavroulaki, D., Katamay, R., Weier, K., Mehling, M., cellular stress in the retina and optic nerve in rat cerebral ischemia and
Kappos, L., Flammer, J., Orgul, S., 2009. Retinal vessels in patients with multiple hypoperfusion models. Neuroscience 155, 937e947.
sclerosis: baseline diameter and response to flicker light stimulation. Klin. Kantarci, K., 2005. Magnetic resonance markers for early diagnosis and progression
Monbl Augenheilkd. 226, 272e275. of Alzheimer's disease. Expert Rev. Neurother. 5, 663e670.
Haan, M., Espeland, M.A., Klein, B.E., Casanova, R., Gaussoin, S.A., Jackson, R.D., Kawasaki, R., Che Azemin, M.Z., Kumar, D.K., Tan, A.G., Liew, G., Wong, T.Y.,
Millen, A.E., Resnick, S.M., Rossouw, J.E., Shumaker, S.A., Wallace, R., Yaffe, K., Mitchell, P., Wang, J.J., 2011. Fractal dimension of the retinal vasculature and risk
2012. Cognitive function and retinal and ischemic brain changes: the Women's of stroke: a nested case-control study. Neurology 76, 1766e1767.
Health Initiative. Neurology 78, 942e949. Kawasaki, R., Xie, J., Cheung, N., Lamoureux, E., Klein, R., Klein, B.E., Cotch, M.F.,
Hammes, H.P., Feng, Y., Pfister, F., Brownlee, M., 2011. Diabetic retinopathy: targeting Sharrett, A.R., Shea, S., Wong, T.Y., Mesa, 2012. Retinal microvascular signs and
vasoregression. Diabetes 60, 9e16. risk of stroke: the Multi-Ethnic Study of Atherosclerosis (MESA). Stroke 43,
Hardin, J.S., Taibbi, G., Nelson, S.C., Chao, D., Vizzeri, G., 2015. Factors affecting 3245e3251.
cirrus-HD OCT optic disc scan quality: a review with case examples. Kergoat, H., Kergoat, M.J., Justino, L., Chertkow, H., Robillard, A., Bergman, H., 2001.
J. Ophthalmol. 2015, 746150. An evaluation of the retinal nerve fiber layer thickness by scanning laser
Hasegawa, T., Ooto, S., Takayama, K., Makiyama, Y., Akagi, T., Ikeda, H.O., polarimetry in individuals with dementia of the Alzheimer type. Acta Oph-
Nakanishi, H., Suda, K., Yamada, H., Uji, A., Yoshimura, N., 2016 Nov. Cone thalmol. Scand. 79, 187e191.

Please cite this article in press as: Cheung, C.Y.-l., et al., Imaging retina to study dementia and stroke, Progress in Retinal and Eye Research (2016),
http://dx.doi.org/10.1016/j.preteyeres.2017.01.001
 C.Y.-l. Cheung et al. / Progress in Retinal and Eye Research xxx (2016) 1e19 17

Kernt, M., Hadi, I., Pinter, F., Seidensticker, F., Hirneiss, C., Haritoglou, C., Kampik, A., 921e925.
Ulbig, M.W., Neubauer, A.S., 2012. Assessment of diabetic retinopathy using Lim, M., Sasongko, M.B., Ikram, M.K., Lamoureux, E., Wang, J.J., Wong, T.Y.,
nonmydriatic ultra-widefield scanning laser ophthalmoscopy (Optomap) Cheung, C.Y., 2013. Systemic associations of dynamic retinal vessel analysis: a
compared with ETDRS 7-field stereo photography. Diabetes Care 35, review of current literature. Microcirculation 20, 257e268.
2459e2463. Lindley, R.I., Wang, J.J., Wong, M.C., Mitchell, P., Liew, G., Hand, P., Wardlaw, J., De
Kesler, A., Vakhapova, V., Korczyn, A.D., Naftaliev, E., Neudorfer, M., 2011. Retinal Silva, D.A., Baker, M., Rochtchina, E., Chen, C., Hankey, G.J., Chang, H.M.,
thickness in patients with mild cognitive impairment and Alzheimer's disease. Fung, V.S., Gomes, L., Wong, T.Y., 2009. Retinal microvasculature in acute
Clin. Neurol. Neurosurg. 113, 523e526. lacunar stroke: a cross-sectional study. Lancet Neurol. 8, 628e634.
Kim, D.H., Newman, A.B., Hajjar, I., Strotmeyer, E.S., Klein, R., Newton, E., Lombardo, M., Serrao, S., Devaney, N., Parravano, M., Lombardo, G., 2013. Adaptive
Sarnak, M.J., Burke, G.L., Lipsitz, L.A., 2011a. Retinal microvascular signs and optics technology for high-resolution retinal imaging. Sensors (Basel) 13,
functional loss in older persons: the cardiovascular health study. Stroke 42, 334e366.
1589e1595. London, A., Benhar, I., Schwartz, M., 2013. The retina as a window to the brain-from
Kim, M., Park, K.H., Kwon, J.W., Jeoung, J.W., Kim, T.W., Kim, D.M., 2011b. Retinal eye research to CNS disorders. Nat. Rev. Neurol. 9, 44e53.
nerve fiber layer defect and cerebral small vessel disease. Invest Ophthalmol. Longstreth Jr., W., Larsen, E.K., Klein, R., Wong, T.Y., Sharrett, A.R., Lefkowitz, D.,
Vis. Sci. 52, 6882e6886. Manolio, T.A., 2007. Associations between findings on cranial magnetic reso-
Kim, N.R., Lee, H., Lee, E.S., Kim, J.H., Hong, S., Je Seong, G., Kim, C.Y., 2012. Influence nance imaging and retinal photography in the elderly: the Cardiovascular
of cataract on time domain and spectral domain optical coherence tomography Health Study. Am. J. Epidemiol. 165, 78e84.
retinal nerve fiber layer measurements. J. Glaucoma 21, 116e122. Lu, Y., Li, Z., Zhang, X., Ming, B., Jia, J., Wang, R., Ma, D., 2010. Retinal nerve fiber layer
Kirbas, S., Turkyilmaz, K., Anlar, O., Tufekci, A., Durmus, M., 2013. Retinal nerve fiber structure abnormalities in early Alzheimer's disease: evidence in optical
layer thickness in patients with Alzheimer disease. J. Neuroophthalmol. 33, coherence tomography. Neurosci. Lett. 480, 69e72.
58e61. Maberley, D., Morris, A., Hay, D., Chang, A., Hall, L., Mandava, N., 2004. A comparison
Kling, M.A., Trojanowski, J.Q., Wolk, D.A., Lee, V.M., Arnold, S.E., 2013. Vascular of digital retinal image quality among photographers with different levels of
disease and dementias: paradigm shifts to drive research in new directions. training using a non-mydriatic fundus camera. Ophthalmic Epidemiol. 11,
Alzheimers Dement. 9, 76e92. 191e197.
Kocaoglu, O.P., Lee, S., Jonnal, R.S., Wang, Q., Herde, A.E., Derby, J.C., Gao, W., MacGillivray, T.J., Trucco, E., Cameron, J.R., Dhillon, B., Houston, J.G., van Beek, E.J.,
Miller, D.T., 2011. Imaging cone photoreceptors in three dimensions and in time 2014. Retinal imaging as a source of biomarkers for diagnosis, characterization
using ultrahigh resolution optical coherence tomography with adaptive optics. and prognosis of chronic illness or long-term conditions. Br. J. Radiol. 87,
Biomed. Opt. Express 2, 748e763. 20130832.
Koch, E., Rosenbaum, D., Brolly, A., Sahel, J.A., Chaumet-Riffaud, P., Girerd, X., Marziani, E., Pomati, S., Ramolfo, P., Cigada, M., Giani, A., Mariani, C., Staurenghi, G.,
Rossant, F., Paques, M., 2014. Morphometric analysis of small arteries in the 2013. Evaluation of retinal nerve fiber layer and ganglion cell layer thickness in
human retina using adaptive optics imaging: relationship with blood pressure Alzheimer's disease using spectral-domain optical coherence tomography.
and focal vascular changes. J. Hypertens. 32, 890e898. Invest Ophthalmol. Vis. Sci. 54, 5953e5958.
Koronyo-Hamaoui, M., Koronyo, Y., Ljubimov, A.V., Miller, C.A., Ko, M.K., Black, K.L., McGeechan, K., Macaskill, P., Irwig, L., Liew, G., Wong, T.Y., 2008. Assessing new
Schwartz, M., Farkas, D.L., 2011. Identification of amyloid plaques in retinas from biomarkers and predictive models for use in clinical practice: a clinician's guide.
Alzheimer's patients and noninvasive in vivo optical imaging of retinal plaques Arch. Intern. Med. 168, 2304e2310.
in a mouse model. Neuroimage 54 (Suppl. 1), S204eS217. McGeechan, K., Liew, G., Macaskill, P., Irwig, L., Klein, R., Klein, B.E., Wang, J.J.,
Kromer, R., Serbecic, N., Hausner, L., Froelich, L., Aboul-Enein, F., Beutelspacher, S.C., Mitchell, P., Vingerling, J.R., de Jong, P.T., Witteman, J.C., Breteler, M.M., Shaw, J.,
2014. Detection of retinal nerve fiber layer defects in Alzheimer's disease using Zimmet, P., Wong, T.Y., 2009. Prediction of incident stroke events based on
SD-OCT. Front. Psychiatry 5, 22. retinal vessel caliber: a systematic review and individual-participant meta-
Kwa, V.I., van der Sande, J.J., Stam, J., Tijmes, N., Vrooland, J.L., 2002. Retinal arterial analysis. Am.J Epidemiol. 170, 1323e1332.
changes correlate with cerebral small-vessel disease. Neurology 59, 1536e1540. Meier, M.H., Shalev, I., Moffitt, T.E., Kapur, S., Keefe, R.S., Wong, T.Y., Belsky, D.W.,
La Morgia, C., Ross-Cisneros, F.N., Koronyo, Y., Hannibal, J., Gallassi, R., Cantalupo, G., Harrington, H., Hogan, S., Houts, R., Caspi, A., Poulton, R., 2013. Microvascular
Sambati, L., Pan, B.X., Tozer, K.R., Barboni, P., Provini, F., Avanzini, P., abnormality in schizophrenia as shown by retinal imaging. Am. J. Psychiatry
Carbonelli, M., Pelosi, A., Chui, H., Liguori, R., Baruzzi, A., Koronyo-Hamaoui, M., 170, 1451e1459.
Sadun, A.A., Carelli, V., 2016. Melanopsin retinal ganglion cell loss in Alzheimer Meier, M.H., Gillespie, N.A., Hansell, N.K., Hewitt, A.W., Hickie, I.B., Lu, Y.,
disease. Ann. Neurol. 79, 90e109. MacGregor, S., Medland, S.E., Sun, C., Wong, T.Y., Wright, M.J., Zhu, G.,
Lammie, G.A., 2002. Hypertensive cerebral small vessel disease and stroke. Brain Martin, N.G., Mackey, D.A., 2014. Associations between depression and anxiety
Pathol. 12, 358e370. symptoms and retinal vessel caliber in adolescents and young adults. Psycho-
Larrosa, J.M., Garcia-Martin, E., Bambo, M.P., Pinilla, J., Polo, V., Otin, S., Satue, M., som. Med. 76, 732e738.
Herrero, R., Pablo, L.E., 2014. Potential new diagnostic tool for Alzheimer's Mitchell, P., Wang, J.J., Wong, T.Y., Smith, W., Klein, R., Leeder, S.R., 2005. Retinal
disease using a linear discriminant function for Fourier domain optical coher- microvascular signs and risk of stroke and stroke mortality. Neurology 65,
ence tomography. Invest Ophthalmol. Vis. Sci. 55, 3043e3051. 1005e1009.
Larson, E.B., Langa, K.M., 2008. The rising tide of dementia worldwide. Lancet 372, Moreno-Ramos, T., Benito-Leon, J., Villarejo, A., Bermejo-Pareja, F., 2013. Retinal
430e432. nerve fiber layer thinning in dementia associated with Parkinson's disease,
Lau, P., Lee, M.L., Hsu, W., Wong, T.Y., 2014. The Singapore eye vessel assessment dementia with Lewy bodies, and Alzheimer's disease. J. Alzheimers Dis. 34,
system. In: Ng, E.Y.K., Acharya, U.R., Campilho, A., Suri, J.S. (Eds.), Image Analysis 659e664.
and Modeling in Opthalmology. CRC Press, pp. 143e160. Moschos, M.M., Tagaris, G., Markopoulos, I., Margetis, I., Tsapakis, S., Kanakis, M.,
Leitgeb, R.A., Werkmeister, R.M., Blatter, C., Schmetterer, L., 2014. Doppler optical Koutsandrea, C., 2011. Morphologic changes and functional retinal impairment
coherence tomography. Prog. Retin Eye Res. 41, 26e43. in patients with Parkinson disease without visual loss. Eur. J. Ophthalmol. 21,
Leung, C.K., Cheung, C.Y., Weinreb, R.N., Qiu, Q., Liu, S., Li, H., Xu, G., Fan, N., 24e29.
Huang, L., Pang, C.P., Lam, D.S., 2009. Retinal nerve fiber layer imaging with Moschos, M.M., Markopoulos, I., Chatziralli, I., Rouvas, A., Papageorgiou, S.G.,
spectral-domain optical coherence tomography: a variability and diagnostic Ladas, I., Vassilopoulos, D., 2012. Structural and functional impairment of the
performance study. Ophthalmology 116 (1257e1263), 1263 e1251e1252. retina and optic nerve in Alzheimer's disease. Curr. Alzheimer Res. 9, 782e788.
Leung, C.K., Yu, M., Weinreb, R.N., Ye, C., Liu, S., Lai, G., Lam, D.S., 2012. Retinal nerve Mrejen, S., Spaide, R.F., 2013. Optical coherence tomography: imaging of the choroid
fiber layer imaging with spectral-domain optical coherence tomography: a and beyond. Surv. Ophthalmol. 58, 387e429.
prospective analysis of age-related loss. Ophthalmology 119, 731e737. Murray, C.D., 1926. The physiological principle of minimum work: I. The vascular
Leung, C.K., Ye, C., Weinreb, R.N., Yu, M., Lai, G., Lam, D.S., 2013. Impact of age- system and the cost of blood volume. Proc. Natl. Acad. Sci. U.S.A 12, 207e214.
related change of retinal nerve fiber layer and macular thicknesses on evalua- Mwanza, J.C., Durbin, M.K., Budenz, D.L., Sayyad, F.E., Chang, R.T., Neelakantan, A.,
tion of glaucoma progression. Ophthalmology 120, 2485e2492. Godfrey, D.G., Carter, R., Crandall, A.S., 2012. Glaucoma diagnostic accuracy of
Li, H., Mitchell, P., Liew, G., Rochtchina, E., Kifley, A., Wong, T.Y., Hsu, W., Lee, M.L., ganglion cell-inner plexiform layer thickness: comparison with nerve fiber
Zhang, Y.P., Wang, J.J., 2010. Lens opacity and refractive influences on the layer and optic nerve head. Ophthalmology 119, 1151e1158.
measurement of retinal vascular fractal dimension. Acta Ophthalmol. 88, Nguyen, T.T., Wong, T.Y., 2006. Retinal vascular manifestations of metabolic disor-
e234ee240. ders. Trends Endocrinol. Metab. 17, 262e268.
Liang, J., Williams, D.R., Miller, D.T., 1997. Supernormal vision and high-resolution Nguyen, U.T., Bhuiyan, A., Park, L.A., Kawasaki, R., Wong, T.Y., Wang, J.J., Mitchell, P.,
retinal imaging through adaptive optics. J. Opt. Soc. Am. A Opt. Image Sci. Vis. Ramamohanarao, K., 2013. Automated quantification of retinal arteriovenous
14, 2884e2892. nicking from colour fundus images. Conf. Proc. IEEE Eng. Med. Biol. Soc. 2013,
Liew, G., Wang, J.J., Mitchell, P., Wong, T.Y., 2008. Retinal vascular imaging: a new 5865e5868.
tool in microvascular disease research. Circ. Cardiovasc. Imaging 1, 156e161. O'Brien, J.T., Thomas, A., 2015. Vascular dementia. Lancet 386, 1698e1706.
Liew, G., Mitchell, P., Wong, T.Y., Lindley, R.I., Cheung, N., Kaushik, S., Wang, J.J., Ong, Y.T., De Silva, D.A., Cheung, C.Y., Chang, H.M., Chen, C.P., Wong, M.C., Wong, T.Y.,
2009. Retinal microvascular signs and cognitive impairment. J. Am. Geriatr. Soc Ikram, M.K., 2013. Microvascular structure and network in the retina of patients
57, 1892e1896. with ischemic stroke. Stroke 44, 2121e2127.
Liew, G., Baker, M.L., Wong, T.Y., Hand, P.J., Wang, J.J., Mitchell, P., De Silva, D.A., Ong, Y.T., Hilal, S., Cheung, C.Y., Xu, X., Chen, C., Venketasubramanian, N., Wong, T.Y.,
Wong, M.C., Rochtchina, E., Lindley, R.I., Wardlaw, J.M., Hankey, G.J., Multi- Ikram, M.K., 2014. Retinal vascular fractals and cognitive impairment. Dement.
Centre Retinal Stroke Study, G., 2014. Differing associations of white matter Geriatr. Cogn. Dis. Extra 4, 305e313.
lesions and lacunar infarction with retinal microvascular signs. Int. J. Stroke 9, Ong, Y.T., Hilal, S., Cheung, C.Y., Venketasubramanian, N., Niessen, W.J., Vrooman, H.,

Please cite this article in press as: Cheung, C.Y.-l., et al., Imaging retina to study dementia and stroke, Progress in Retinal and Eye Research (2016),
http://dx.doi.org/10.1016/j.preteyeres.2017.01.001
18 C.Y.-l. Cheung et al. / Progress in Retinal and Eye Research xxx (2016) 1e19

Anuar, A.R., Chew, M., Chen, C., Wong, T.Y., Ikram, M.K., 2015. Retinal neuro- Green, A.J., Nguyen, Q.D., Calabresi, P.A., 2012. Microcystic macular oedema,
degeneration on optical coherence tomography and cerebral atrophy. Neurosci. thickness of the inner nuclear layer of the retina, and disease characteristics in
Lett. 584, 12e16. multiple sclerosis: a retrospective study. Lancet Neurol. 11, 963e972.
Pantoni, L., 2010. Cerebral small vessel disease: from pathogenesis and clinical Schrijvers, E.M., Buitendijk, G.H., Ikram, M.K., Koudstaal, P.J., Hofman, A.,
characteristics to therapeutic challenges. Lancet Neurol. 9, 689e701. Vingerling, J.R., Breteler, M.M., 2012. Retinopathy and risk of dementia: the
Pantoni, L., Garcia, J.H., 1997. Pathogenesis of leukoaraiosis: a review. Stroke 28, Rotterdam Study. Neurology 79, 365e370.
652e659. Schuman, S.G., Koreishi, A.F., Farsiu, S., Jung, S.H., Izatt, J.A., Toth, C.A., 2009.
Paquet, C., Boissonnot, M., Roger, F., Dighiero, P., Gil, R., Hugon, J., 2007a. Abnormal Photoreceptor layer thinning over drusen in eyes with age-related macular
retinal thickness in patients with mild cognitive impairment and Alzheimer's degeneration imaged in vivo with spectral-domain optical coherence tomog-
disease. Neurosci. Lett. 420, 97e99. raphy. Ophthalmology 116 (488e496), e482.
Paquet, C., Boissonnot, M., Roger, F., Dighiero, P., Gil, R., Hugon, J., 2007b. Abnormal Sherry, L.M., Wang, J.J., Rochtchina, E., Wong, T., Klein, R., Hubbard, L., Mitchell, P.,
retinal thickness in patients with mild cognitive impairment and Alzheimer's 2002. Reliability of computer-assisted retinal vessel measurementin a popula-
disease. Neurosci. Lett. 420, 97e99. tion. Clin.Experiment. Ophthalmol. 30, 179e182.
Parisi, V., Restuccia, R., Fattapposta, F., Mina, C., Bucci, M.G., Pierelli, F., 2001. Silva, P.S., Cavallerano, J.D., Tolls, D., Omar, A., Thakore, K., Patel, B., Sehizadeh, M.,
Morphological and functional retinal impairment in Alzheimer's disease pa- Tolson, A.M., Sun, J.K., Aiello, L.M., Aiello, L.P., 2014. Potential efficiency benefits
tients. Clin. Neurophysiol. 112, 1860e1867. of nonmydriatic ultrawide field retinal imaging in an ocular telehealth diabetic
Patton, N., Aslam, T., Macgillivray, T., Pattie, A., Deary, I.J., Dhillon, B., 2005. Retinal retinopathy program. Diabetes Care 37, 50e55.
vascular image analysis as a potential screening tool for cerebrovascular dis- Silva, P.S., Horton, M.B., Clary, D., Lewis, D.G., Sun, J.K., Cavallerano, J.D., Aiello, L.P.,
ease: a rationale based on homology between cerebral and retinal microvas- 2016. Identification of diabetic retinopathy and ungradable image rate with
culatures. J. Anat. 206, 319e348. ultrawide field imaging in a National Teleophthalmology Program. Ophthal-
Patton, N., Aslam, T.M., Macgillivray, T., Deary, I.J., Dhillon, B., Eikelboom, R.H., mology 123, 1360e1367.
Yogesan, K., Constable, I.J., 2006. Retinal image analysis: concepts, applications Smith, J.J., Sorensen, A.G., Thrall, J.H., 2003. Biomarkers in imaging: realizing radi-
and potential. Prog.Retin.Eye Res. 25, 99e127. ology's future. Radiology 227, 633e638.
Patton, N., Pattie, A., Macgillivray, T., Aslam, T., Dhillon, B., Gow, A., Starr, J.M., Soderstrom, M., 2001. Optic neuritis and multiple sclerosis. Acta Ophthalmol. Scand.
Whalley, L.J., Deary, I.J., 2007. The association between retinal vascular network 79, 223e227.
geometry and cognitive ability in an elderly population. Invest Ophthalmol. Vis. Soliman, M.K., Sadiq, M.A., Agarwal, A., Sarwar, S., Hassan, M., Hanout, M., Graf, F.,
Sci. 48, 1995e2000. High, R., Do, D.V., Nguyen, Q.D., Sepah, Y.J., 2016. High-resolution imaging of
Petzold, A., de Boer, J.F., Schippling, S., Vermersch, P., Kardon, R., Green, A., parafoveal cones in different stages of diabetic retinopathy using adaptive op-
Calabresi, P.A., Polman, C., 2010. Optical coherence tomography in multiple tics fundus camera. PLoS One 11, e0152788.
sclerosis: a systematic review and meta-analysis. Lancet Neurol. 9, 921e932. Sperling, R.A., Rentz, D.M., Johnson, K.A., Karlawish, J., Donohue, M., Salmon, D.P.,
Polo, V., Garcia-Martin, E., Bambo, M.P., Pinilla, J., Larrosa, J.M., Satue, M., Otin, S., Aisen, P., 2014. The A4 study: stopping AD before symptoms begin? Sci. Transl.
Pablo, L.E., 2014. Reliability and validity of Cirrus and Spectralis optical coher- Med. 6, 228fs213.
ence tomography for detecting retinal atrophy in Alzheimer's disease. Eye Sun, C., Wang, J.J., Mackey, D.A., Wong, T.Y., 2009. Retinal vascular caliber: systemic,
(Lond) 28, 680e690. environmental, and genetic associations. Surv. Ophthalmol. 54, 74e95.
Qiu, C., Cotch, M.F., Sigurdsson, S., Klein, R., Jonasson, F., Klein, B.E., Garcia, M., Tan, O., Chopra, V., Lu, A.T., Schuman, J.S., Ishikawa, H., Wollstein, G., Varma, R.,
Jonsson, P.V., Harris, T.B., Eiriksdottir, G., Kjartansson, O., van Buchem, M.A., Huang, D., 2009. Detection of macular ganglion cell loss in glaucoma by Fourier-
Gudnason, V., Launer, L.J., 2009. Microvascular lesions in the brain and retina: domain optical coherence tomography. Ophthalmology 116 (2305e2314),
the age, gene/environment susceptibility-Reykjavik study. Ann. Neurol. 65, e2301e2302.
569e576. Taylor, A.M., MacGillivray, T.J., Henderson, R.D., Ilzina, L., Dhillon, B., Starr, J.M.,
Qiu, C., Cotch, M.F., Sigurdsson, S., Jonsson, P.V., Jonsdottir, M.K., Sveinbjrnsdottir, S., Deary, I.J., 2015. Retinal vascular fractal dimension, childhood IQ, and cognitive
Eiriksdottir, G., Klein, R., Harris, T.B., van Buchem, M.A., Gudnason, V., ability in old age: the Lothian Birth Cohort Study 1936. PLoS One 10, e0121119.
Launer, L.J., 2010. Cerebral microbleeds, retinopathy, and dementia: the AGES- Thom, S., Stettler, C., Stanton, A., Witt, N., Tapp, R., Chaturvedi, N., Allemann, S.,
Reykjavik Study. Neurology 75, 2221e2228. Mayet, J., Sever, P., Poulter, N., O'Brien, E., Hughes, A., 2009. Differential effects of
Ratchford, J.N., Saidha, S., Sotirchos, E.S., Oh, J.A., Seigo, M.A., Eckstein, C., antihypertensive treatment on the retinal microcirculation: an anglo-
Durbin, M.K., Oakley, J.D., Meyer, S.A., Conger, A., Frohman, T.C., Newsome, S.D., scandinavian cardiac outcomes trial substudy. Hypertension 54, 405e408.
Balcer, L.J., Frohman, E.M., Calabresi, P.A., 2013. Active MS is associated with Thomas, G.N., Ong, S.Y., Tham, Y.C., Hsu, W., Lee, M.L., Lau, Q.P., Tay, W., Alessi-
accelerated retinal ganglion cell/inner plexiform layer thinning. Neurology 80, Calandro, J., Hodgson, L., Kawasaki, R., Wong, T.Y., Cheung, C.Y., 2014. Mea-
47e54. surement of macular fractal dimension using a computer-assisted program.
Reid, C.M., Storey, E., Wong, T.Y., Woods, R., Tonkin, A., Wang, J.J., Kam, A., Janke, A., Invest Ophthalmol. Vis. Sci. 55, 2237e2243.
Essex, R., Abhayaratna, W.P., Budge, M.M., Group, A.S., 2012. Aspirin for the Thomson, K.L., Yeo, J.M., Waddell, B., Cameron, J.R., Pal, S., 2015. A systematic review
prevention of cognitive decline in the elderly: rationale and design of a neuro- and meta-analysis of retinal nerve fiber layer change in dementia, using optical
vascular imaging study (ENVIS-ion). BMC Neurol. 12, 3. coherence tomography. Alzheimer's Dement. Diagn. Assess. Dis. Monit. 1,
Rha, J., Jonnal, R.S., Thorn, K.E., Qu, J., Zhang, Y., Miller, D.T., 2006. Adaptive optics 136e143.
flood-illumination camera for high speed retinal imaging. Opt. Express 14, Tirsi, A., Bruehl, H., Sweat, V., Tsui, W., Reddy, S., Javier, E., Lee, C., Convit, A., 2009.
4552e4569. Retinal vessel abnormalities are associated with elevated fasting insulin levels
Rimmer, E., Stave, C., Sganga, A., O'Connell, B., 2005. Implications of the facing and cerebral atrophy in nondiabetic individuals. Ophthalmology 116,
dementia survey for policy makers and third-party organisations across Europe. 1175e1181.
Int. J. Clin. Pract. Suppl. 34e38. Tomita, Y., Kubis, N., Calando, Y., Tran, D.A., Meric, P., Seylaz, J., Pinard, E., 2005.
Rochtchina, E., Wang, J.J., Taylor, B., Wong, T.Y., Mitchell, P., 2008. Ethnic variability Long-term in vivo investigation of mouse cerebral microcirculation by fluo-
in retinal vessel caliber: a potential source of measurement error from ocular rescence confocal microscopy in the area of focal ischemia. J.Cereb. Blood Flow.
pigmentation?ethe Sydney Childhood Eye Study. Invest Ophthalmol. Vis.Sci. Metab. 25, 858e867.
49, 1362e1366. van Veluw, S.J., Hilal, S., Kuijf, H.J., Ikram, M.K., Xin, X., Yeow, T.B.,
Roorda, A., Romero-Borja, F., Donnelly Iii, W., Queener, H., Hebert, T., Campbell, M., Venketasubramanian, N., Biessels, G.J., Chen, C., 2015. Cortical microinfarcts on
2002. Adaptive optics scanning laser ophthalmoscopy. Opt. Express 10, 3T MRI: clinical correlates in memory-clinic patients. Alzheimers Dement. 11,
405e412. 1500e1509.
Rosenberg, G.A., Wallin, A., Wardlaw, J.M., Markus, H.S., Montaner, J., Wolfson, L., Venketasubramanian, N., Tan, L.C., Sahadevan, S., Chin, J.J., Krishnamoorthy, E.S.,
Iadecola, C., Zlokovic, B.V., Joutel, A., Dichgans, M., Duering, M., Schmidt, R., Hong, C.Y., Saw, S.M., 2005. Prevalence of stroke among Chinese, Malay, and
Korczyn, A.D., Grinberg, L.T., Chui, H.C., Hachinski, V., 2016. Consensus state- Indian Singaporeans: a community-based tri-racial cross-sectional survey.
ment for diagnosis of subcortical small vessel disease. J. Cereb. Blood Flow. Stroke 36, 551e556.
Metab. 36, 6e25. Vermeer, S.E., Hollander, M., van Dijk, E.J., Hofman, A., Koudstaal, P.J., Breteler, M.M.,
Sachdev, P., Kalaria, R., O'Brien, J., Skoog, I., Alladi, S., Black, S.E., Blacker, D., Rotterdam Scan, S., 2003. Silent brain infarcts and white matter lesions increase
Blazer, D.G., Chen, C., Chui, H., Ganguli, M., Jellinger, K., Jeste, D.V., Pasquier, F., stroke risk in the general population: the Rotterdam Scan Study. Stroke 34,
Paulsen, J., Prins, N., Rockwood, K., Roman, G., Scheltens, P., Internationlal So- 1126e1129.
ciety for Vascular, B., Cognitive, D., 2014. Diagnostic criteria for vascular Veronese, C., Marcheggiani, E.B., Tassi, F., Gallelli, I., Armstrong, G.W., Ciardella, A.P.,
cognitive disorders: a VASCOG statement. Alzheimer Dis. Assoc. Disord. 28, 2013. Fundus autofluorescence imaging in hereditary ATTR amyloidosis with
206e218. ocular involvement. Amyloid 20, 269e271.
Sagong, M., van Hemert, J., Olmos de Koo, L.C., Barnett, C., Sadda, S.R., 2015. Vuong, L.N., Thulasi, P., Biousse, V., Garza, P., Wright, D.W., Newman, N.J., Bruce, B.B.,
Assessment of accuracy and precision of quantification of ultra-widefield im- 2015. Ocular fundus photography of patients with focal neurologic deficits in an
ages. Ophthalmology 122, 864e866. emergency department. Neurology 85, 256e262.
Sahadevan, S., Saw, S.M., Gao, W., Tan, L.C., Chin, J.J., Hong, C.Y., Wainwright, A., Liew, G., Burlutsky, G., Rochtchina, E., Zhang, Y.P., Hsu, W., Lee, J.M.,
Venketasubramanian, N., 2008. Ethnic differences in Singapore's dementia Wong, T.Y., Mitchell, P., Wang, J.J., 2010. Effect of image quality, color, and format
prevalence: the stroke, Parkinson's disease, epilepsy, and dementia in on the measurement of retinal vascular fractal dimension. Invest Ophthalmol.
Singapore study. J. Am. Geriatr. Soc. 56, 2061e2068. Vis.Sci. 51, 5525e5529.
Saidha, S., Sotirchos, E.S., Ibrahim, M.A., Crainiceanu, C.M., Gelfand, J.M., Sepah, Y.J., Walton, O.B.t., Garoon, R.B., Weng, C.Y., Gross, J., Young, A.K., Camero, K.A., Jin, H.,
Ratchford, J.N., Oh, J., Seigo, M.A., Newsome, S.D., Balcer, L.J., Frohman, E.M., Carvounis, P.E., Coffee, R.E., Chu, Y.I., 2016. Evaluation of automated teleretinal

Please cite this article in press as: Cheung, C.Y.-l., et al., Imaging retina to study dementia and stroke, Progress in Retinal and Eye Research (2016),
http://dx.doi.org/10.1016/j.preteyeres.2017.01.001
 C.Y.-l. Cheung et al. / Progress in Retinal and Eye Research xxx (2016) 1e19 19

screening program for diabetic retinopathy. JAMA Ophthalmol. 134, 204e209. 2310e2317.
Wang, T.J., 2011. Assessing the role of circulating, genetic, and imaging biomarkers Wong, T.Y., Klein, R., Couper, D.J., Cooper, L.S., Shahar, E., Hubbard, L.D.,
in cardiovascular risk prediction. Circulation 123, 551e565. Wofford, M.R., Sharrett, A.R., 2001. Retinal microvascular abnormalities and
Wang, J.J., Liew, G., Klein, R., Rochtchina, E., Knudtson, M.D., Klein, B.E., Wong, T.Y., incident stroke: the atherosclerosis risk in Communities Study. Lancet 358,
Burlutsky, G., Mitchell, P., 2007. Retinal vessel diameter and cardiovascular 1134e1140.
mortality: pooled data analysis from two older populations. Eur. Heart J. 28, Wong, T.Y., Klein, R., Sharrett, A.R., Couper, D.J., Klein, B.E., Liao, D.P., Hubbard, L.D.,
1984e1992. Mosley, T.H., 2002. Cerebral white matter lesions, retinopathy, and incident
Wang, J.J., Baker, M.L., Hand, P.J., Hankey, G.J., Lindley, R.I., Rochtchina, E., Wong, T.Y., clinical stroke. JAMA 288, 67e74.
Liew, G., Mitchell, P., 2011. Transient ischemic attack and acute ischemic stroke: Wong, T.Y., Knudtson, M.D., Klein, R., Klein, B.E., Meuer, S.M., Hubbard, L.D., 2004a.
associations with retinal microvascular signs. Stroke 42, 404e408. Computer-assisted measurement of retinal vessel diameters in the Beaver Dam
Wang, D., Li, Y., Wang, C., Xu, L., You, Q.S., Wang, Y.X., Zhao, L., Wei, W.B., Zhao, X., Eye Study: methodology, correlation between eyes, and effect of refractive er-
Jonas, J.B., 2014. Localized retinal nerve fiber layer defects and stroke. Stroke 45, rors. Ophthalmology 111, 1183e1190.
1651e1656. Wong, T.Y., Knudtson, M.D., Klein, R., Klein, B.E., Meuer, S.M., Hubbard, L.D., 2004b.
Wardlaw, J.M., Smith, E.E., Biessels, G.J., Cordonnier, C., Fazekas, F., Frayne, R., Computer-assisted measurement of retinal vessel diameters in the Beaver Dam
Lindley, R.I., O'Brien, J.T., Barkhof, F., Benavente, O.R., Black, S.E., Brayne, C., Eye Study: methodology, correlation between eyes, and effect of refractive er-
Breteler, M., Chabriat, H., Decarli, C., de Leeuw, F.E., Doubal, F., Duering, M., rors. Ophthalmology 111, 1183e1190.
Fox, N.C., Greenberg, S., Hachinski, V., Kilimann, I., Mok, V., Oostenbrugge, R., Wong, T.Y., Kamineni, A., Klein, R., Sharrett, A.R., Klein, B.E., Siscovick, D.S.,
Pantoni, L., Speck, O., Stephan, B.C., Teipel, S., Viswanathan, A., Werring, D., Cushman, M., Duncan, B.B., 2006. Quantitative retinal venular caliber and risk of
Chen, C., Smith, C., van Buchem, M., Norrving, B., Gorelick, P.B., Dichgans, M., cardiovascular disease in older persons: the cardiovascular health study. Arch.
nEuroimaging, S.T.f.R.V.c.o, 2013. Neuroimaging standards for research into Intern Med. 166, 2388e2394.
small vessel disease and its contribution to ageing and neurodegeneration. Yatsuya, H., Folsom, A.R., Wong, T.Y., Klein, R., Klein, B.E., Sharrett, A.R., 2010. Retinal
Lancet Neurol. 12, 822e838. microvascular abnormalities and risk of lacunar stroke: atherosclerosis Risk in
Wassle, H., 2004. Parallel processing in the mammalian retina. Nat. Rev. Neurosci. 5, Communities Study. Stroke 41, 1349e1355.
747e757. Yildiz, M., Alim, S., Batmaz, S., Demir, S., Songur, E., Ortak, H., Demirci, K., 2016.
Wieberdink, R.G., Ikram, M.K., Koudstaal, P.J., Hofman, A., Vingerling, J.R., Duration of the depressive episode is correlated with ganglion cell inner
Breteler, M.M., 2010. Retinal vascular calibers and the risk of intracerebral plexifrom layer and nasal retinal fiber layer thicknesses: optical coherence to-
hemorrhage and cerebral infarction: the Rotterdam Study. Stroke 41, mography findings in major depression. Psychiatry Res. 251, 60e66.
2757e2761. Yip, W., Siantar, R., Perera, S.A., Milastuti, N., Ho, K.K., Tan, B., Wong, T.Y.,
Williams, M.A., McGowan, A.J., Cardwell, C.R., Cheung, C.Y., Craig, D., Passmore, P., Cheung, C.Y., 2014. Reliability and determinants of retinal vessel oximetry
Silvestri, G., Maxwell, A.P., McKay, G.J., 2015. Retinal microvascular network measurements in healthy eyes. Invest Ophthalmol. Vis. Sci. 55, 7104e7110.
attenuation in Alzheimer's disease. Alzheimer's Dementia Diagnosis, Assess. Yiu, G., Chiu, S.J., Petrou, P.A., Stinnett, S., Sarin, N., Farsiu, S., Chew, E.Y., Wong, W.T.,
Dis. Monit. 1, 229e235. Toth, C.A., 2015. Relationship of central choroidal thickness with age-related
Witt, N., Wong, T.Y., Hughes, A.D., Chaturvedi, N., Klein, B.E., Evans, R., macular degeneration status. Am. J. Ophthalmol. 159, 617e626.
McNamara, M., Thom, S.A., Klein, R., 2006. Abnormalities of retinal microvas- Zayit-Soudry, S., Duncan, J.L., Syed, R., Menghini, M., Roorda, A.J., 2013. Cone
cular structure and risk of mortality from ischemic heart disease and stroke. structure imaged with adaptive optics scanning laser ophthalmoscopy in eyes
Hypertension 47, 975e981. with nonneovascular age-related macular degeneration. Invest Ophthalmol. Vis.
Wong, T.Y., Mitchell, P., 2004. Hypertensive retinopathy. N. Engl. J. Med. 351, Sci. 54, 7498e7509.

Please cite this article in press as: Cheung, C.Y.-l., et al., Imaging retina to study dementia and stroke, Progress in Retinal and Eye Research (2016),
http://dx.doi.org/10.1016/j.preteyeres.2017.01.001
 International Journal of
Molecular Sciences

Review
Machine Learning and Novel Biomarkers for the Diagnosis of
Alzheimer’s Disease
Chun-Hung Chang 1,2,3 , Chieh-Hsin Lin 1,4,5,6, * and Hsien-Yuan Lane 1,2,4,7, *

1 Institute of Clinical Medical Science, China Medical University, Taichung 40402, Taiwan;
[email protected]
2 Department of Psychiatry & Brain Disease Research Center, China Medical University Hospital,
Taichung 40402, Taiwan
3 An Nan Hospital, China Medical University, Tainan 709025, Taiwan
4 Graduate Institute of Biomedical Sciences, China Medical University, Taichung 40402, Taiwan
5 Kaohsiung Chang Gung Memorial Hospital, Chang Gung University College of Medicine,
Kaohsiung 83301, Taiwan
6 School of Medicine, Chang Gung University, Taoyuan 83301, Taiwan
7 Department of Psychology, College of Medical and Health Sciences, Asia University, Taichung 41354, Taiwan
* Correspondence: [email protected] (C.-H.L.); [email protected] (H.-Y.L.);
Tel.: +886-7-7317123 (ext. 8753) (C.-H.L.); Fax: +886-7-7326817 (C.-H.L.)

Abstract: Background: Alzheimer’s disease (AD) is a complex and severe neurodegenerative disease
that still lacks effective methods of diagnosis. The current diagnostic methods of AD rely on cogni-
tive tests, imaging techniques and cerebrospinal fluid (CSF) levels of amyloid-β1-42 (Aβ42), total
tau protein and hyperphosphorylated tau (p-tau). However, the available methods are expensive
and relatively invasive. Artificial intelligence techniques like machine learning tools have being
increasingly used in precision diagnosis. Methods: We conducted a meta-analysis to investigate the





 machine learning and novel biomarkers for the diagnosis of AD. Methods: We searched PubMed, the
Citation: Chang, C.-H.; Lin, C.-H.;
Cochrane Central Register of Controlled Trials, and the Cochrane Database of Systematic Reviews
Lane, H.-Y. Machine Learning and for reviews and trials that investigated the machine learning and novel biomarkers in diagnosis
Novel Biomarkers for the Diagnosis of AD. Results: In additional to Aβ and tau-related biomarkers, biomarkers according to other
of Alzheimer’s Disease. Int. J. Mol. mechanisms of AD pathology have been investigated. Neuronal injury biomarker includes neurofilia-
Sci. 2021, 22, 2761. https://doi.org/ ment light (NFL). Biomarkers about synaptic dysfunction and/or loss includes neurogranin, BACE1,
10.3390/ijms22052761 synaptotagmin, SNAP-25, GAP-43, synaptophysin. Biomarkers about neuroinflammation includes
sTREM2, and YKL-40. Besides, D-glutamate is one of coagonists at the NMDARs. Several machine
Academic Editor: Matthias Schmitz learning algorithms including support vector machine, logistic regression, random forest, and naïve
Bayes) to build an optimal predictive model to distinguish patients with AD from healthy controls.
Received: 31 January 2021
Conclusions: Our results revealed machine learning with novel biomarkers and multiple variables
Accepted: 5 March 2021
may increase the sensitivity and specificity in diagnosis of AD. Rapid and cost-effective HPLC for
Published: 9 March 2021
biomarkers and machine learning algorithms may assist physicians in diagnosing AD in outpatient
clinics.
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in
published maps and institutional affil-
Keywords: machine learning; deep learning; AI; biomarker; Alzheimer’s disease
iations.

1. Introduction
Copyright: © 2021 by the authors. Alzheimer’s disease (AD), characterized by progressive memory loss and cognitive
Licensee MDPI, Basel, Switzerland. impairment, is the most common cause of dementia [1]. Approximately 5.7 million Ameri-
This article is an open access article can patients suffer from AD. In 2015, AD was ranked as the sixth leading cause of death in
distributed under the terms and the United States. This care is valued at more than $232 billion and contributes as a factor
conditions of the Creative Commons of extended risk of emotional anxiety and negative mental and physical health issues of
Attribution (CC BY) license (https:// caregivers. However, early and accurate diagnosis of AD may help save up to $7.9 trillion
creativecommons.org/licenses/by/ in medical and care expenses [2].
4.0/).

Int. J. Mol. Sci. 2021, 22, 2761. https://doi.org/10.3390/ijms22052761 https://www.mdpi.com/journal/ijms

Int. J. Mol. Sci. 2021, 22, 2761 2 of 12

Current diagnostic approaches for AD mainly depend on neurocognitive tests, brain

imaging, and cerebrospinal fluid (CSF) assays [3,4]. Deposition of amyloid plaques, neu-
rofibrillary tangles and significant synapse loss are noted in brain pathology in patients with
AD [5]. Diagnostic guidelines have included cerebrospinal fluid (CSF) levels of amyloid-
β1-42 (Aβ42), total tau protein and hyperphosphorylated tau (p-tau) [6,7]. CSF biomarkers
like Aβ42 and p-tau have been used for research purposes. However, these methods are
expensive and relatively invasive [8,9]. Besides, sensitivity and specificity of CSF Aβ42
and p-tau biomarkers have raised concerns about their clinical implication [6,10,11]. The
sensitivity of CSF Aβ42 ranges from 0.69 to 0.81 and specificity ranges from 0.44 to 0.89 [12].
Moreover, patients with AD are generally diagnosed late. If AD can be detected in early
stages before major brain damage develop, patients may benefit more from treatment.
Therefore, identifying biomarkers that can assist in detecting AD early or at onset is crit-
ical. Biomarkers can improve diagnostics and enable treatment initiation at the earliest
possible stage [13].
However, the causal link between Aβ and AD has not been proved. The role of
the amyloid-β in the definition, etiology and diagnosis of Alzheimer’s disease is ques-
tioned [14]. Pathological levels of Aβ and tau are noted in cognitively normal people.
About 20% of cognitively normal elderly exhibit neuropathological AD when restrictive
diagnostic criteria for Aβ and tau pathology are applied [15]. On the contrary, some
individuals clinically diagnosed with AD do not have Aβ pathology. Studies have re-
ported limited evidence of cerebral AD pathology in approximately 10–20% of individuals
clinically diagnosed with AD [16,17].
Therefore, biomarkers according to other mechanisms of AD pathology have been
investigated. Neuronal injury biomarker includes neurofiliament light (NFL). Biomarkers
about synaptic dysfunction and/or loss includes neurogranin, BACE1, synaptotagmin,
SNAP-25, GAP-43, synaptophysin. Biomarkers about neuroinflammation includes sTREM2,
and YKL-40 [12]. Besides, abnormal hyperfunction of the N-methyl-D-aspartate receptor
(NMDAR) has been found to be involved in synapse dysfunction and neurotoxicity of
AD mechanisms [18–20]. D-Serine, one of major coagonists at the NMDARs, has been
found to be related with NMDAR-mediated neurotoxicity [21,22]. On the other hand,
D -serine has been shown to increase adult hippocampal neurogenesis [23]. Several trials
have investigated CSF D-serine levels. Two studies have reported that CSF D-serine levels
in an AD group were significantly higher than in a control group [24,25].
A machine learning algorithm is one of artificial intellectual techniques for selecting the
best model from a set of alternatives to fit a set of observations. Machine learning algorithms
have several merits, including nonlinearity, fault tolerance, and real-time operation, thereby
making them suitable for complex applications [26]. Machine learning tools are being
increasingly used in precision psychiatry [27–29]. However, most of current studies have
used brain image equipment that is not feasible in clinical practice. A pilot study enrolling
31 healthy controls, 21 patients with MCI, and 133 patients with AD has used machine
learning models to predict AD. D-Glutamate is one of agonists at the NMDARs. The naïve
Bayes model and random forest model appeared to be the best models for determining
MCI and AD susceptibility, respectively (area under the receiver operating characteristic
curve: 0.8207 and 0.7900; sensitivity: 0.8438 and 0.6997; and specificity = 0.8158 and 0.9188,
respectively). Therefore, in the context of machine learning methods, we review various
research studies with novel biomarkers on diagnosis of AD.

2. Methods
Search Strategy
PubMed, Cochrane Systematic Reviews, and Cochrane Collaboration Central Register
of Controlled Clinical Trials databases were searched for studies on machine learning
and novel biomarkers for the diagnosis of Alzheimer’s disease from the earliest record to
January 2021. Review studies that investigated biomarker for dementia or AD patients
were analyzed and included trials and related review articles were manually reviewed for
Int. J. Mol. Sci. 2021, 22, 2761 3 of 12

relevant references. The search strings used are the following: “(machine learning OR deep
learning OR AI) AND biomarker AND (dementia OR Alzheimer’s disease.” This article
reviews and summarizes these clinical trials.

3. Results
3.1. Machine Learning Models in Alzheimer’s Disease
Machine learning is one of artificial intellectual techniques, used for classification,
regression, clustering, or normative modeling. Machine learning algorithms can be divided
into supervised models where the data are labeled, unsupervised algorithms where the
aim is to separate an unlabeled data into groups of related cases, and semi-supervised algo-
rithms including both labeled and unlabeled data [30]. A machine learning algorithm is a
procedure for selecting the best model from a set of alternatives that fit a set of observations.
Machine learning algorithms have several merits, including nonlinearity, fault tolerance,
and real-time operation, thereby making them suitable for complex applications [26].
There are dozens of machine learning algorithms have been developed. Common
algorithms are listed in Table 1. In this review, relevant studies with biomarkers and
machine learning models are listed in Table 2.

Table 1. Common machine learning algorithms.

Algorithm Learning Type Class Restriction Bias Preference Bias

Generally suitable for measuring
K-Nearest distance-based approximations; Preferred for distance-based
Supervised Instance based
Neighbors however, it is subject to problems
dimensionality
Preferred for problems in
Works on problems where the
which the probability is
Naive Bayes Supervised Probabilistic inputs are independent from
always greater than zero for
each other
each class
Decision Trees/ Becomes less useful on problems Preferred for problems with
Supervised Tree
Random Forests with low covariance categorical data
Works where there is a definite
Support Vector Decision Preferred for binary
Supervised distinction between two
Machines boundary classification problems
classification
Nonlinear
Neural Networks Supervised functional Little restriction bias Preferred for binary inputs
approximation
Generally works well for system
Hidden Markov Supervised/ Preferred for time-series data
Markovian information where the Markov
Models Unsupervised and memoryless information
assumption holds
Preferred for data that is in
groupings given some form
Clustering Unsupervised Clustering No restriction
of distance (Euclidean,
Manhattan, or others)
Depending on algorithm can
Matrix
Feature Selection Unsupervised No restriction prefer data with high mutual
factorization
information
Will work much better on
Feature Matrix
Unsupervised Must be a nondegenerate matrix matrices that don’t have
Transformation factorization
inversion issues
Preferred for data that is not
Bagging Meta-heuristic Meta-heuristic Will work on just about anything
highly variable

3.2. Amyloid-β1-42 (Aβ42), Tau Protein and Hyperphosphorylated Tau (p-tau) with Machine Learning
Several AD biomarkers have been studied including the deposition of pathological
amyloid (Aβ) and tau in the cerebrospinal fluid (CSF) [31], the brain metabolic change
derived from fluorodeoxyglucose positron emission tomography (FDGPET) [32], and the
Int. J. Mol. Sci. 2021, 22, 2761 4 of 12

structural change in the brain morphology measured from the magnetic resonance imaging
(MRI) [33].
MRI provides a direct measurement of brain structure as possible biomarker to differ-
ential a normal control brain and an AD brain. Numerous studies have been conducted to
explore the potential of MRI-based AD biomarkers [34]. The recent availability of a large
database of individuals with DAT in the Alzheimer’s Disease Neuroimaging Initiative
(ADNI) and other openly available databases has improved MRI-based AD biomarkers.
Machine learning provide a powerful tool to explore huge brain image data. Pilot studies
have tried to use machine learning to quantify structural MRI neurodegeneration pat-
terns of AD into dementia score on 8,834 images from ADNI, AIBL, OASIS, and MIRIAD
databases
Popuri and his colleagues propose a novel biomarker using structural MRI volume-
based features. They computed a similarity score for the participant’s structural patterns
relative to those observed in the DAT group. They employed ensemble-learning framework
combining structural features in most discriminative ROIs to create an aggregate measure
of neurodegeneration in the brain. The classifier model was trained on 423 stable normal
control (NC) and 330 DAT subjects. In this dataset, clinical diagnosis has the highest
certainty. In the next step, they independently validated on 8,834 unseen images from
ADNI, AIBL, OASIS, and MIRIAD Alzheimer’s disease (AD) databases. The results
revealed promising potential to predict the development of DAT depending on the time-
to-conversion (TTC). The prediction of classification on stable versus progressive mild
cognitive impairment (MCI) groups achieved an AUC of 0.81 for TTC of 6 months and
0.73 for TTC of up to 7 years. Therefore, their findings may help assessing the presence
of AD structural atrophy patterns in normal aging and MCI stages, and monitoring the
progression of the patient’s brain along with the disease course.
Abate and his colleagues evaluated a p53-misfolding conformation recognized by the
antibody 2D3A8, also named Unfolded p53 (U-p532D3A8+), in 375 plasma samples derived
from InveCe.Ab and PharmaCog/E-ADNI longitudinal trials [35]. They used machine
learning models with U-p532D3A8+ plasma levels, Mini-Mental State Examination (MMSE)
and apolipoprotein E epsilon-4 (APOEε4) to predict AD. Their results showed likelihood
risk in InveCe.Ab with an overall 86.67% agreement with clinical diagnosis. These ML
models predicted (AUC = 0.92) Aβ+—amnestic Mild Cognitive Impairment (aMCI) patients
who will develop AD in PharmaCog/E-ADNI, where subjects were stratified according
to Cerebrospinal fluid (CSF) AD markers (Aβ42 and p-Tau). These findings may support
U-p532D3A8+ plasma level as a promising additional candidate blood-based biomarker
for AD.

3.3. PET-Based Tau Biomarker with Machine Learning

Deep learning, developed from machine learning, has been used in a variety of
applications in the rapidly growing huge and complicated amount of medical imaging
data [36]. Many studies have been focusing on the application of deep learning to AD
research. The prediction of AD mainly rely on deep learning using neuroimaging data
such as magnetic resonance imaging (MRI) and/or amyloid positron emission tomography
(PET). However, MRI scans cannot visualize molecular pathological hallmarks of AD,
and amyloid PET cannot, without difficulty, visualize the progression of AD due to the
accumulation of amyloid-β early in the disease course with a plateau in later stages [37].
The presence and location of pathological tau deposition in the human brain are
well established [38]. Studies have analyzed AD-related neuropathology and generated a
staging algorithm to describe the anatomical distribution of tau [39]. Subsequent studies
have revealed that the topography of tau corresponds with the pathological stages of
neurofibrillary tangle deposition. Cross-sectional autopsy data indicated that AD-related
tau pathology begins with tau deposition in the medial temporal lobe (Braak stages I and
II), moves to the lateral temporal cortex and part of the medial parietal lobe (stages III
and IV), and eventually proceeds to broader neocortical regions (stages V and VI) [40].
Int. J. Mol. Sci. 2021, 22, 2761 5 of 12

Convolutional neural networks (CNNs) are novel and frequently used in deep learning.
A CNN combined with tau PET is novel because the resulting spatial characteristics
and interpretation differ considerably with those of amyloid PET, FDG PET, or MRI.
Furthermore, the regional location and topography of tau PET signals are considered
more crucial than those of other molecular imaging modalities, resulting in implications
for how a CNN interacts with such complex inputs as well as for the visualization of
informative features.
Choi et al. developed a deep CNN–based automatic image interpretation system that
could accurately predict future cognitive decline in patients with MCI by using FDG and
florbetapir PET [41]. Their PET images included those of 139 patients with AD, 171 patients
with MCI, and 182 healthy controls obtained from the ADNI database. Their deep CNN was
trained using three-dimensional (3D) PET volumes of AD and healthy controls as inputs.
By contrast, manually defined image feature extraction methods, such as quantification
using predefined regions of interest, were unnecessary for our approach. Furthermore, their
CNN used minimally processed images without spatial normalization, which is commonly
used in conventional quantitative analyses. The cognitive outcomes of patients with MCI
were predicted using this CNN. The prediction accuracy of the conversion of MCI to AD
was compared with the conventional feature-based quantification approach. Prediction
accuracy (84.2%) for conversion to AD in patients with MCI outperformed conventional
feature-based quantification approaches. ROC analyses revealed that the performance of
the CNN was significantly higher than that of the conventional quantification methods
(p < 0.05). The output scores of the CNN were strongly correlated with the longitudinal
changes in cognitive measurements (p < 0.05). These results demonstrate the feasibility of
deep learning as a practical tool for identifying predictive neuroimaging biomarkers.
Jo and his colleagues developed a novel deep learning-based framework. They used
tau PET to identify the morphological phenotypes of tau deposition [42]. With these
tau PET images, they used deep learning models to differentiate patients with AD from
healthy controls. This 3D CNN-based classification model yielded an average accuracy
of 90.8% according to five-fold cross-validation. In addition, the researchers used a layer-
wise relevance propagation (LRP) model to identify the brain regions in tau PET images
that contributed the most to the classification results. The most-identified brain regions
were the hippocampus, parahippocampus, thalamus, and fusiform. The LRP results
were consistent with those from the voxel-wise analysis in SPM12, indicating significant
focal AD–associated regional tau deposition in the bilateral temporal lobes, including the
entorhinal cortex. The AD probability scores calculated by the classifier were correlated
with brain tau deposition in the medial temporal lobe in patients with MCI (r = 0.43 and
r = 0.49 for early and late MCI, respectively). A deep learning framework combining
3D CNN and LRP algorithms can be used with tau PET images to identify informative
features for AD classification and may be feasible for the early detection of AD during the
prodromal stages.
Patients with AD exhibit early changes in the structural integrity of white matter
(WM). A pilot study investigated the use of diffusion tensor imaging (DTI) in assessing
WM alterations in the predementia stage of mild cognitive impairment (MCI) [43]. They
applied a Support Vector Machine (SVM) with DTI and volumetric magnetic resonance
imaging data from 35 amyloid-β42 negative MCI subjects (MCI-Aβ42−), 35 positive MCI
subjects (MCI-Aβ42+), and 25 healthy controls (HC) retrieved from the European DTI Study
on Dementia. The SVM was applied to DTI-derived fractional anisotropy, mean diffusivity
(MD), and mode of anisotropy (MO) maps. For comparison, they studied classification
based on gray matter (GM) and WM volume. The accuracies were up to 68% for MO and
63% for GM volume when distinguishing between MCI-Aβ42− and MCI-Aβ42+. When
separating MCI-Aβ42+ from HC, the accuracy was up to 77% for MD and a significantly
lower accuracy of 68% for GM volume. Therefore, their findings suggest that DTI data
provide better prediction accuracy than GM volume in predementia AD.
Int. J. Mol. Sci. 2021, 22, 2761 6 of 12

Qiu and his colleagues developed and validated an interpretable deep learning frame-
work for Alzheimer’s disease classification. The model was trained using clinically diag-
nosed Alzheimer’s disease and cognitively normal subjects from the Alzheimer’s Disease
Neuroimaging Initiative (ADNI) dataset (n = 417). They validated on three independent
cohorts: the Australian Imaging, Biomarker and Lifestyle Flagship Study of Ageing (AIBL)
(n = 382), the Framingham Heart Study (n = 102), and the National Alzheimer’s Coordinat-
ing Center (NACC) (n = 582). Their predicting model was consistent across datasets, with
mean area under curve values of 0.996, 0.974, 0.876 and 0.954 for the ADNI study, AIBL,
Framingham Heart Study and NACC datasets, respectively. Furthermore, their approach
exceeded the diagnostic performance of practicing neurologists (n = 11). The high-risk
cerebral regions recognized by the model closely with post-mortem histopathological
findings. Their framework provides a clinically adaptable strategy for using routinely
available imaging techniques such as MRI to generate nuanced neuroimaging signatures
for Alzheimer’s disease diagnosis. This approach may be generalized for linking deep
learning to pathophysiological processes in human disease [44].

3.4. N-Methyl-D-Aspartate Receptor (NMDAR)-Mediated Biomarkers with Machine Learning

N-Methyl-D-aspartate (NMDA) receptors play a very important role in cognitive
functions [45]. NMDAR-mediated glutamate is a major excitatory neurotransmitter in the
mammalian central nervous system (CNS) [46]. NMDAR and glutamate play fundamental
roles in synaptic plasticity and in the underlying molecular mechanisms of learning and
memory [47]. Their crucial roles in excitatory neurotransmission indicate that normal
signaling disruption through iGluRs is implicated in a wide range of neuropathological
disorders and diseases, especially Alzheimer’s disease (AD) [48,49].
L -Amino acids are predominant in Nature. But their molecular chirality, D -amino
acids, have been noted to affect protein folding, neuronal proliferation, and brain functional
laterality. Recently, D-form amino acids have been found to play a crucial role in cognitive
functions and psychiatric disorders [50]. Pilot studies have shown that D-amino acids
are novel neurotransmitters [51]. However, studies on D-glutamate are few. Its role in
neurocognitive function remains unclear. A pilot human study showed that D-glutamate
levels are associated with cognitive functions in patients with AD or MCI [52].
Previous trials have revealed that decreased plasma D-glutamate levels are associ-
ated with cognitive impairment in AD [52,53]. A study of 397 participants reported that
D -glutamate levels in patients with MCI, and AD were significantly decreased than those
of healthy controls (MCI: 1097.8 ± 284.0, mild AD: 1031.9 ± 775.8, moderate to severe
AD: 598.3 ± 551.9, healthy elderly: 1620.4 ± 558.2,). Furthermore, they found that MMSE
score was significantly correlated with D-glutamate level (adjusted R square = 0.344) [53].
Another trial enrolling 144 patients showed that the D-glutamate level was negatively corre-
lated with the cognitive functions using Alzheimer’s Disease Assessment Scale-Cognitive
Subscale (ADAS-cog) behavior scores (r = −0.177, p = 0.034) [52].
These findings raise the next question about the reason why the lower peripheral
D -glutamate levels are correlated with cognitive impairment. Brain image studies have
explored the relationship between glutamate and cognition. Wong and his colleagues
included eight patients with MCI, nine patients with AD, and 16 healthy elderly controls.
They found that reduced hippocampal glutamate in MCI and AD was associated with
episodic memory performance [54]. Vijayakumari and his colleagues using functional
magnetic resonance imaging (fMRI) study to study 15 patients with amnestic MCI and
22 age-, sex-, and education-matched healthy controls. They noted a significant increase
was observed in glutamate during a working memory task (both 0 back and 1 back) in
healthy controls, but no significant changes were detected in patients with MCI [55].
Pilot studies have used machine learning models with NMDAR-mediated biomarker
like D-amino acid oxidase activator (DAOA, also known as G72) protein level to detect
schizophrenia. They enrolled 149 participants including 89 patients with schizophrenia
and 60 healthy controls. The naive Bayes model using two factors like G72 rs1421292
Int. J. Mol. Sci. 2021, 22, 2761 7 of 12

and G72 protein. The naive Bayes showed the best model for disease susceptibility
(sensitivity = 0.7969, specificity = 0.9372, area under the receiver operating characteris-
tic curve (AUC) = 0.9356) [56]. Another machine learning- ensemble approach to predict
schizophrenia using biomarker in the N-methyl-D-aspartate receptor (NMDAR) and tryp-
tophan catabolic pathways [57]. The analysis revealed that the ensemble boosting model
with random under sampling [AUC = 0.9242 ± 0.0652; sensitivity = 0.8580 ± 0.0770;
specificity = 0.8594 ± 0.0760] performed best to predict complicated relationship between
schizophrenia and biomarkers.
Chang and his colleagues [58] enrolled 31 healthy controls, 21 patients with MCI, and
133 patients with AD. They measured serum D-glutamate levels using high-performance
liquid chromatography (HPLC). Cognitive deficit severity was assessed using the Clinical
Dementia Rating scale and the MMSE. The researchers employed four machine learn-
ing algorithms (SVM, logistic regression, random forest, and naïve Bayes) to build an
optimal predictive model for distinguishing patients with MCI or AD from healthy con-
trols. They found that the MCI and AD groups had lower plasma D-glutamate levels
(1097.79 ± 283.99 and 785.10 ± 720.06 ng/mL, respectively) than did healthy controls
(1620.08 ± 548.80 ng/mL). They found that the naïve Bayes model and random forest
model were the best models for determining MCI and AD susceptibility, respectively (area
under the receiver operating characteristic curve: 0.8207 and 0.7900; sensitivity: 0.8438
and 0.6997; and specificity = 0.8158 and 0.9188, respectively). Moreover, the total MMSE
score was positively correlated with D-glutamate levels (r = 0.368, p < 0.001). Multivariate
regression analysis indicated that D-glutamate levels were significantly associated with
the total MMSE score (B = 0.003, confidence interval: 0.002–0.005, p < 0.001). Their re-
sults showed that peripheral plasma D-glutamate levels were associated with cognitive
impairment and may therefore be a suitable peripheral biomarker for detecting MCI and
AD. They suggested rapid and cost-effective HPLC for biomarkers with machine learning
models may improve diagnosis of MCI and AD in outpatient clinics.
In addition to predicting Alzheimer’s disease, machine learning models may help
to predict NMDAR antagonists for new medication development. Recently, generative
deep learning models have been applied to de novo drug design as a means to expand the
amount of chemical space that can be explored for potential drug-like compounds. Schultz
and his colleagues assessed the application of a generative model to the NMDAR. They
investigated two primary objectives. First, the creation and release of a comprehensive
library of experimentally validated NMDAR phencyclidine (PCP) site antagonists to assist
the drug discovery community. Second, an analysis of both the advantages conferred
by applying such generative artificial intelligence models to drug design and the current
limitations of the approach. They applied and provided source code for, a variety of ligand-
and structure-based assessment techniques used in standard drug discovery analyses to
the deep learning- generated compounds. Finally, they found twelve candidate antagonists.
Further synthesis and experimental validation of these compounds are still required [59].

3.5. Metabolites Biomarkers with Machine Learning

Recently, pilot studies have examined blood metabolites as potential biomarkers for
AD [60]. It is easier to evaluate metabolites at blood than at CSF or brain tissue. Blood
metabolites may represent an essential aspect of the phenotype of an organism. Therefore,
they might act as a molecular fingerprint of AD.
Stamate and hiscolleagues [61] used the data from the European Medical Information
Framework for AD Multimodal Biomarker Discovery (EMIF-AD) [62]. They explored eight
hundred metabolites using machine learning (ML) algorithms to identify those individuals
with AD from dataset. They also compared the effectiveness of blood-based metabolites as
a predictor of AD with CSF markers.
In this study, they used ML algorithms including deep learning (DL), extreme gradient
boosting (XGBoost), and random forest (RF) algorithm. They enrolled 242 cognitively
normal (CN) participants and 115 patients with AD-type dementia. Three machine learning
Int. J. Mol. Sci. 2021, 22, 2761 8 of 12

models like DL, XGBoost, and RF were used to differentiate AD from CN. They internally
validated these models using nested cross validation (NCV). When using the test data,
DL showed the AUC of 0.85 (0.80 to 0.89), XGBoost showed 0.88 (0.86 to 0.89) and RF
showed 0.85 (0.83 to 0.87). When using CSF amyloid, p-tau and t-tau (together with age
and gender), XGBoost revealed the AUC values of 0.78, 0.83 and 0.87, respectively.

Table 2. Relevant studies on the biomarkers and predictive models of AD.

Study Biomarker Model Results

Classification performance on stable
versus progressive mild cognitive
impairment (MCI) groups achieved an
Popuri et al., 2020 [63] CSF [t-tau/Aβ1-42] ensemble-learning
AUC of 0.81 for TTC of 6 months and 0.73
for TTC of up to 7 years, achieving
state-of-the-art results.
These algorithms also accurately classify
(AUC = 0.92) Aβ+—amnestic Mild
Abate et al., 2020 [35] p53 Regression Tree (RT)
Cognitive Impairment (aMCI) patients
who will develop AD
Accuracy of prediction (84.2%) for
convolutional neural network conversion to AD in MCI patients
Choi et al., 2018 [41] amyloid
(CNN) outperformed conventional feature-based
quantification approaches.
Deep learning-based classification model
convolutional neural network
Jo et al., 2020 [42] tau of AD from CN yielded an average
(CNN)
accuracy of 90.8%
accuracies of up to 68% for MO and 63%
Support Vector Machine for GM volume when it came to
Dyrba et al., 2015 [43] amyloid-β42
(SVM) distinguishing between
MCI-Aβ42− and MCI-Aβ42+.
The naïve Bayes model and random
forest model appeared to be the best
models for determining MCI and AD
support vector machine,
susceptibility, respectively (area under
Chang et al. (2021) [1] D-glutamate logistic regression, random
the receiver operating characteristic
forest, and naïve Bayes
curve: 0.8207 and 0.7900; sensitivity:
0.8438 and 0.6997; and specificity = 0.8158
and 0.9188, respectively).
Deep Learning (DL), Extreme
DL produced the AUC of 0.85 (0.80–0.89),
Gradient
Stamate et al. 2019 [61] Metabolites biomarkers XGBoost produced 0.88 (0.86–0.89) and
Boosting (XGBoost) and
RF produced 0.85 (0.83–0.87).
Random Forest (RF)

4. Outlook and Future Direction

AD is human diseases. In addition to biomarkers in brain, CSF or blood, multiple risk
factors with several markers are under study. In addition to medications, multidomain
interventions, targeting multiple risk factors simultaneously, could be effective dementia
prevention strategies. However, multidomain interventions may be burdensome and not
universally acceptable. Coley and his colleague [64] investigated adherence rates and
predictors for all intervention components separately and simultaneously in the Finnish
Geriatric Intervention Study to Prevent Cognitive Impairment and Disability (FINGER)
and Multidomain Alzheimer Preventive Trial (MAPT). FINGER participants received a
2-year multidomain lifestyle intervention (physical training, cognitive training, nutritional
counseling, and cardiovascular monitoring). MAPT participants received a 3-year multido-
main lifestyle intervention (cognitive training, physical activity counseling, and nutritional
Int. J. Mol. Sci. 2021, 22, 2761 9 of 12

counseling) with either an omega-3 supplement or placebo. Adherence decreased with

increasing intervention complexity and intensity. It was highest for cardiovascular moni-
toring, nutritional counseling, and the omega-3 supplement, and lowest for unsupervised
computer-based cognitive training. The most consistent baseline predictors of adherence
were smoking and depressive symptoms. Reducing participant burden, maintaining in-
person contacts, and taking into account participant characteristics may increase adherence
in future trials with large sample size.
Increasing studies are investigating machine learning techniques with novel biomark-
ers as promising approaches to predict AD. Some studies have shown promising results.
However, some lack large sample sizes and the appropriate power, or not hypothesis-driven.
This review examined novel biomarkers including amyloid, tau protein, NMDAR-mediated
biomarkers, and metabolites biomarkers. Because many machine learning models have
no standard settings and guidelines, a robust comparison of these trials remains incom-
plete. However, brain image-based biomarkers with machine learning models especially
deep learning such as CNN are promising. Moreover, machine learning combined with
NMDAR-mediated biomarkers appear to be a new approach to predict the long-term cog-
nitive outcome. Therefore, although the field of machine learning is relatively immature,
such techniques, especially deep learning, warrant further study for their diagnostic and
therapeutic implications on patients with AD.

Author Contributions: Conceptualization: C.-H.L. and H.-Y.L.; writing, C.-H.C. All authors have
read, revised, and agreed to the published version of the manuscript.
Funding: This work was supported by the Ministry of Science and Technology, Taiwan (MOST
107-2628-B-182A-002; 108-2628-B-182A-002; 109-2628-B-182A-002; 109-2314-B-039-039-MY3), the
National Health Research Institutes (NHRI-EX109-10816NC; NHRI-EX110-10816NC), Kaohsiung
Chang Gung Me-morial Hospital (CMRPG8G1391; CMRPG8K1461), and China Medical University
Hospital (HHC-109-13).
Conflicts of Interest: The authors declare no conflict of interest. The sponsors were not involved
in the design of the study; the collection, analysis, and interpretation of the data; the writing of the
report; and the decision to submit the article for publication.

References
1. Schneider:, L. Alzheimer’s disease and other dementias: Update on research. Lancet Neurol. 2017, 16, 4–5. [CrossRef]
2. Weuve, J.; Hebert, L.E.; Scherr, P.A.; Evans, D.A. Prevalence of Alzheimer disease in US states. Epidemiology 2015, 26, e4–e6.
[CrossRef] [PubMed]
3. Zeng, H.M.; Han, H.B.; Zhang, Q.F.; Bai, H. Application of modern neuroimaging technology in the diagnosis and study of
Alzheimer’s disease. Neural Regen. Res. 2021, 16, 73–79. [CrossRef] [PubMed]
4. D’Abramo, C.; D’Adamio, L.; Giliberto, L. Significance of Blood and Cerebrospinal Fluid Biomarkers for Alzheimer’s Disease:
Sensitivity, Specificity and Potential for Clinical Use. J. Pers Med. 2020, 10, 116. [CrossRef] [PubMed]
5. Mattson, M.P. Pathways towards and away from Alzheimer’s disease. Nature 2004, 430, 631–639. [CrossRef] [PubMed]
6. McKhann, G.M.; Knopman, D.S.; Chertkow, H.; Hyman, B.T.; Jack, C.R., Jr.; Kawas, C.H.; Klunk, W.E.; Koroshetz, W.J.; Manly,
J.J.; Mayeux, R.; et al. The diagnosis of dementia due to Alzheimer’s disease: Recommendations from the National Institute on
Aging-Alzheimer’s Association workgroups on diagnostic guidelines for Alzheimer’s disease. Alzheimer’s Dement. J. Alzheimer’s
Assoc. 2011, 7, 263–269. [CrossRef]
7. Blennow, K.; Hampel, H.; Weiner, M.; Zetterberg, H. Cerebrospinal fluid and plasma biomarkers in Alzheimer disease. Nat. Rev.
Neurol. 2010, 6, 131–144. [CrossRef] [PubMed]
8. Atri, A. The Alzheimer’s Disease Clinical Spectrum: Diagnosis and Management. Med. Clin. N. Am. 2019, 103, 263–293. [CrossRef]
9. Davda, N.; Corkill, R. Biomarkers in the diagnosis and prognosis of Alzheimer’s disease. J. Neurol 2020, 267, 2475–2477. [CrossRef]
10. Jack, C.R., Jr.; Albert, M.S.; Knopman, D.S.; McKhann, G.M.; Sperling, R.A.; Carrillo, M.C.; Thies, B.; Phelps, C.H. Introduction to
the recommendations from the National Institute on Aging-Alzheimer’s Association workgroups on diagnostic guidelines for
Alzheimer’s disease. Alzheimer’s Dement. J. Alzheimer’s Assoc. 2011, 7, 257–262. [CrossRef]
11. Hampel, H.; Burger, K.; Teipel, S.J.; Bokde, A.L.; Zetterberg, H.; Blennow, K. Core candidate neurochemical and imaging
biomarkers of Alzheimer’s disease. Alzheimer’s Dement. J. Alzheimer’s Assoc. 2008, 4, 38–48. [CrossRef] [PubMed]
12. Balasa, A.F.; Chircov, C.; Grumezescu, A.M. Body Fluid Biomarkers for Alzheimer’s Disease-An Up-To-Date Overview.
Biomedicines 2020, 8, 421. [CrossRef]
Int. J. Mol. Sci. 2021, 22, 2761 10 of 12

13. Auso, E.; Gomez-Vicente, V.; Esquiva, G. Biomarkers for Alzheimer’s Disease Early Diagnosis. J. Pers Med. 2020, 10, 114.
[CrossRef] [PubMed]
14. Morris, G.P.; Clark, I.A.; Vissel, B. Questions concerning the role of amyloid-beta in the definition, aetiology and diagnosis of
Alzheimer’s disease. Acta Neuropathol. 2018, 136, 663–689. [CrossRef]
15. Price, J.L.; McKeel, D.W., Jr.; Buckles, V.D.; Roe, C.M.; Xiong, C.; Grundman, M.; Hansen, L.A.; Petersen, R.C.; Parisi, J.E.; Dickson,
D.W.; et al. Neuropathology of nondemented aging: Presumptive evidence for preclinical Alzheimer disease. Neurobiol. Aging
2009, 30, 1026–1036. [CrossRef] [PubMed]
16. Beach, T.G.; Monsell, S.E.; Phillips, L.E.; Kukull, W. Accuracy of the clinical diagnosis of Alzheimer disease at National Institute
on Aging Alzheimer Disease Centers, 2005–2010. J. Neuropathol. Exp. Neurol. 2012, 71, 266–273. [CrossRef]
17. Serrano-Pozo, A.; Qian, J.; Monsell, S.E.; Blacker, D.; Gomez-Isla, T.; Betensky, R.A.; Growdon, J.H.; Johnson, K.A.; Frosch, M.P.;
Sperling, R.A.; et al. Mild to moderate Alzheimer dementia with insufficient neuropathological changes. Ann. Neurol. 2014,
75, 597–601. [CrossRef]
18. Li, S.; Jin, M.; Koeglsperger, T.; Shepardson, N.E.; Shankar, G.M.; Selkoe, D.J. Soluble Abeta oligomers inhibit long-term
potentiation through a mechanism involving excessive activation of extrasynaptic NR2B-containing NMDA receptors. J. Neurosci.
2011, 31, 6627–6638. [CrossRef]
19. Li, S.; Hong, S.; Shepardson, N.E.; Walsh, D.M.; Shankar, G.M.; Selkoe, D. Soluble oligomers of amyloid Beta protein facilitate
hippocampal long-term depression by disrupting neuronal glutamate uptake. Neuron 2009, 62, 788–801. [CrossRef]
20. Shankar, G.M.; Bloodgood, B.L.; Townsend, M.; Walsh, D.M.; Selkoe, D.J.; Sabatini, B.L. Natural oligomers of the Alzheimer
amyloid-beta protein induce reversible synapse loss by modulating an NMDA-type glutamate receptor-dependent signaling
pathway. J. Neurosci. 2007, 27, 2866–2875. [CrossRef]
21. Radzishevsky, I.; Sason, H.; Wolosker, H. D-serine: Physiology and pathology. Curr. Opin. Clin. Nutr. Metab. Care 2013, 16, 72–75.
[CrossRef]
22. Wolosker, H.; Dumin, E.; Balan, L.; Foltyn, V.N. D-amino acids in the brain: Dserine in neurotransmission and neurodegeneration.
FEBS J. 2008, 275, 3514–3526. [CrossRef]
23. Sultan, S.; Gebara, E.G.; Moullec, K.; Toni, N. D-serine increases adult hippocampal neurogenesis. Front. Neurosci. 2013, 7, 155.
[CrossRef] [PubMed]
24. Madeira, C.; Lourenco, M.V.; Vargas-Lopes, C.; Suemoto, C.K.; Brandao, C.O.; Reis, T.; Leite, R.E.; Laks, J.; Jacob-Filho, W.;
Pasqualucci, C.A.; et al. D-serine levels in Alzheimer’s disease: Implications for novel biomarker development. Transl. Psychiatry
2015, 5, e561. [CrossRef] [PubMed]
25. Fisher, G.; Lorenzo, N.; Abe, H.; Fujita, E.; Frey, W.H.; Emory, C.; Di Fiore, M.M.; D’Aniello, A. Free D- and L-amino acids in
ventricular cerebrospinal fluid from Alzheimer and normal subjects. Amino Acids 1998, 15, 263–269. [CrossRef]
26. Li, R. Data Mining and Machine Learning Methods for Dementia Research. Methods Mol. Biol. 2018, 1750, 363–370. [CrossRef]
[PubMed]
27. Yang, S.; Bornot, J.M.S.; Wong-Lin, K.; Prasad, G. M/EEG-based Bio-markers to predict the Mild Cognitive Impairment and
Alzheimer’s disease: A Review from the Machine Learning Perspective. IEEE Trans. Biomed. Eng. 2019, 66, 2924–2935. [CrossRef]
[PubMed]
28. Duffy, I.R.; Boyle, A.J.; Vasdev, N. Improving PET Imaging Acquisition and Analysis with Machine Learning: A Narrative Review
With Focus on Alzheimer’s Disease and Oncology. Mol. Imaging 2019, 18, 1536012119869070. [CrossRef]
29. Lin, E.; Lin, C.H.; Lane, H.Y. Precision Psychiatry Applications with Pharmacogenomics: Artificial Intelligence and Machine
Learning Approaches. Int. J. Mol. Sci. 2020, 21, 969. [CrossRef]
30. Dwyer, D.B.; Falkai, P.; Koutsouleris, N. Machine Learning Approaches for Clinical Psychology and Psychiatry. Annu Rev. Clin.
Psychol. 2018, 14, 91–118. [CrossRef]
31. Ritchie, C.; Smailagic, N.; Noel-Storr, A.H.; Ukoumunne, O.; Ladds, E.C.; Martin, S. CSF tau and the CSF tau/ABeta ratio for
the diagnosis of Alzheimer’s disease dementia and other dementias in people with mild cognitive impairment (MCI). Cochrane
Database Syst. Rev. 2017, 3, CD010803. [CrossRef]
32. Popuri, K.; Balachandar, R.; Alpert, K.; Lu, D.; Bhalla, M.; Mackenzie, I.R.; Hsiung, R.G.; Wang, L.; Beg, M.F.; Alzheimer’s
Disease Neuroimaging, I. Development and validation of a novel dementia of Alzheimer’s type (DAT) score based on metabolism
FDG-PET imaging. NeuroImage. Clin. 2018, 18, 802–813. [CrossRef]
33. Frisoni, G.B.; Fox, N.C.; Jack, C.R., Jr.; Scheltens, P.; Thompson, P.M. The clinical use of structural MRI in Alzheimer disease. Nat.
Rev. Neurol. 2010, 6, 67–77. [CrossRef] [PubMed]
34. Weiner, M.W.; Veitch, D.P.; Aisen, P.S.; Beckett, L.A.; Cairns, N.J.; Green, R.C.; Harvey, D.; Jack, C.R., Jr.; Jagust, W.; Morris, J.C.;
et al. Recent publications from the Alzheimer’s Disease Neuroimaging Initiative: Reviewing progress toward improved AD
clinical trials. Alzheimer’s Dement. J. Alzheimer’s Assoc. 2017, 13, e1–e85. [CrossRef] [PubMed]
35. Abate, G.; Vezzoli, M.; Polito, L.; Guaita, A.; Albani, D.; Marizzoni, M.; Garrafa, E.; Marengoni, A.; Forloni, G.; Frisoni, G.B.; et al.
A Conformation Variant of p53 Combined with Machine Learning Identifies Alzheimer Disease in Preclinical and Prodromal
Stages. J. Pers Med. 2020, 11, 14. [CrossRef]
36. Jo, T.; Nho, K.; Saykin, A.J. Deep Learning in Alzheimer’s Disease: Diagnostic Classification and Prognostic Prediction Using
Neuroimaging Data. Front. Aging Neurosci. 2019, 11, 220. [CrossRef] [PubMed]
Int. J. Mol. Sci. 2021, 22, 2761 11 of 12

37. Palmqvist, S.; Scholl, M.; Strandberg, O.; Mattsson, N.; Stomrud, E.; Zetterberg, H.; Blennow, K.; Landau, S.; Jagust, W.; Hansson,
O. Earliest accumulation of beta-amyloid occurs within the default-mode network and concurrently affects brain connectivity.
Nat. Commun. 2017, 8, 1214. [CrossRef] [PubMed]
38. Cho, H.; Choi, J.Y.; Hwang, M.S.; Kim, Y.J.; Lee, H.M.; Lee, H.S.; Lee, J.H.; Ryu, Y.H.; Lee, M.S.; Lyoo, C.H. In vivo cortical
spreading pattern of tau and amyloid in the Alzheimer disease spectrum. Ann. Neurol. 2016, 80, 247–258. [CrossRef] [PubMed]
39. Schwarz, A.J.; Yu, P.; Miller, B.B.; Shcherbinin, S.; Dickson, J.; Navitsky, M.; Joshi, A.D.; Devous, M.D., Sr.; Mintun, M.S. Regional
profiles of the candidate tau PET ligand 18F-AV-1451 recapitulate key features of Braak histopathological stages. Brain 2016,
139, 1539–1550. [CrossRef] [PubMed]
40. Braak, H.; Alafuzoff, I.; Arzberger, T.; Kretzschmar, H.; Del Tredici, K. Staging of Alzheimer disease-associated neurofibrillary
pathology using paraffin sections and immunocytochemistry. Acta Neuropathol. 2006, 112, 389–404. [CrossRef] [PubMed]
41. Choi, H.; Jin, K.H.; Alzheimer’s Disease Neuroimaging, I. Predicting cognitive decline with deep learning of brain metabolism
and amyloid imaging. Behav. Brain Res. 2018, 344, 103–109. [CrossRef] [PubMed]
42. Jo, T.; Nho, K.; Risacher, S.L.; Saykin, A.J.; Alzheimer’s Neuroimaging, I. Deep learning detection of informative features in tau
PET for Alzheimer’s disease classification. BMC Bioinform. 2020, 21, 496. [CrossRef] [PubMed]
43. Dyrba, M.; Barkhof, F.; Fellgiebel, A.; Filippi, M.; Hausner, L.; Hauenstein, K.; Kirste, T.; Teipel, S.J.; EDSD Study Group.
Predicting Prodromal Alzheimer’s Disease in Subjects with Mild Cognitive Impairment Using Machine Learning Classification of
Multimodal Multicenter Diffusion-Tensor and Magnetic Resonance Imaging Data. J. Neuroimaging Off. J. Am. Soc. Neuroimaging
2015, 25, 738–747. [CrossRef] [PubMed]
44. Qiu, S.; Joshi, P.S.; Miller, M.I.; Xue, C.; Zhou, X.; Karjadi, C.; Chang, G.H.; Joshi, A.S.; Dwyer, B.; Zhu, S.; et al. Development
and validation of an interpretable deep learning framework for Alzheimer’s disease classification. Brain 2020, 143, 1920–1933.
[CrossRef]
45. Li, F.; Tsien, J.Z. Memory and the NMDA receptors. N. Engl. J. Med. 2009, 361, 302–303. [CrossRef] [PubMed]
46. Niciu, M.J.; Kelmendi, B.; Sanacora, G. Overview of glutamatergic neurotransmission in the nervous system. Pharm. Biochem.
Behav. 2012, 100, 656–664. [CrossRef] [PubMed]
47. Riedel, G.; Platt, B.; Micheau, J. Glutamate receptor function in learning and memory. Behav. Brain Res. 2003, 140, 1–47. [CrossRef]
48. Bleich, S.; Romer, K.; Wiltfang, J.; Kornhuber, J. Glutamate and the glutamate receptor system: A target for drug action. Int. J.
Geriatr. Psychiatry 2003, 18, S33–S40. [CrossRef]
49. Wang, R.; Reddy, P.H. Role of Glutamate and NMDA Receptors in Alzheimer’s Disease. J. Alzheimer’s Dis. 2017, 57, 1041–1048.
[CrossRef]
50. Genchi, G. An overview on D-amino acids. Amino Acids 2017, 49, 1521–1533. [CrossRef]
51. Mothet, J.P.; Snyder, S.H. Brain D-amino acids: A novel class of neuromodulators. Amino Acids 2012, 43, 1809–1810. [CrossRef]
52. Lin, C.H.; Yang, H.T.; Lane, H.Y. D-glutamate, D-serine, and D-alanine differ in their roles in cognitive decline in patients with
Alzheimer’s disease or mild cognitive impairment. Pharm. Biochem. Behav. 2019, 185, 172760. [CrossRef] [PubMed]
53. Lin, C.H.; Yang, H.T.; Chiu, C.C.; Lane, H.Y. Blood levels of D-amino acid oxidase vs. D-amino acids in reflecting cognitive aging.
Sci. Rep. 2017, 7, 14849. [CrossRef]
54. Wong, D.; Atiya, S.; Fogarty, J.; Montero-Odasso, M.; Pasternak, S.H.; Brymer, C.; Borrie, M.J.; Bartha, R. Reduced Hippocampal
Glutamate and Posterior Cingulate N-Acetyl Aspartate in Mild Cognitive Impairment and Alzheimer’s Disease Is Associated with
Episodic Memory Performance and White Matter Integrity in the Cingulum: A Pilot Study. J. Alzheimer’s Dis. 2020, 73, 1385–1405.
[CrossRef] [PubMed]
55. Vijayakumari, A.A.; Menon, R.N.; Thomas, B.; Arun, T.M.; Nandini, M.; Kesavadas, C. Glutamatergic response to a low load
working memory paradigm in the left dorsolateral prefrontal cortex in patients with mild cognitive impairment: A functional
magnetic resonance spectroscopy study. Brain Imaging Behav. 2019, 14, 451–459. [CrossRef]
56. Lin, E.; Lin, C.H.; Lai, Y.L.; Huang, C.H.; Huang, Y.J.; Lane, H.Y. Combination of G72 Genetic Variation and G72 Protein Level to
Detect Schizophrenia: Machine Learning Approaches. Front. Psychiatry 2018, 9, 566. [CrossRef] [PubMed]
57. Lin, E.; Lin, C.H.; Hung, C.C.; Lane, H.Y. An Ensemble Approach to Predict Schizophrenia Using Protein Data in the N-
methyl-D-Aspartate Receptor (NMDAR) and Tryptophan Catabolic Pathways. Front. Bioeng Biotechnol. 2020, 8, 569. [CrossRef]
[PubMed]
58. Chang, C.H.; Lin, C.H.; Liu, C.Y.; Huang, C.S.; Chen, S.J.; Lin, W.C.; Yang, H.T.; Lane, H.Y. Plasma D-glutamate levels
for detecting mild cognitive impairment and Alzheimer’s disease: Machine learning approaches. J. Psychopharmacol. 2021,
35, 0269881120972331. [CrossRef] [PubMed]
59. Schultz, K.J.; Colby, S.M.; Yesiltepe, Y.; Nunez, J.R.; McGrady, M.Y.; Renslow, R.S. Application and assessment of deep learning
for the generation of potential NMDA receptor antagonists. Phys. Chem. Chem. Phys. 2021, 23, 1197–1214. [CrossRef] [PubMed]
60. Varma, V.R.; Oommen, A.M.; Varma, S.; Casanova, R.; An, Y.; Andrews, R.M.; O’Brien, R.; Pletnikova, O.; Troncoso, J.C.; Toledo,
J.; et al. Brain and blood metabolite signatures of pathology and progression in Alzheimer disease: A targeted metabolomics
study. PLoS Med. 2018, 15, e1002482. [CrossRef]
61. Stamate, D.; Kim, M.; Proitsi, P.; Westwood, S.; Baird, A.; Nevado-Holgado, A.; Hye, A.; Bos, I.; Vos, S.J.B.; Vandenberghe, R.;
et al. A metabolite-based machine learning approach to diagnose Alzheimer-type dementia in blood: Results from the European
Medical Information Framework for Alzheimer disease biomarker discovery cohort. Alzheimers Dement. 2019, 5, 933–938.
[CrossRef] [PubMed]
Int. J. Mol. Sci. 2021, 22, 2761 12 of 12

62. Kim, M.; Snowden, S.; Suvitaival, T.; Ali, A.; Merkler, D.J.; Ahmad, T.; Westwood, S.; Baird, A.; Proitsi, P.; Nevado-Holgado, A.;
et al. Primary fatty amides in plasma associated with brain amyloid burden, hippocampal volume, and memory in the European
Medical Information Framework for Alzheimer’s Disease biomarker discovery cohort. Alzheimer’s Dement. J. Alzheimer’s Assoc.
2019, 15, 817–827. [CrossRef]
63. Popuri, K.; Ma, D.; Wang, L.; Beg, M.F. Using machine learning to quantify structural MRI neurodegeneration patterns of
Alzheimer’s disease into dementia score: Independent validation on 8,834 images from ADNI, AIBL, OASIS, and MIRIAD
databases. Hum. Brain Mapp. 2020, 41, 4127–4147. [CrossRef] [PubMed]
64. Coley, N.; Ngandu, T.; Lehtisalo, J.; Soininen, H.; Vellas, B.; Richard, E.; Kivipelto, M.; Andrieu, S.; van Gool, P.; van Charante, E.M.;
et al. Adherence to multidomain interventions for dementia prevention: Data from the FINGER and MAPT trials. Alzheimer’s
Dement. J. Alzheimer’s Assoc. 2019, 15, 729–741. [CrossRef] [PubMed]
 Neuropsychiatric Disease and Treatment Dovepress
open access to scientific and medical research

Open Access Full Text Article

ORIGINAL RESEARCH

Retinal Imaging Techniques Based on Machine

Learning Models in Recognition and Prediction of
Mild Cognitive Impairment
Neuropsychiatric Disease and Treatment downloaded from https://www.dovepress.com/ on 26-Sep-2023

Qian Zhang 1,2, * Background and Purpose: Mild Cognitive Impairment (MCI) is thought to be the signal
Jun Li 3, * of many progressive diseases but is easily ignored. Therefore, a simple and easy screening
Minjie Bian 1,2 method for recognizing and predicting MCI is urgently needed. The study aimed to establish
Qin He 1,2 machine learning models of retinal vascular features to categorize and predict MCI.
For personal use only.

Patients and Methods: Subjects enrolled underwent cognitive function assessment and were
Yuxian Shen 1,2
divided into a normal group, an MCI group, and a dementia group, and fundus photography was
Yue Lan 4
1,2 performed. MATLAB 2019b was used for fundus image preprocessing and vascular segmenta­
Dongfeng Huang
tion. Via the Green channel, adaptive histogram equalization (AHE), image binarization, and
1
Department of Rehabilitation Medicine, median filtering, we obtained the original and segmentation retinal vessel images. Afterwards,
The Seventh Affiliated Hospital, Sun Yat-
sen University, Shenzhen, People’s
the histogram of oriented gradient (HOG) was used for image feature extraction. Support vector
Republic of China; 2Guangdong machine (SVM) and extreme learning machine (ELM) were selected for training models in the
Engineering and Technology Research fundus original images and fundus vascular segmentation images, respectively. Among the three
Center for Rehabilitation Medicine and
Translation, Guangzhou, People’s cognitive groups, sensitivity, specificity, the receiver operating characteristic (ROC) curves, and
Republic of China; 3Department of the area under the curve (AUC) were used to evaluate and compare the predictive performance of
Urology, Kidney and Urology Center, The the two models in the fundus original and vascular segmentation images, respectively.
Seventh Affiliated Hospital, Sun Yat-sen
University, Shenzhen, People’s Republic of Results: A total of 86 eligible subjects were enrolled in the study. After a clinical cognitive
China; 4Department of Rehabilitation assessment, the participants were divided into the normal group (N = 38), the MCI group
Medicine, Guangzhou First People’s
(N = 26), and the dementia group (N = 22). A total of 332 qualified fundus images were
Hospital, Guangzhou Medical University,
Guangzhou, People’s Republic of China adopted after screening. Comparing the models among the three groups showed that the
SVM model had more advantages than the ELM model in the fundus original images and
*These authors contributed equally to
this work vascular segmentation images. Meanwhile, we found that the original images performed
better than the segmentation images in the same prediction model. Among the three groups,
the SVM model of the fundus original images had the best performance.
Conclusion: The establishment of a predictive model based on vascular-related feature
Correspondence: Yue Lan extraction from fundus images has high recognition and prediction abilities for cognitive
Department of Rehabilitation Medicine,
Guangzhou First People’s Hospital, function and can be used as a screening method for MCI.
Guangzhou Medical University, 151 Clinical Trial Registration: ChiCTR.org.cn (ChiCTR1900027404), Registered on Nov 12,
Yanjiang Road, Guangzhou, 510030,
People’s Republic of China 2019.
Tel +86-18988991916 Keywords: retinal imaging techniques, mild cognitive impairment, machine learning,
Email [email protected]
support vector machine, extreme learning machine
Dongfeng Huang
Department of Rehabilitation Medicine,
The Seventh Affiliated Hospital, Sun Yat-
sen University, 628 Zhenyuan Road, Introduction
Shenzhen, 518107, People’s Republic of
China Mild Cognitive Impairment
Tel +86-13322800919 With the ageing of the population, the number of people with mild cognitive
Fax +86-0755-81206511
Email [email protected] impairment (MCI) or dementia is expected to increase. There are over

Neuropsychiatric Disease and Treatment 2021:17 3267–3281 3267

Received: 11 August 2021 © 2021 Zhang et al. This work is published and licensed by Dove Medical Press Limited. The full terms of this license are available at https://www.dovepress.com/terms.
Accepted: 27 October 2021 php and incorporate the Creative Commons Attribution – Non Commercial (unported, v3.0) License (http://creativecommons.org/licenses/by-nc/3.0/). By accessing the
work you hereby accept the Terms. Non-commercial uses of the work are permitted without any further permission from Dove Medical Press Limited, provided the work is properly attributed. For
Published: 6 November 2021 permission for commercial use of this work, please see paragraphs 4.2 and 5 of our Terms (https://www.dovepress.com/terms.php).
 Zhang et al Dovepress

55 million people worldwide living with dementia in Machine Learning

2020. This number will almost double every 20 years, In recent years, machine learning (ML) has become the hot
reaching 78 million in 2030 and 139 million in 2050.1 spot in neurocognition study due to great advantages and
Such growth will inevitably lead to a substantial wide applications in feature training and group classifica­
increase in the economic and health burden on societies tion. ML is an artificial intelligence instrument that pre­
worldwide. dicts future data based on learning data from the past.
MCI, which is defined by the Diagnostic and There are two main types of ML: supervised learning
Statistical Manual of Mental Disorders 5th Edition and unsupervised learning. Supervised learning methods,
(DSM-V), refers to a clinical condition between normal such as neural networks (NNs), support vector machines
ageing and dementia. Patients have a subjective and (SVMs), K-nearest algorithms, linear and logistic regres­
objective decline from a previous level of functioning sion, and random forest (RF) algorithms, build a model by
in one or more of the six cognitive domains but their
learning from known classes (labelled training data). In
activities of daily living are not significantly affected
contrast, unsupervised learning methods, including princi­
and the problems are not severe enough to meet the
pal component analysis (PCA), are to classify to predict
diagnosis criteria for dementia.2 There are more than
for unknown data (unlabelled training data).8
six main cognitive domains that could be affected poten­
With advances in ML and computing techniques,
tially, such as learning and memory, complex attention,
developing computerized-aided diagnostics to assist spe­
executive function, language, visuospatial and structural
cialists in the diagnosis and prediction of diseases has been
function, and social cognition. These domains are hier­
recognized as an important area of medical research and
archical in nature, with the bottom referring to more
development. As a branch of ML, deep learning (DL) is an
basic sensory and perceptual processes and the top
algorithm of representational learning based on artificial
referring to elements of executive functioning and cog­
neural networks. Simulating the structure and function of
nitive control. They are not independent of each other
the nervous system of the human brain, DL learns repre­
and executive functioning exerts control over the utili­
sentation information automatically by transforming the
zation of more basic processes. Among the domains,
input data into multiple levels of abstraction. In the deeply
learning and memory is considered to be the most
stacked neural network, the output results of the upper
vulnerable.3
neural layer serve as the input data of the next, and the
In the past, MCI was thought simply to be a “precursor”
phase of Alzheimer’s disease (AD).4,5 Current research has final layer reveals the output diagnostic results. Each neu­
shown that MCI is not synonymous with AD and does not ron in the layer has its own weight for each input connec­
always progress to dementia; fortunately, some patients revert tion that represents the strength for the particular
to the range of normal cognition.6 However, there are many connection and the bias value, which allows us to shift
diseases known to have a positive association with MCI, such the activation function to the neuron with the weighted
as Parkinson’s disease, cerebrovascular diseases, and sum of the inputs and to control the trigger value of the
Huntington’s disease.7 Initially, these progressive diseases activation function. In order to create the output for the
may appear as mild cognitive behavioural impairment, which neuron and introduce non-linearity to the neuron decision,
is easily overlooked. In principle, the diagnosis of MCI is one of the activation functions is applied to the neuron
decided on the basis of clinical judgment whether the patients output.9 There are several types of deep neural networks,
are absence or presence of cognitive impairment. Afterwards, including convolutional neural networks (CNNs) and
specialists conduct the comprehensive neuropsychological test recurrent neural networks (RNNs). CNNs are known to
that can help to diagnose MCI and assess impaired cognitive deal with image data, while RNNs are utilized for sequen­
domains further, which should take into account age and tial input data.
education of the patients. Therefore, universal screening for As a clinical decision-making technology, DL can pro­
MCI is important to facilitate earlier diagnosis, and potential vide decision support to clinicians and improving the
progression may be prevented or postponed with treating accuracy and efficiency of various diagnostic and treat­
reversible causes and making lifestyle changes. Nowadays, ment processes by extracting features directly from the
the early diagnosis of MCI has been the focus of the recent raw data and generate an appropriate model for the target
research. task.10 DL is also frequently applied to improve the

3268 https://doi.org/10.2147/NDT.S333833 Neuropsychiatric Disease and Treatment 2021:17

DovePress

Powered by TCPDF (www.tcpdf.org)

 Dovepress Zhang et al

sensitivity and specificity of disease detection and surveil­ fluid.19 Recently, several clinical investigations have
lance, increase objectivity in clinical decision making, attempted to determine whether the retinal vasculature
significantly enhance identification and diagnostic ima­ could be an imaging biomarker for cognitive
ging, and contribute to a wide range of screening impairment.20,21 They found that the amplitude of retinal
programs.11 venous pulsations correlated negatively with the neocorti­
cal Aβ scores while the amplitude of retinal arterial pulsa­
Retinal Imaging Techniques tions correlated positively with neocortical Aβ scores.
Retinal imaging technique is commonly used in the field Meanwhile, they also found that the retinal ganglion cell
of neurology and ophthalmology. Currently, coloured digi­ layer (RGCL) thickness was significantly lower in the
tal imaging technique of retina captured by fundus camera clinical AD group.22 Studies have also shown that retinal
is one of the standard methods to record the vascular vascular geometry changes, including retinal vascular cali­
appearance of the retina. Fundus vascular morphology bre, retinal vascular tortuosity, retinal vascular fractal
can be observed directly by ophthalmoscope. Optical dimension,23 and retinal nerve layer thickness,20 are cor­
coherence tomography angiography (OCT-A) is a non- related with cognitive impairment.
invasive imaging technique providing depth-resolved Changes in the retinal vessels to a certain extent can
images of blood flow in the retina, choroid, and optic reflect the extent of a lesion by fundus examination.
nerve.12 The three methods mentioned above have been Clinicians can extract specific features from fundus images
applied in the clinical practice. to suggest reactions or changes from systemic diseases,
Literature has indicated that the retina and the central such as diabetes, multiple sclerosis, and cognitive impair­
nervous system are homologous. Embryologically, the ment, and to assist in making correct diagnoses and
retina, the epitaxial part of the diencephalon, originates treatments.24
from the neural tube. During early development, the retina
has a similar pattern of angiogenesis to the brain.13,14 Retinal Vessel Segmentation Technology
Meanwhile, the retinal vasculature is the only part of an Machine learning based retinal vessel segmentation tech­
individual’s cerebral microcirculation that can be directly nology for fundus images is one of the most important
visualized.15 Therefore, the observation of changes in the segmentation methods at present and is the basis of clin­
retinal vasculature can be used as a window to evaluate the ical auxiliary diagnosis systems and large-scale screening
continuous changes in the central microcirculation. systems.25
Recently, a growing body of evidence has demon­ Due to the interference of noise sources during image
strated that the nerve fiber and vascular pathologic changes acquisition, transmission and storage, and the characteris­
of the retina and brain are significantly correlated.16 The tics of blood vessels in fundus images, it is difficult to
study reported that retinal nerve fiber layer (RNFL) thin­ extract and segment blood vessels as areas of interest
ning was associated with cingulate cortex atrophy and effectively. In order to extract the vascular components
episodic memory decline in old participants.17 Some evi­ in fundus images and obtain the ideal segmentation effect,
dence has also shown that the RNFL was thinner in the the contrast between the blood vessels and the background
superior quadrant in patients with AD compared to the was increased, the greyscale of the background part of the
healthy groups, while the RNFL thickness was not differed vasculature consistent was made, and necessary image
significantly in the inferior, nasal, and temporal preprocessing was carried out before vascular segmenta­
quadrants.18 tion. In the MATLAB environment, blood vessels in fun­
In addition to the changes of nerve fiber layer thick­ dus image pretreatment technology as well as theoretical
ness, there is a difference in vascular morphology with analysis and research combined with fundus image back­
cognitive impairment. The researchers found that the arter­ grounds are studied. Fundus images, by the red, green, and
ial dilation and reaction amplitude of retinal vessels were blue (RGB) channels, enhance image contrast, greyscale
decreased in the AD group in comparison with the control inversion and enhancement, and image filtering, realize
group, and it demonstrated that Alzheimer’s and MCI fundus image preprocessing to obtain clear vascular eye­
subjects were characterized by significant impairment of ground veins from digital images and enhance vascular
the retinal neurovascular coupling which was inversely characteristics.26 Computer image processing technology
correlated with the level of amyloid β in the cerebrospinal is applied to the analysis and processing of coloured

Neuropsychiatric Disease and Treatment 2021:17 https://doi.org/10.2147/NDT.S333833

3269
DovePress

Powered by TCPDF (www.tcpdf.org)

 Zhang et al Dovepress

fundus images, and the blood vessels of the images are (MMSE), Montreal Cognitive Assessment scale (MoCA),
segmented and quantified. This approach greatly improves Clinical Dementia Rating scale (CDR), Physical Activities
the utilization of medical resources, which has become an of Daily Living (PADL), Instrumental Activities of Daily
important direction in the research field of computer-aided Living (IADL), and fundus photography.
diagnosis in recent years. The study was approved by the ethic committee at the
In this study, features from coloured retinal vascular Seventh Affiliated Hospital of Sun Yat-sen University,
images are extracted and analysed in a comprehensive where all data collection occurred, complies with the prin­
manner based on ML to identify and predict MCI. This ciples of the Declaration of Helsinki. The study has been
article aims to screen out populations at high risk for MCI registered in the China Clinical Trial Registry
as early as possible by ML models, to make clinical (ChiCTR1900027404).
diagnoses in time, and to determine effective intervention
means, which is of great significance for achieving the Cognitive Impairment Grouping
secondary prevention goals of early detection, early diag­ MCI Diagnostic Criteria
nosis, and early treatment, delaying the progression of the In the 11th Revision of the International Classification of
disease, and improving the lives of elderly individuals. Diseases (ICD-11), the World Health Organization (WHO)
Moreover, examining the pathophysiology of the retina adopted the definition MCI as ‘mild neurocognitive
in these earliest stages of dementia is scientifically valu­ disorders’.27,28 It includes the following four main points:
able for the development of a screening measure that can a) subjective experience of a decline from a previous level
be administered by point-of-care clinicians to large popu­ of cognitive functioning; b) accompanying objective evi­
lations. It is important for the coming ageing of the popu­ dence of impairment in performance on one or more
lation in 2050 as well. cognitive domains relative to that expected given the indi­
vidual’s age and general level of intellectual functioning,
and that is not sufficiently severe to significantly interfere
Materials and Methods
with independence in the person’s performance of activ­
Study Design ities of daily living; c) cognitive impairment that is not
This study was a cross-sectional study. Eighty-six subjects
entirely attributable to normal ageing; and d) cognitive
were recruited from the outpatient department and the impairment that may be attributable to an underlying dis­
inpatient department of rehabilitation medicine at the ease of the nervous system, a trauma, an infection, or other
Seventh Affiliated Hospital of Sun Yat-sen University disease processes affecting specific areas of the brain, to
from November 2019 to August 2020. All participants, chronic use of specific substances or medications, or to an
between the age of 18 and 80 years, including stroke etiology that may be undetermined. These points align
patients and non-stroke subjects, were provided with writ­ with the DSM-5 diagnostic guidelines as well.2
ten consent for study inclusion and offered information for
the case report form (CRF). The exclusion criteria were Criteria for Grouping Cognitive Functions
unstable vital signs, complications with severe chronic The subjects were grouped based on the MCI diagnostic
diseases which are difficult to complete clinical assess­ criteria and a clinical cognitive scale assessment.
ments, and contraindications for completing fundus According to the clinical experts’ advice and the relative
photography. literature for analysis, the subjects who were divided into
The CRF was used to collect the demographic infor­ MCI group met the following requirements: the range of
mation and the clinical information of the subjects, such as MMSE score was a loss of 3 from normal (normal standard
age, sex, and time of education. All enrolled stroke according to educational level: illiterate group >19, primary
patients conformed to the relevant diagnostic criteria for group >22, junior secondary group and above >26); MoCA
cerebral infarction and cerebral haemorrhage, which were scale scores ranged from 19 to 26, the CDR score was 0.5,
formulated by MRI or CT examination, and they also and PADL and IADL were not apparently abnormal. The
completed the internal carotid and vertebral artery ultra­ rest of the subjects were classified into the normal group or
sonography and heart ultrasonography simultaneously. cognitive impairment group.29–31 Afterwards, each cognitive
Meanwhile, they all underwent cognitive function assess­ group was divided into a training set and validation set at
ments, including the Mini-Mental State Examination scale a ratio of 2:1 according to the random tables.

3270 https://doi.org/10.2147/NDT.S333833 Neuropsychiatric Disease and Treatment 2021:17

DovePress

Powered by TCPDF (www.tcpdf.org)

 Dovepress Zhang et al

Personnel Training and Quality Monitoring MATLAB 2019b (MathWorks, USA) was used for
Researchers were organized to conduct training of clinical image preprocessing and segmentation, as presented in
cognition scales before the study began and discussed pro­ Figure 1, via the Green channel, adaptive histogram equal­
blems and inconsistent opinions throughout the survey to ization (AHE), image binarization, and median filtering, to
ensure the identical evaluation criteria. During the evalua­ obtain the original and segmentation retinal vessel images.
tion, the investigators introduced the purpose and content of The preprocessed fundus images were randomly horizon­
the study for participants and obtained informed consent and tally translated, flipped, randomly enlarged, miscut trans­
signatures from the subjects. Then, each subjects completed formed and filled to increase the diversity of the training
all the assessments and fundus photography in the study with images. Then, the images were normalized. Finally, the
the same day. The researchers completed the CRF within one retinal vessel images were inputted for feature extraction
week. Meanwhile, one person in charge and another person and constructed machine learning models.32
as quality controller of the study supervised the research
process and controlled the quality of all the research data. AHE
Histogram equalization (HE) is a commonly used histo­
Fundus Image Acquisition, Image gram method. The basic idea is to determine a mapping
Preprocessing and Segmentation curve through the histogram of the grey distribution of the
RetiCam 3100, the fully automatic and free mydriatic fundus image, which is used to transform the grey level of the
camera, was used for fundus imaging. Total four fundus image to achieve the purpose of improving the image
images were taken, two of which belong to one eye with contrast. AHE is a computer image processing technique
their centre on the optic disc and the macula, respectively. used to improve image contrast. Different from ordinary

Figure 1 Fundus images centred on the macula and optic disc. (A) Original fundus images centred on the macula; (B) adaptive histogram equalization (AHE) fundus images
centred on the macula; (C) segmentation fundus images centred on the macula; (a) Original fundus images centred on the optic disc; (b) adaptive histogram equalization
(AHE) fundus images centred on the optic disc; (c) segmentation fundus images centred on the optic disc.

Neuropsychiatric Disease and Treatment 2021:17 https://doi.org/10.2147/NDT.S333833

3271
DovePress

Powered by TCPDF (www.tcpdf.org)

 Zhang et al Dovepress

HE algorithms, the AHE algorithm calculates the local (b) Image gradient calculation: the gradient of the
histogram of the image and then redistributes the bright­ abscissa and ordinate directions of the image was
ness to change the image contrast. calculated, and the gradient direction value of each
pixel position was calculated accordingly. The deri­
Median Filtering vation not only can capture contours and some
Median filtering based on the theory of order statistics of textural information but also further attenuate the
a nonlinear signal processing technology can effectively effect of lighting.
restrain noise, and it is the basic principle of the digital (c) Histogram of gradient direction construction by cell
image or the value of the point in the sequence in which units: the images were divided into multiple cell
the values at various points in the field of a median units of the same size, the gradient direction was
instead. Then, the surrounding pixel values close to the divided into 9 bins, and the interval of each direc­
real value are ordered to remove isolated noise points. tion was 0–20 degrees.
(d) Block normalization: the cells were grouped into
Extract Image Features and Train Machine large, spatially connected sections. In this way, the
Learning Models feature vectors of the cells within a block were
Model Selection connected in series to obtain the HOG feature of
The support vector machine (SVM) is a general type of the block. The gradient intensity varied over a wide
feedforward network that is mainly used to solve the range due to local variations in illumination and
problem of data classification in pattern recognition fields. variations in the foreground-background contrast.
The SVM is based on nonlinear mapping theory to find the Therefore, this required the normalization of the
optimal hyperplane of feature space division to help in gradient intensity.
finding the key samples and eliminating the redundant (e) HOG feature calculation: the last step was to col­
samples with good robustness.33 lect all overlapping blocks in the detection window
Extreme learning machines (ELM) are learning algorithms for HOG feature collection and combine them into
for solving a single hidden layer feedforward neural network the final feature vector for classification.
that does not need to adjust the hidden nodes. Its greatest
advantage is that it is faster than the traditional learning algo­ In this paper, the size of the normalized image was 80×80
rithm on the premise of ensuring the learning accuracy.34 pixels, the size of the block area was 40×40 pixels, the size
The above two models were selected for training in this of each cell was 8×8 pixels, and the pixel with sliding step
study. length was 8×8 pixels. The gradient direction was divided
into 9 bins. That is, the number of eigenvectors was [(80–
Extract Features and Model Training 8)/8+1]2*(2*2) *9=3600. Finally, the feature vector was
Extracting image features and training machine learning input into the model for training as well as self-verification
models were implemented in Python 3.6 with OpenCV. In and verification.
this study, feature extraction and model training were
carried out on the original and segmentation images of Statistical Analysis
fundus images, respectively. The workflow is presented Python 3.6 was used for data processing. The result of the
in Figure 2. machine learning models was estimated using a ten-fold
The histogram of oriented gradient (HOG) feature is cross-validation approach. The receiver operating charac­
a feature descriptor used for object detection in computer teristic (ROC) curves of sensitivity and specificity and the
vision and image processing. It calculates and statistically area under the curve (AUC) were evaluated to identify and
analyses a histogram of gradient direction in the local area predict impaired cognitive function. Sensitivity was used
of the image to form the feature. The feature extraction to describe the proportion of actual cognitive impairment
process was as follows: and judged as positive (sensitivity = true positive/(true
positive + false negative)), and specificity was used to
(a) Gamma standardization: the image was greyed first, describe the proportion of actual non-cognitive impairment
and then the gamma space was processed to reduce and judged as negative (specificity = true negative/(true
the sensitivity of the light. negative + false positive)). This paper compared the

3272 https://doi.org/10.2147/NDT.S333833 Neuropsychiatric Disease and Treatment 2021:17

DovePress

Powered by TCPDF (www.tcpdf.org)

 Dovepress Zhang et al

Figure 2 The workflow of the study. (A) Flowchart of the study; (B) graphical abstracts of the study.

Neuropsychiatric Disease and Treatment 2021:17 https://doi.org/10.2147/NDT.S333833

3273
DovePress

Powered by TCPDF (www.tcpdf.org)

 Zhang et al Dovepress

sensitivity, specificity and AUC of the original retinal model were 0.85, 0.81, and 0.81 for the three groups in
vascular fundus images and segmentation images for clas­ the fundus vascular segmentation images, while the AUCs
sifying and predicting cognitive impairment and compared of the ELM model were 0.73, 0.80, and 0.79, suggesting
the classification effects of different machine learning that the fundus vascular segmentation images achieved
models. A two-sided t-test was used for statistical analysis, good predictive efficacy as well. Also, the AUCs of the
and P < 0.05 was considered statistically significant. fundus vascular segmentation images between the two
models were compared, and the results showed that the
Results difference was statistically significant (P = 0.0441),
Demographic Characteristics demonstrating the SVM model was better than ELM
A total of 86 eligible subjects were included in the study, model.
with an average age of 46.37 ± 1.79 years. Of these In addition, the fundus original and vascular segmenta­
patients, the following cases were noted: 25 with cerebral tion images were compared between the same prediction
infarction, 13 with cerebral hemorrhage, and 48 with non- models. For the SVM model, the AUCs of the fundus
stroke subjects. After comprehensive cognitive assess­ original images were 0.85, 0.87, and 0.86 for the normal
ment, the participants were divided into a normal group group, MCI group, and dementia group, while the AUCs
(N = 38), an MCI group (N = 26), and a dementia group of the segmentation images were 0.85, 0.81, and 0.81, and
(N = 22), and then each group was randomly divided into the results showed that the difference was statistically
a training set and a validation set, as shown in Table 1. significant (P = 0.0388), suggesting the fundus original
Significant differences were observed among the three images performed better than the vascular segmentation
groups in terms of age, educational status, MMSE, images. For the ELM model, the results indicated that the
MoCA, CDR, PADL and IADL, but not sex. A total of difference was statistically significant (P = 0.0052), sug­
332 qualified fundus images were adopted after screening. gesting the fundus original images performed better than
Twelve images were removed due to poor exposure, over- the vascular segmentation images as well. In the fundus
exposure, blurriness, or out-of-focus. original images, the AUCs of the SVM model in the above
three groups were 0.87, 0.88 and 0.90 in the training set
Comparison of Each Cognitive Function and 0.85, 0.87, and 0.86 in the validation set, respectively,
Group Based on Models and the Fundus which was the best performance of all.

Original and Vascular Segmentation

Images Discussion
The fundus original and vascular segmentation images of Episodic memory loss is the most common early manifes­
each group were selected to establish the SVM and ELM tation of dementia and is often overlooked. Later, terrain
models, respectively. The model was self-validated and difficulties gradually emerge and are often accompanied
validated apart in the training set and validation set. The by difficulties with multitasking. As the disease continues
sensitivity, specificity and AUC results are shown in to progress, cognitive impairment becomes more severe
Table 2 and Figure 3. and interferes with activities in daily life. Most dementia
In the fundus original images, the AUCs of the SVM patients are identified in the middle and late period often
model were 0.85, 0.87, and 0.86 for the normal group, by family members. Disease-modifying treatments that
MCI group, and dementia group, while the AUCs of the alter the underlying disease pathology or disease course
ELM model were 0.81, 0.83, and 0.84, suggesting that the are not yet available.35,36 Therefore, dementia seriously
fundus original vascular images achieved good predictive affects the quality of life of patients and their families
efficacy. Meanwhile, the AUCs of the fundus original and increases the social burden. At present, some studies
images between the SVM and ELM models among three have shown that early intervention, such as cognitive
groups were compared, and the results showed that the function training,37 motor function training38 and psycho­
difference was statistically significant (P = 0.0012), indi­ logical and behavioural intervention,39 in patients with
cating that the SVM model had more advantages than the MCI can delay or prevent the progression of cognitive
ELM model. Similar results were found in the fundus impairment and can even promote a reversion to
vascular segmentation images. The AUCs of the SVM a normal cognitive state in some patients. Therefore,

3274 https://doi.org/10.2147/NDT.S333833 Neuropsychiatric Disease and Treatment 2021:17

DovePress

Powered by TCPDF (www.tcpdf.org)

 Dovepress Zhang et al

Table 1 Baseline Characteristics of the Training and Validation Sets

Characteristics of Total Normal Group MCI Group Dementia Group P-value F /χ2 -value

Age (years) 35.11 ± 1.85 55.59 ± 2.16 57.61 ± 3.21 <0.0001* 33.11
Sample size(M/F) (n) 38 (15/23) 26 (15/11) 22 (12/10) 0.2958 2.436
Cerebral infarction (M/F) (n) 0 12 (8/4) 13 (7/6) 0.5133 0.4274
Cerebral hemorrhage (M/F) (n) 0 7 (3/4) 6 (3/3) 0.7968 0.06633
Non-stroke(M/F) (n) 38 (15/23) 7 (4/3) 3 (2/1) 0.4885 1.433
Time of education(years) 14.72 ± 0.76 9.66 ± 0.98 9.58 ± 0.92 <0.0001* 12.60
MMSE (score) 29.43 ± 0.16 27.88 ± 0.40 21.67 ± 0.82 <0.0001* 43.64
MoCA (score) 27.61 ± 0.36 22.00 ± 0.77 14.50 ± 1.30 <0.0001* 79.42
CDR (score) 0.00 ± 0.00 0.50 ± 0.00 1.50 ± 0.15 <0.0001* 151.6
PADL (score) 5.97 ± 0.02 4.41 ± 0.51 1.61 ± 0.41 <0.0001* 48.31
IADL (score) 8.00 ± 0.00 6.50 ± 0.49 2.72 ± 0.55 <0.0001* 59.67
Number of fundus images (n) 149 93 90 - -

Characteristics of Training Set Normal Group MCI Group Dementia Group P-value F /χ2 -value

Age (years) 37.50 ± 2.46 56.60 ± 2.63 55.50 ± 3.21 <0.0001* 17.22
Sample size (M/F) (n) 26 (10/16) 17 (9/8) 15 (8/7) 0.5379 1.240
Cerebral infarction (M/F) (n) 0 8 (5/3) 9 (5/4) 0.7715 0.08433
Cerebral hemorrhage (M/F) (n) 0 5 (2/3) 4 (2/2) 0.7642 0.0900
Non-stroke(M/F) (n) 26 (10/16) 4 (2/2) 2 (1/1) 0.8741 0.2691
Time of education(years) 13.58 ± 0.84 8.63 ± 1.29 9.63 ± 1.12 0.0021* 7.013
MMSE (score) 29.38 ± 0.29 27.33 ± 0.74 24.33 ± 0.92 <0.0001* 17.96
MoCA (score) 27.63 ± 0.52 21.47 ± 1.10 17.00 ± 1.41 <0.0001* 35.09
CDR (score) 0.00 ± 0.00 0.50 ± 0.00 1.17 ± 0.11 <0.0001* 158.7
PADL (score) 5.96 ± 0.04 3.67 ± 0.67 1.75 ± 0.51 <0.0001* 27.59
IADL (score) 8.00 ± 0.00 5.80 ± 0.65 3.33 ± 0.67 <0.0001* 29.15
Number of fundus images (n) 102 66 56 - -

Characteristics of Validation Set Normal Group MCI Group Dementia Group P-value F /χ2 -value

Age (years) 30.33 ± 2.04 53.43 ± 3.93 61.83 ± 7.39 <0.0001* 18.49
Sample size (M/F) (n) 12 (5/7) 9 (6/3) 7 (4/3) 0.5117 1.340
Cerebral infarction (M/F) (n) 0 4 (3/1) 4 (2/2) 0.4652 0.5333
Cerebral hemorrhage (M/F) (n) 0 2 (1/1) 2 (1/1) 1.000 0.0000
Non-stroke(M/F) (n) 12 (5/7) 3 (2/1) 1 (1/0) 0.4346 1.667
Time of education(years) 17.00 ± 1.34 11.86 ± 1.06 9.50 ± 1.77 0.0031* 7.605
MMSE (score) 29.33 ± 0.22 28.29 ± 0.29 16.33 ± 1.41 <0.0001* 115.7
MoCA (score) 27.58 ± 0.34 23.14 ± 0.34 9.50 ± 1.06 <0.0001* 272.7
CDR (score) 0.00 ± 0.00 0.50 ± 0.00 2.17 ± 0.17 <0.0001* 251.2
PADL (score) 6.00 ± 0.00 6.00 ± 0.00 1.33 ± 0.71 <0.0001* 71.24
IADL (score) 8.00 ± 0.00 7.86 ± 0.14 1.50 ± 0.85 <0.0001* 98.57
Number of fundus images (n) 47 27 34 - -
Notes: -: not applicable; *Statistically significant among the Normal group, MCI group, and Dementia group.
Abbreviations: F, female; M, male; MMSE, the Mini-Mental State Examination scale; MoCA, Montreal Cognitive Assessment scale; CDR, Clinical Dementia Rating scale;
PADL, Physical Activities of Daily Living; IADL, Instrumental Activities of Daily Living.

early prediction and diagnosis of dementia is important for language, executive function, complex attention, and
clinical decision-making and the prognosis of patients. social cognition. However, different environmental condi­
There are many clinical instruments available to screen tions, time periods, and qualified evaluators require speci­
for cognitive dysfunction. There are clinical assessments fic tools to screen for cognitive impairment. The MMSE40
that follow the DSM-V criteria for specific tests in six is still the most widely used in screening dementia, and the
cognitive domains, including learning and memory, MoCA41 has better performance than the other dementia

Neuropsychiatric Disease and Treatment 2021:17 https://doi.org/10.2147/NDT.S333833

3275
DovePress

Powered by TCPDF (www.tcpdf.org)

 Zhang et al Dovepress

Table 2 Sensitivity, Specificity and AUC for the SVM and ELM Models of the Fundus Original and Vascular Segmentation Images
Original Images P-value

SVM

Normal Group MCI Group Dementia Group

Self-Validation Validation Self-Validation Validation Self- Validation

Validation

AUC 0.87 0.85 0.88 0.87 0.90 0.86

Sensitivity 0.8488 0.7609 0.7941 0.6901 0.8684 0.9277
Specificity 0.7222 0.7500 0.8871 0.9844 0.8750 0.6964
PS-E1=0.0012*

ELM

Normal Group MCI Group Dementia Group

Self-Validation Validation Self-Validation Validation Self- Validation

Validation

AUC 0.82 0.81 0.86 0.83 0.85 0.84

Sensitivity 0.8488 0.7391 0.8159 0.6100 0.9737 0.7222
Specificity 0.6389 0.8571 0.7661 0.9844 0.5833 0.9107

Segmentation Images

SVM

Normal Group MCI Group Dementia Group

Self-Validation Validation Self-Validation Validation Self- Validation

Validation

AUC 0.86 0.85 0.87 0.81 0.84 0.81

Sensitivity 0.7958 0.8913 0.7647 0.7248 0.8158 0.6111
Specificity 0.7917 0.7500 0.9113 0.8281 0.7750 0.9286
PS-E2=0.0441*

ELM

Normal Group MCI Group Dementia Group

Self-Validation Validation Self-Validation Validation Self- Validation

Validation

AUC 0.75 0.73 0.83 0.80 0.82 0.79

Sensitivity 0.6860 0.5217 0.6745 0.7500 0.6553 0.8933
Specificity 0.6806 0.8929 0.9032 0.7813 0.8750 0.6607
PS-S=0.0388†
PE-E=0.0052†
Notes: The PS-E1 and PS-E2 values were obtained, respectively, from the AUCs of the fundus original and vascular segmentation images between the SVM and ELM models.
The PS-S value was obtained from the AUC of the SVM model between the fundus original and vascular segmentation images, respectively. The PE-E value was obtained from
the AUC of the ELM model between the fundus original and vascular segmentation images, respectively. *Statistically significant between the SVM and ELM model;
†
Statistically significant between the fundus original and vascular segmentation images.
Abbreviations: MCI, Mild Cognitive Impairment; AUC, the area under the curve; SVM, support vector machine; ELM, extreme learning machine.

screening tests for MCI. Additionally, neuropsychological magnetic resonance imaging (MRI), diffusion tensor ima­
assessments,42 which include six main cognitive domains, ging (DTI), arterial spin labelling (ASL), magnetic reso­
may predict preclinical diseases, such as AD. Recently, nance spectroscopy (MRS), functional MRI (fMRI), and
various neuroimaging technologies, including structural positron emission tomography (PET), have emerged and

3276 https://doi.org/10.2147/NDT.S333833 Neuropsychiatric Disease and Treatment 2021:17

DovePress

Powered by TCPDF (www.tcpdf.org)

 Dovepress Zhang et al

Figure 3 The ROC curves and areas under the curve for the SVM and ELM models of the fundus original and segmentation images in the training set and validation set. (A)
The ROC curve and area under the curve for the SVM model of the fundus original images in the training set; (B) the ROC curve and area under the curve for the SVM
model of the fundus original images in the validation set; (C) the ROC curve and area under the curve for the ELM model of the fundus original images in the training set; (D)
the ROC curve and area under the curve for the ELM model of the fundus original images in the validation set; (a) The ROC curve and area under the curve for the SVM
model of the fundus segmentation images in the training set; (b) the ROC curve and area under the curve for the SVM model of the fundus segmentation images in the
validation set; (c) the ROC curve and area under the curve for the ELM model of the fundus segmentation images in the training set; (d) the ROC curve and area under the
curve for the ELM model of the fundus segmentation images in the validation set.

Neuropsychiatric Disease and Treatment 2021:17 https://doi.org/10.2147/NDT.S333833

3277
DovePress

Powered by TCPDF (www.tcpdf.org)

 Zhang et al Dovepress

been applied in clinical practice. There is also growing Now, it has been widely and successfully used in image
evidence supporting the validity of MRI as a biomarker for recognition, especially in pedestrian detection. Because
MCI and dementia. Early identification and diagnosis of quantization of position and direction space by the HOG
MCI or dementia can be achieved by detecting tau protein algorithm can suppress the influence of translation and
aggregation sites, atrophy of the temporal lobe and hippo­ rotation to a certain extent and normalize it to offset the
campus, and changes in blood flow and metabolism.43,44 changes brought by illumination, the HOG algorithm was
However, most clinical assessments of cognitive function applied to extract features in this study. Compared with the
are based on medical knowledge and MRI techniques are segmentation images, original images performed better in
often used in hospitals. With the increasing degree of the prediction of MCI, this result indicated that some
global ageing, it is difficult to effectively carry out large- information in the picture may be lost in the process of
scale screening. image segmentation. Alternatively, we need to expand the
Fundus photography is simple and easy to perform, sample size further and improve image segmentation tech­
cost-effective, and most hospitals have gradually included niques and the performance of the models to verify this
it in physical examination programs or popularized it in result.
social health programs. In this study, a machine learning Furthermore, there are some inconsistencies in the
intelligent diagnostic model was combined with fundus choice of cognitive assessment scales in the studies.
photography technology based on the theoretical basis of MMSE is the most widely applied test for cognitive func­
the homology of retinal blood vessels and cerebral blood tion screening throughout decades. However, it has been
vessels during embryonic development. By extracting the shown inadequate to detect MCI and clinical signs of
features of blood vessels from the fundus images, we dementia because of its ceiling effect.45 Then, MoCA
attempted to establish a prediction model to screen people was developed by Nasreddine et al in 2005 and has been
at high risk of cognitive impairment. The study found that shown to be a highly capable tracking tool for the recogni­
the models based on the vascular features of the fundus tion of normal cognitive function, MCI, and early-onset
images had good recognition and prediction ability, and dementia. The sensitivity of MoCA to identify MCI was
the normal group, the MCI group, and the dementia group 90%, which is superior to MMSE when evaluating patients
showed good calibration and discrimination ability in the with MCI.30 But some studies have demonstrated that
self-verification and verification set as well. Therefore, MoCA is not better than MMSE.46 Consequently, there is
clinicians could refer to the predicted results and the no consensus on which tool is more accurate in detecting
patient’s situation to perform further targeted evaluation cognitive function decline. CDR is a global assessment
or examination as well as targeted cognitive function tool to evaluate both cognition and function. Also, it is
training. a highly effective clinically based interview to distinguish
In the fundus original and vascular segmentation between cognitive and functional decline in the spectrum
images, we found that the SVM model had more advan­ of AD. In general, a CDR score of 0.5 is considered to be
tages than the ELM model. Moreover, compared to fundus the definition of MCI.47 Studies have consistently shown
vascular segmentation images, we found that the original that ADL is impaired in patients with cognitive impair­
images had a better ability to recognize and predict cog­ ment. Therefore, the impact of cognitive impairment on
nitive impairment for the same prediction model. In this ADL, especially IADL, is being used as a criterion to
paper, taking monocular fundus photos centred on optic distinguish between MCI and dementia.48 Combining
disc and the macula allowed us to extract vascular features with diagnostic criteria for MCI based on ICD-11, we
from different angles and in a larger range, avoiding the used MMSE, MOCA, CDR, PADL, and IADL to compre­
loss of vascular information. The AHE algorithm and hensively evaluate the cognitive and functional status of
median filtering method used in this study for segmenta­ subjects, and then classified into a normal group, an MCI
tion were more suitable for improving the local contrast of group, and a dementia group.
the image to obtain more image details and for removing There are several limitations to our study that may
noise to protect the edge of the signal from being blurred. represent a direction for further research. One important
Additionally, the method of pedestrian detection based on limitation is the lack of external validation for the model.
HOG feature combined with SVM classifiers was first To obtain a high level of clinical application evidence,
proposed by French researcher Dalal in CVPR in 2005. multicentre validation with a larger sample size is needed.

3278 https://doi.org/10.2147/NDT.S333833 Neuropsychiatric Disease and Treatment 2021:17

DovePress

Powered by TCPDF (www.tcpdf.org)

 Dovepress Zhang et al

In addition, the studies showed that blood sugar,49 blood Ethics

pressure,50 and language51 are associated with cognitive The study was approved by the ethic committee at the
function. Whether those factors had good recognition and Seventh Affiliated Hospital of Sun Yat-sen University,
predictive ability for MCI, or whether they could improve where all data collection occurred, complies with the prin­
the predictive performance of the model in combination ciples of the Declaration of Helsinki. The study has been
with fundus vessel features, requires further exploration. registered in the China Clinical Trial Registry
Moreover, different diseases and different subtypes of the (ChiCTR1900027404).
stroke may lead to different degrees and domains of cog­
nitive impairment. In order to improve the early recogni­
Acknowledgments
tion and prediction of cognitive impairment in different
This study was supported by the Sun Yat-Sen University
diseases and different types of the same disease, further
Clinical Research 5010 (Grant Number: 2014001).
research on a large sample size is needed.

Conclusion Author Contributions

All authors contributed to data analysis, drafting or revis­
In the study, the establishment of a predictive model based on
ing the article, have agreed on the journal to which the
vascular-related feature extraction from fundus images by
article will be submitted, gave final approval of the version
machine learning has high recognition and prediction abil­
to be published, and agree to be accountable for all aspects
ities for cognitive function and can be used as a screening
of the work.
method for MCI. In conclusion, we believe this study gives
a preliminary conclusion and provides a basis for further
research in this field. Compared with the cognitive function Disclosure
assessment and imaging examination commonly used in The authors declare that they have no conflicts of interest.
clinical screening and diagnosis of dementia, the predictive
model has greater advantages in operation and implementa­ References
tion, and is more suitable as a screening method for MCI or 1. Alzheimer’s Disease International [homepage on the Internet]; 2021.
dementia in the community or physical examination centres. Available from: https://www.alzint.org/about/dementia-facts-figures
/dementia-statistics. Accessed October 14, 2021.
In future research, we will conduct a larger sample size study 2. Nemeroff CB, Weinberger D, Rutter M, et al. DSM-5: a collection of
in multiple centres and try a variety of image segmentation psychiatrist views on the changes, controversies, and future
and feature extraction methods to improve the recognition directions. BMC Med. 2013;11:202. doi:10.1186/1741-7015-11-202
3. Harvey PD. Domains of cognition and their assessment. Dialogues
ability. In the process of further research, the performance of Clin Neurosci. 2019;21(3):227–237.
the prediction model will decline due to different factors, 4. Morris JC. Mild cognitive impairment is early-stage Alzheimer dis­
ease: time to revise diagnostic criteria. Arch Neurol. 2006;63:15–16.
such as equipment and region. How to train the model with doi:10.1001/archneur.63.1.15
more stable and better performance through the transfer 5. Fiorini R, Luzzi S, Vignini A. Perspectives on mild cognitive impair­
learning method is our next research direction. ment as a precursor of Alzheimer’s disease. Neural Regen Res.
2020;15(11):2039–2040. doi:10.4103/1673-5374.282256
6. Sanford AM. Mild cognitive impairment. Clin Geriatr Med. 2017;33
Data Sharing Statement (3):325–337. doi:10.1016/j.cger.2017.02.005
7. Sachdev PS, Blacker D, Blazer DG, et al. Classifying neurocognitive
We are willing to share all individual deidentified partici­ disorders: the DSM-5 approach. Nat Rev Neurol. 2014;10
pant data related to the article, including the data of (11):634–642. doi:10.1038/nrneurol.2014.181
models, the datasets of the fundus images, the Case 8. Deo RC. Machine learning in medicine. Circulation. 2015;132
(20):1920–1930. doi:10.1161/CIRCULATIONAHA.115.001593
Report Form (CRF), and so on. The data of the SVM 9. Kalmet PHS, Sanduleanu S, Primakov S, et al. Deep learning in
and ELM models are within the manuscript. The datasets fracture detection: a narrative review. Acta Orthop. 2020;91
(2):215–220. doi:10.1080/17453674.2019.1711323
of the fundus images will be uploaded to the public repo­ 10. LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;521
sitory after the clinical trial is completed within six months (7553):436–444. doi:10.1038/nature14539
11. Anwar SM, Majid M, Qayyum A, Awais M, Alnowami M,
for further analysis for other researchers. The data are
Khan MK. Medical image analysis using convolutional neural net­
available on request from the corresponding author works: a review. J Med Syst. 2018;42(11):226. doi:10.1007/s10916-
(Dongfeng Huang, E-mail: [email protected]) 018-1088-1
12. Pellegrini M, Vagge A, Ferro Desideri LF, et al. Optical coherence
and the first author (Qian Zhang, E-mail: zhangq226@­ tomography angiography in neurodegenerative disorders. J Clin Med.
mail.sysu.edu.cn). 2020;9(6):1706. doi:10.3390/jcm9061706

Neuropsychiatric Disease and Treatment 2021:17 https://doi.org/10.2147/NDT.S333833

3279
DovePress

Powered by TCPDF (www.tcpdf.org)

 Zhang et al Dovepress

13. Risau W. Mechanisms of angiogenesis. Nature. 1997;386 31. Breton A, Casey D, Arnaoutoglou NA. Cognitive tests for the detec­
(6626):671–674. doi:10.1038/386671a0 tion of mild cognitive impairment (MCI), the prodromal stage of
14. Dorrell MI, Aguilar E, Friedlander M. Retinal vascular development dementia: meta-analysis of diagnostic accuracy studies.
is mediated by endothelial filopodia, a preexisting astrocytic template Int J Geriatr Psychiatry. 2019;34(2):233–242. doi:10.1002/gps.5016
and specific R-cadherin adhesion. Invest Ophthalmol Vis Sci. 2002;43 32. Lakhani P, Gray DL, Pett CR, Nagy P, Shih G. Hello world deep
(11):3500–3510. learning in medical imaging. J Digit Imaging. 2018;31(3):283–289.
15. Heringa SM, Bouvy WH, van den Berg E, Moll AC, Kappelle LJ, doi:10.1007/s10278-018-0079-6
Biessels GJ. Associations between retinal microvascular changes and 33. Noble WS. What is a support vector machine? Nat Biotechnol.
dementia, cognitive functioning, and brain imaging abnormalities: 2006;24(12):1565–1567. doi:10.1038/nbt1206-1565
a systematic review. J Cereb Blood Flow Metab. 2013;33 34. Jammoussi I, Ben Nasr M, Hybrid Method A. Based on extreme
(7):983–995. doi:10.1038/jcbfm.2013.58 learning machine and self organizing map for pattern classification.
16. Zhang Y, Wang Y, Shi C, Shen M, Lu F. Advances in retina imaging as Comput Intell Neurosci. 2020;2020:2918276. doi:10.1155/2020/
potential biomarkers for early diagnosis of Alzheimer’s disease. Transl 2918276
Neurodegener. 2021;10(1):6. doi:10.1186/s40035-021-00230-9 35. Mangialasche F, Solomon A, Winblad B, Mecocci P,
17. Shi Z, Zheng H, Hu J, et al. Retinal nerve fiber layer thinning is Kivipelto M. Alzheimer’s disease: clinical trials and drug devel­
associated with brain atrophy: a longitudinal study in nondemented opment. Lancet Neurol. 2010;9(7):702–716. doi:10.1016/S1474-
older adults. Front Aging Neurosci. 2019;11:69. doi:10.3389/ 4422(10)70119-8
fnagi.2019.00069 36. Costandi M. Ways to stop the spread of Alzheimer’s disease. Nature.
18. Kwon JY, Yang JH, Han JS, Kim DG. Analysis of the retinal nerve 2018;559(7715):S16–S17. doi:10.1038/d41586-018-05723-8
fiber layer thickness in Alzheimer disease and mild cognitive 37. Gates NJ, Rutjes AW, Di Nisio M, et al. Computerised cognitive
impairment. Korean J Ophthalmol. 2017;31(6):548–556. training for maintaining cognitive function in cognitively healthy
doi:10.3341/kjo.2016.0118 people in midlife. Cochrane Database Syst Rev. 2019;3(3):
19. Querques G, Borrelli E, Sacconi R, et al. Functional and morpholo­ CD012278.
gical changes of the retinal vessels in Alzheimer’s disease and mild 38. Carrasco-Poyatos M, Rubio-Arias JA, Ballesta-García I, Ramos-
cognitive impairment. Sci Rep. 2019;9(1):63. doi:10.1038/s41598- Campo DJ. Pilates vs. muscular training in older women. Effects in
018-37271-6 functional factors and the cognitive interaction: a randomized con­
20. Chan VTT, Sun Z, Tang S, et al. Spectral-domain OCT measurements trolled trial. Physiol Behav. 2019;201:157–164. doi:10.1016/j.
in Alzheimer’s disease: a systematic review and meta-analysis. physbeh.2018.12.008
Ophthalmology. 2019;126(4):497–510. doi:10.1016/j. 39. Robert P, Manera V, Derreumaux A, et al. Efficacy of a web app for
ophtha.2018.08.009 cognitive training (MeMo) regarding cognitive and behavioral per­
21. Cheung CY, Chan VTT, Mok VC, Chen C, Wong TY. Potential formance in people with neurocognitive disorders: randomized con­
retinal biomarkers for dementia: what is new? Curr Opin Neurol.
trolled trial. J Med Internet Res. 2020;22(3):e17167. doi:10.2196/
2019;32(1):82–91. doi:10.1097/WCO.0000000000000645
17167
22. Golzan SM, Goozee K, Georgevsky D, et al. Retinal vascular and
40. Li H, Jia J, Yang Z, Moreau N. Mini-mental state examination in
structural changes are associated with amyloid burden in the elderly:
elderly Chinese: a population-based normative study. J Alzheimers
ophthalmic biomarkers of preclinical Alzheimer’s disease.
Dis. 2016;53(2):487–496. doi:10.3233/JAD-160119
Alzheimers Res Ther. 2017;9(1):13. doi:10.1186/s13195-017-0239-9
41. Liew TM, Feng L, Gao Q, Ng TP, Yap P. Diagnostic utility of
23. Cheung CY, Ong YT, Ikram MK, et al. Microvascular network altera­
Montreal Cognitive Assessment in the Fifth Edition of Diagnostic
tions in the retina of patients with Alzheimer’s disease. Alzheimers
and Statistical Manual of Mental Disorders: major and mild neuro­
Dement. 2014;10(2):135–142. doi:10.1016/j.jalz.2013.06.009
cognitive disorders. J Am Med Dir Assoc. 2015;16(2):144–148.
24. Nadal J, Deverdun J, de Champfleur NM, et al. Retinal vascular
fractal dimension and cerebral blood flow, a pilot study. Acta doi:10.1016/j.jamda.2014.07.021
Ophthalmol. 2020;98(1):e63–e71. doi:10.1111/aos.14232 42. Tosi G, Borsani C, Castiglioni S, Daini R, Franceschi M, Romano D.
25. Srinidhi L, Aparna P, Rajan J. Recent advancements in retinal vessel Complexity in neuropsychological assessments of cognitive impair­
segmentation. J Med Syst. 2017;41(4):70. doi:10.1007/s10916-017- ment: a network analysis approach. Cortex. 2020;124:85–96.
0719-2 doi:10.1016/j.cortex.2019.11.004
26. Rasta SH, Partovi ME, Seyedarabi H, Javadzadeh A. A comparative 43. Chandra A, Dervenoulas G, Politis M. Alzheimer’s disease neuroi­
study on preprocessing techniques in diabetic retinopathy retinal maging initiative. magnetic resonance imaging in Alzheimer’s dis­
images: illumination correction and contrast enhancement. J Med ease and mild cognitive impairment. J Neurol. 2019;266
Signals Sens. 2015;5(1):40–48. doi:10.4103/2228-7477.150414 (6):1293–1302. doi:10.1007/s00415-018-9016-3
27. World Health Organization. International statistical classification of 44. Mamun AA, Uddin MS, Mathew B, Ashraf GM. Toxic tau: structural
diseases and related health problems (11th Revision); 2018. Available origins of tau aggregation in Alzheimer’s disease. Neural Regen Res.
from: https://www.who.int/classifications/icd/en. Accessed November 2020;15(8):1417–1420. doi:10.4103/1673-5374.274329
15, 2018. 45. Tsoi KK, Chan JY, Hirai HW, Wong SY, Kwok TC. Cognitive tests to
28. Kasper S, Bancher C, Eckert A, et al. Management of mild cognitive detect dementia: a systematic review and meta-analysis. JAMA Intern
impairment (MCI): the need for national and international guidelines. Med. 2015;175(9):1450–1458. doi:10.1001/
World J Biol Psychiatry. 2020;21(8):579–594. doi:10.1080/ jamainternmed.2015.2152
15622975.2019.1696473 46. Zhou S, Zhu J, Zhang N, et al. The influence of education on Chinese
29. Folstein MF, Folstein SE, McHugh PR. “Mini-mental state”. version of Montreal cognitive assessment in detecting amnesic mild
A practical method for grading the cognitive state of patients for cognitive impairment among older people in a Beijing rural
the clinician. J Psychiatr Res. 1975;12(3):189–198. doi:10.1016/ community. Sci World J. 2014;2014:689456. doi:10.1155/2014/
0022-3956(75)90026-6 689456
30. Nasreddine ZS, Phillips NA, Bédirian V, et al. The Montreal 47. Saxton J, Snitz BE, Lopez OL, et al. Functional and cognitive criteria
Cognitive Assessment, MoCA: a brief screening tool for mild cogni­ produce different rates of mild cognitive impairment and conversion
tive impairment. J Am Geriatr Soc. 2005;53(4):695–699. to dementia. J Neurol Neurosurg Psychiatry. 2009;80(7):737–743.
doi:10.1111/j.1532-5415.2005.53221.x doi:10.1136/jnnp.2008.160705

3280 https://doi.org/10.2147/NDT.S333833 Neuropsychiatric Disease and Treatment 2021:17

DovePress

Powered by TCPDF (www.tcpdf.org)

 Dovepress Zhang et al

48. Perneczky R, Pohl C, Sorg C, et al. Complex activities of daily living 50. Hughes D, Judge C, Murphy R, et al. Association of blood pressure
in mild cognitive impairment: conceptual and diagnostic issues. Age lowering with incident dementia or cognitive impairment:
Ageing. 2006;35(3):240–245. doi:10.1093/ageing/afj054 a systematic review and meta-analysis. JAMA. 2020;323
49. Fickweiler W, Wolfson EA, Paniagua SM, et al. Association of (19):1934–1944. doi:10.1001/jama.2020.4249
cognitive function and retinal neural and vascular structure in type 51. Bidet-Ildei C, Beauprez SA, Boucard G. The link between language
1 diabetes. J Clin Endocrinol Metab. 2021;106(4):1139–1149. and action in aging. Arch Gerontol Geriatr. 2020;90:104099.
doi:10.1210/clinem/dgaa921 doi:10.1016/j.archger.2020.104099

Neuropsychiatric Disease and Treatment Dovepress

Publish your work in this journal
Neuropsychiatric Disease and Treatment is an international, peer- is the official journal of The International Neuropsychiatric
reviewed journal of clinical therapeutics and pharmacology focusing Association (INA). The manuscript management system is comple­
on concise rapid reporting of clinical or pre-clinical studies on a tely online and includes a very quick and fair peer-review system,
range of neuropsychiatric and neurological disorders. This journal is which is all easy to use. Visit http://www.dovepress.com/testimo­
indexed on PubMed Central, the ‘PsycINFO’ database and CAS, and nials.php to read real quotes from published authors.
Submit your manuscript here: https://www.dovepress.com/neuropsychiatric-disease-and-treatment-journal

Neuropsychiatric Disease and Treatment 2021:17 DovePress 3281

Powered by TCPDF (www.tcpdf.org)

 RESEARCH ARTICLE

Retinal texture biomarkers may help to

discriminate between Alzheimer’s,
Parkinson’s, and healthy controls
Ana Nunes ID1,2,3, Gilberto Silva1,2,3, Cristina Duque ID4, Cristina Januário3,4,5,
Isabel Santana3,5,6,7, António Francisco Ambrósio3,6,8, Miguel Castelo-Branco1,2,3,
Rui Bernardes1,2,3*
1 Coimbra Institute for Biomedical Imaging and Translational Research (CIBIT), Health Sciences Campus,
Polo III, Azinhaga de Santa Comba, Coimbra, Portugal, 2 Institute of Nuclear Sciences Applied to Health
a1111111111 (ICNAS), Health Sciences Campus, Polo III, Azinhaga de Santa Comba, Coimbra, Portugal, 3 Faculty of
a1111111111 Medicine, University of Coimbra, Health Sciences Campus, Polo III, Azinhaga de Santa Comba, Coimbra,
a1111111111 Portugal, 4 Movement Disorders Clinic, Department of Neurology, Centro Hospitalar e Universitário de
a1111111111 Coimbra (CHUC), Praceta Prof. Mota Pinto, Coimbra, Portugal, 5 CNC.IBILI Consortium, Health Sciences
a1111111111 Campus, Polo III, Azinhaga de Santa Comba, Coimbra, Portugal, 6 Dementia Clinic, Department of
Neurology, Centro Hospitalar e Universitário de Coimbra (CHUC), Praceta Prof. Mota Pinto, Coimbra,
Portugal, 7 Center for Neuroscience and Cell Biology (CNC), Zoology Department, University of Coimbra,
Coimbra, Portugal, 8 Coimbra Institute for Clinical and Biomedical Research (iCBR), Health Sciences
Campus, Polo III, Azinhaga de Santa Comba, Coimbra, Portugal

OPEN ACCESS * [email protected]

Citation: Nunes A, Silva G, Duque C, Januário C,

Santana I, Ambrósio AF, et al. (2019) Retinal
texture biomarkers may help to discriminate Abstract
between Alzheimer’s, Parkinson’s, and healthy
controls. PLoS ONE 14(6): e0218826. https://doi. A top priority in biomarker development for Alzheimer’s disease (AD) and Parkinson’s dis-
org/10.1371/journal.pone.0218826
ease (PD) is the focus on early diagnosis, where the use of the retina is a promising avenue
Editor: Demetrios G. Vavvas, Massachusetts Eye & of research. We computed fundus images from optical coherence tomography (OCT) data
Ear Infirmary, Harvard Medical School, UNITED
and analysed the structural arrangement of the retinal tissue using texture metrics. We built
STATES
clinical class classification models to distinguish between healthy controls (HC), AD, and
Received: April 3, 2019
PD, using machine learning (support vector machines). Median sensitivity is 88.7%, 79.5%
Accepted: June 11, 2019 and 77.8%, for HC, AD, and PD eyes, respectively. When the same subject has the same
Published: June 21, 2019 classification for both eyes, 94.4% (median) of the classifications are correct. A significant
Copyright: © 2019 Nunes et al. This is an open amount of information discriminating between multiple neurodegenerative states is con-
access article distributed under the terms of the veyed by OCT imaging of the human retina, even when differences in thickness are not yet
Creative Commons Attribution License, which present. This technique may allow for simultaneously diagnosing Alzheimer’s and Parkin-
permits unrestricted use, distribution, and
son’s diseases.
reproduction in any medium, provided the original
author and source are credited.

Data Availability Statement: All relevant data are

within the manuscript and its Supporting
Information files. Introduction
Funding: Authors AFA, MCB, RB were supported Diagnosis of both Alzheimer’s disease (AD) and Parkinson’s disease (PD) remains a major
by: The changing brain in Alzheimer’s disease: is challenge. Simple diagnostic tests serving as biomarkers for tracking the onset and progression
the retina a reliable mirror of disease onset and
of both diseases would be extremely valuable.
progression? Mantero Belard Award - Santa Casa
Awards – Neurosciences (3rd edition), MB-1049-
AD affects 5 to 7% of people over the age of sixty [1], and a total of 5.4 million patients
2015 (Funder name: Santa Casa da Misericórdia de in the United States [2]. For PD, the incidence has been reported to be 4.5–19 per 100 000
Lisboa Funder URL: http://www.scml.pt/). Authors population, per year [3]; while its age-adjusted prevalence is estimated to range between

PLOS ONE | https://doi.org/10.1371/journal.pone.0218826 June 21, 2019 1 / 13

 Retinal biomarkers of neurodegenerative disorders

AN, CJ, AFA, MCB, RB were supported by: 72 and 259 per 100 000. In the US, the direct costs of health and social care associated with
BIGDATAIMAGE - From computational modelling dementia patients have been estimated to reach USD 100 billion per year [3]. Similarly, for
and clinical research to the development of
PD patients the total annual cost is more than the double of that of the control population
neuroimaging big data platforms for discovery of
novel biomarker, CENTRO-01-0145-FEDER- [3].
000016 (Funder name: Comissão de Coordenação Even though epidemiological evidence suggests a potential reduction in dementia incidence
e Desenvolvimento Regional do Centro Funder in recent years, its prevalence, according to Shah and co-workers [4], is expected to increase
URL: http://www.ccdrc.pt/). Authors AN, CJ, IS, in the coming decades due to population ageing. The impact of dementia in patients and the
AFA, MCB, RB were supported by: PEst-UID/NEU/
overall society was acknowledged in the First WHO Ministerial Conference on Global Action
04539/2013, FEDER-COMPETE (POCI-01-0145-
FEDER-007440 and POCI01-0145-FEDER-016428)
Against Dementia (March 2015), bringing together, among others, representatives from 80
(Funder name: Fundação para a Ciência e WHO Member States and four UN agencies [4]. On the top priorities in the domain of diagno-
Tecnologia Funder URL: https://www.fct.pt/). sis, one can find the research aiming to promote timely and accurate diagnosis of dementia
Authors AFA, MCB were supported by: CENTRO- through the development and validation of simple riskless biomarkers that could even be used
01-0145-FEDER-000008: BrainHealth 2020 in prodromal disease stages. Moreover, the focus on the need to “. . . identify similarities and
(Funder name: Centro 2020 Funder URL: http://
differences between diseases and dementia subtypes, and assess progression from premanifest
www.centro.portugal2020.pt/).
(presymptomatic) to late-stage diseases” [4] is stressed in the report. Furthermore, priorities
Competing interests: The authors have declared point to the need to diagnose dementia in the primary care practices, thus putting aside the
that no competing interests exist.
focus on the diagnosis based on brain imaging requiring instrumentation available only at
advanced imaging centres, where the access is only possible to a fraction of the population—
not to mention the associated costs.
The morphological effects of the neurodegenerative process seem to be similar in the ret-
ina and the brain [5]. Therefore, the use of the retina as a window into the brain [6–8] might
provide a simple solution for the establishment of reliable image-based biomarkers for neuro-
degeneration. This concept has been extensively exploited, mostly regarding the onset and
progression of AD [9, 10] and PD [11, 12]. Notably, there is published evidence that the reti-
nal structure might be altered in the course of these two neurodegenerative disorders [13–
16].
When compared to brain imaging, the easy access to and lower operational cost of the oph-
thalmological imaging techniques are definite advantages, bringing the diagnosis closer to the
primary care practices. Even though optical coherence tomography (OCT) has been exten-
sively used in assessing neurodegenerative disorders, the hallmark has been the use of thick-
ness measurements, either at the individual retinal layer level or in an aggregated fashion.
While the thinning of specific layers has been the standard finding for both disorders, some
contradictory findings can be found in the literature [9, 17–21]. Particularly in the case of AD,
it has recently been reported that the thickness of some of the inner retinal layers undergoes
dynamic changes as the disease progresses [22]. Moreover, with the retinal nerve fibre layer
(RNFL) and retinal ganglion cell axons being the top candidates for the occurrence of neuro-
degeneration, it should be kept in mind that the thinning of this layer may occur for other
types of dementia beyond AD or PD, such as Lewy body dementia [8].
Recent research [23–28] has pointed towards texture analysis as a promising tool in the
study of biomarkers for neurodegenerative diseases. The term “texture analysis” encompasses
a wide range of methods that allow for the characterisation of the underlying image patterns
[29–31]. Such methods, when applied to images of the human retina, can help identify changes
in the structural arrangement of the retina or that of specific retinal layers, both in health and
disease.
In the present study, we address these issues by imaging the ocular fundus by OCT of
healthy controls (HC), as well as that of patients diagnosed with AD and PD. Using a single
age-matched control group and two groups of patients, we were able to develop a classification
system based on texture characteristics of fundus images computed from OCT data. This sys-
tem achieved a significant triple clinical classification accuracy.

PLOS ONE | https://doi.org/10.1371/journal.pone.0218826 June 21, 2019 2 / 13

 Retinal biomarkers of neurodegenerative disorders

Materials and methods

Participants
OCT data from 20 patients diagnosed with AD, 28 patients diagnosed with PD and 27 age-
matched HC were gathered from the authors’ institutional OCT database. The two studies
(AD and PD) from which data were collected were approved by the Ethics Committee of the
Faculty of Medicine of the University of Coimbra and were conducted according to the princi-
ples stated in the Declaration of Helsinki [32]. Written informed consents were obtained from
all participants. Before the inclusion in this study, the AD patients underwent a thorough
neuropsychological evaluation process, administered by experienced neurologists. As all these
patients scored a value of 1 in the clinical dementia rating (CDR) scale, they were deemed
capable of signing the written informed consent themselves.
The AD and PD groups were recruited respectively at the Dementia and the Movement
Disorders Units of the Neurological Department of the Centro Hospitalar e Universitário de
Coimbra, where they were assessed by experienced neurologists. All participants had been pre-
viously imaged using the same OCT device and acquisition protocol on both eyes. Due to the
low quality of the signal, four eyes were rejected: two from the AD group and two from the PD
group.
For the diagnosis of AD, the standard criteria [33] was used, and patients were extensively
characterised using formal neuropsychological evaluation as well as specific cognitive and
functional staging scales, including the Montreal cognitive assessment (MoCA) [34, 35]—
please refer to S1 Table for the MoCA scores. Patients were further investigated with obligatory
laboratory, imaging studies and cerebrospinal fluid (CSF) analysis to exclude other forms of
reversible dementia or systemic diseases. According to the main aims of the study, patients
were strictly selected with a probable diagnosis of AD supported by positive CSF biomarkers
(amyloid-β42, total tau and phosphorylated tau) and positive Pittsburgh Compound-B (PiB)
and positron emission tomography (PET).
Further details on the methods used for the collection and analysis of CSF and PiB-PET
data can be found in [36].
All patients had recently converted to AD from a prior stage of mild cognitive impairment.
They were all in mild stage (CDR score = 1), with diagnosis duration ranging from zero to two
years (see S1 Table for more detail) and were in a stable condition. Retinal injuries and reti-
nopathies, optic neuropathies secondary to other factors (e.g. glaucoma, diabetes, age-related
macular degeneration) and severe visual impairment were exclusion criteria.
The PD patients were assessed using the MoCA, the unified Parkinson’s disease rating scale
(UPDRS), and Hoehn and Yahr staging (test scores are available in S2 Table). These patients
were diagnosed by a movement disorder neurologist, according to the criteria defined by the
UK Parkinsons’s Disease Society Brain Bank [37]. Only patients with ages ranging between
40 and 85 years old were included in the study. Subjects showing signs of advanced dementia,
severe depression and history of substance abuse were excluded. Furthermore, a neuro-oph-
thalmologist assessed the patients and those with intrinsic optic nerve or macula pathology
were excluded.
All subjects in the present study were selected towards the best age-match between groups
and a balanced distribution by gender within each group. At the time of OCT scan, the PD
patients had a diagnosis duration that ranged from one to 19 years (Q1/median/Q3: 2.75/5.00/
10.00)—please refer to S2 Table for more detail. These patients’ disease duration is modestly
correlated (R2 = 0.035) to their age (S1 Fig). Demographic data for all subjects in the current
study are presented in Table 1.

PLOS ONE | https://doi.org/10.1371/journal.pone.0218826 June 21, 2019 3 / 13

 Retinal biomarkers of neurodegenerative disorders

Table 1. Demographic data of the control and patient groups.

Healthy Controls Alzheimer’s Disease Parkinson’s Disease
N 27 20 28
Age—mean(std) (years) 64.1(7.1) 66.3(6.8) 63.4(6.6)
Age—min(max) (years) 53(75) 54(76) 53(77)
Male(Female) 13(14) 10(10) 13(15)
Right(Left) Eyes 26(27) 20(19) 27(27)
Total Acquisitions 53 39 54
https://doi.org/10.1371/journal.pone.0218826.t001

Imaging protocol
Retinal imaging was performed using the Cirrus SD-OCT 5000 (Carl Zeiss, Meditec, Dublin,
CA, USA) and the 512x128 Macular Cube protocol, centred on the macula.

Data pre-processing
The gathered volumetric data were further processed using the OCT Explorer software (Reti-
nal Image Analysis Lab, Iowa Institute for Biomedical Imaging, Iowa City, IA, USA) to seg-
ment eleven retinal layers: the RNFL, ganglion cell layer (GCL), inner plexiform layer (IPL),
inner nuclear layer (INL), outer plexiform layer (OPL), outer nuclear layer (ONL), inner seg-
ment/outer segment junction, outer segment, outer photoreceptor, subretinal virtual space,
and retinal pigment epithelium (RPE). These segmentations were visually examined and man-
ually corrected where necessary.
From each of the six inner retinal layers, the RNFL, the GCL, the IPL, the INL, the OPL,
and the ONL, a mean value fundus (MVF) image [38] was computed as the average of the
A-scan values between the two retinal layer interfaces defining the layer at study. These six
images per eye constitute the data to be further processed and analysed throughout this work.
All left eyes were horizontally flipped to match the right ones and to allow metrics to keep the
same relative position. Fig 1 shows an example for reference purposes only, where layers were
intensity corrected and pseudo-colour coded for ease of visualisation.

Retinal thickness
The macular thickness was computed for each of the six retinal layers at study in this work,
along with the full retinal thickness, i.e. the thickness from the inner limiting membrane to the
top of the RPE.

Texture analysis
There are several methods for retrieving textural features from an image. Frequently, statistical
methods such as the gray level co-occurrence matrix (GLCM) [39, 40] are employed. These
methods provide insight into the patterns and relative distribution of the image’s intensity val-
ues. In the particular case of the GLCM, localised grey-level variations are recognised through-
out different directions, and scales, examining pixel pairs iteratively. Alternatively, spectral
texture features can be computed. By evaluating spatial frequencies at multiple scales, the
wavelet transform has a natural application in texture analysis, where it can help to identify
general/coarse texture features that statistical approaches like the GLCM fail to capture [29].
For this reason, wavelet-based texture descriptors have been computed and used in the classifi-
cation of different types of medical images [41–45]. Both GLCM and wavelet-based texture
features were used in this work to capture, respectively, local and global texture information.

PLOS ONE | https://doi.org/10.1371/journal.pone.0218826 June 21, 2019 4 / 13

 Retinal biomarkers of neurodegenerative disorders

Fig 1. Mean value fundus images. Colour-coded MVF images from the right eye of a healthy control. From left to right and top to
bottom: RNFL, GCL, IPL, INL, OPL and ONL layer fundus images.
https://doi.org/10.1371/journal.pone.0218826.g001

Regarding the GLCM, each image was first down-sampled to 128x128 pixels to obtain iso-
tropic sampling in the horizontal and vertical directions and converted to 16 grey-level to limit
the size of the GLCM matrix. These images were then split into 7x7 blocks to be independently
analysed. The supremum over the different directions of each of the 20 metrics computed
from the GLCM, for each block, was chosen as the respective value for the block. This informa-
tion was later aggregated per quadrant, the superior-temporal, the superior-nasal, the inferior-
temporal and the inferior-nasal quadrants. Each quadrant metric is the average of that of the
respective 3x3 blocks, leaving out the row and column passing through the fovea (Fig 2). As
such, a total of 80 GLCM-based features characterise each layer of the retina.
Concerning spectral texture features, the dual-tree complex wavelet transform (DTCWT)
[46] was applied to each of the fundus images. The variance and entropy of the magnitude
of the DTCWT complex coefficients were computed for all image subbands, following the
approach used previously [47, 48]. Exploratory analysis revealed that the variance alone, when
computed for the six directionally selective subbands (±15˚, ±45˚ and ±75˚) at decomposition
level one, leads to the best performance. These six directional variance values were therefore
combined with the 80 GLCM-based features to form the final feature vector, composed of a
total of 86 features per layer.

Classification
The selected method for data classification was the support vector machine (SVM). SVM-
based classification [49, 50] is a two-step procedure where, in the first step, SVMs learn the dif-
ferences between two groups to establish a classification model (training phase). In the second

PLOS ONE | https://doi.org/10.1371/journal.pone.0218826 June 21, 2019 5 / 13

 Retinal biomarkers of neurodegenerative disorders

Fig 2. Mean value fundus images’ blocks. Computed fundus image from the volumetric macular cube scan of the right eye of a
healthy control subject. Each of the 7x7 blocks show the individually analysed areas which results were later aggregated into larger
regions (shaded areas). Image axes are: x-axis (horizontal)—temporal (left) to nasal (right) and y-axis (vertical) superior (top) to
inferior (bottom).
https://doi.org/10.1371/journal.pone.0218826.g002

step, this model is used to classify cases not used during the first step. The validity of the model
can be assessed (test phase) by using known cases in the second-step and comparing the pre-
dicted classifications to the real ones.
Prior to the classification, all features were independently z-score transformed to eliminate
differences in scale that may interfere with the system’s performance.
In total, 18 independent binary SVM models (with the radial basis function kernel) were
developed, one per each of the six retinal layers at study and between any two of the three pos-
sible groups at study: the HC, AD, and PD groups.
Six features were considered for each of the 18 models to achieve the best classification per-
formance while avoiding overtraining, considering the number of cases (eyes) in each of the
groups (classes). The identification of the particular set of features for each of the models was
performed by resorting to a recursive forward selection procedure. In this approach, a single
feature is tested at a time, and the one allowing for the best performance is selected. After that,
at a time, each one of the remaining features is considered, and the performance based on the

PLOS ONE | https://doi.org/10.1371/journal.pone.0218826 June 21, 2019 6 / 13

 Retinal biomarkers of neurodegenerative disorders

two features (the feature currently being considered, plus the previously selected “best” one) is
assessed. The set presenting the best performance is kept, and the process repeats until the best
set with the predefined number of features is found.
The performance of each of the classification processes was assessed based on the k-fold
cross-validation. In this procedure, the dataset at hand is split into k sets. The system is then
trained with k-1 sets and tested on the set left out from the training. The testing consists of
predicting the classes where the cases left out from the training would belong to, based on the
model developed during the training phase, and comparing the predictions with the known
(real) classification. This procedure is repeated k times.
For the classification of an eye in one of the three possible classes, a one-versus-one
approach was used, along with a voting scheme. As such, for each of the six retinal layers at
study, each eye was tested against the model between the control and AD classes, between the
control and PD classes and between the AD and PD classes. For each of these three tests, the
eye received a vote in one of the two classes under evaluation.
At the end of this procedure, the eye was classified into the most voted class. If all three clas-
ses had received the same number of votes, a tie was considered, and the eye was classified as
“Unknown”. If two classes received the same number of votes, the final decision was made by
considering the result of the classification between those two classes. A diagram illustrating the
classification process is shown in S2 Fig.

Statistical analysis
Based on the classifications of all eyes, using the classification process described above, a 3x4
confusion matrix was built, where the three rows represent the real (known) classes, and the
four columns represent the classes according to the data and classification models. The extra
“Unknown” class is composed of the cases with ties on the classification.
From the confusion matrix, the sensitivity and specificity were computed for each class: the
control, the AD, and PD classes. The system accuracy was also computed.
Additionally, the percentage of participants with both eyes receiving the same classification
was determined, along with the percentage of these that were correctly classified.
The last computed parameter was the percentage of eyes in the “Unknown” class, that is,
the percentage of eyes where, from the texture point of view, no substantial differences were
found between the HC, AD, and PD study populations.
All results were computed using Matlab R2017b (The MathWorks Inc., Natick, MA, USA)
running on a personal computer.

Results
Table 2 presents the results of the full retinal thickness measurement distributions for the three
groups. Thickness values are the average thickness within nine regions centred on the fovea.
The central region is the circular area of 1000 micrometres in diameter, inner macular areas
are within the radii 1000 and 1500 micrometres, and outer macular areas are within the radii
1500 and 3000 micrometres.
For each of the nine areas specified in Table 2, a conventional one-way ANOVA was per-
formed, in order to verify whether or not the thickness data alone can differentiate between
the three groups under study. No significant main effects were found between the three
groups, for any of the regions.
Regarding the classification, validation was carried out resorting to the k-fold cross-valida-
tion. k values of two, five and ten were used, where k = 2 represents the most conservative
approach, and k = 5 and k = 10, the figures traditionally used, for comparison purposes only.

PLOS ONE | https://doi.org/10.1371/journal.pone.0218826 June 21, 2019 7 / 13

 Retinal biomarkers of neurodegenerative disorders

Table 2. Full-retinal thickness. Sectors mimic the ETDRS thickness map. Inner macular areas are within the radii 500 and 1500 micrometres, and outer macular areas are
within the radii 1500 and 3000 micrometres.
Healthy Controls Alzheimer’s Disease Parkinson’s Disease p-value (ANOVA)
Central Subfield 258.2 ± 19.6 254.9 ± 17.9 255.3 ± 23.3 0.6806
Nasal Inner 326.5 ± 18.8 320.4 ± 18.4 318.4 ± 15.6 0.0555
Temporal Inner 314.2 ± 18.8 306.3 ± 15.8 307.5 ± 17.7 0.0638
Superior Inner 324.7 ± 19.4 319.7 ± 18.1 319.2 ± 17.0 0.7273
Inferior Inner 322.3 ± 18.5 316.7 ± 18.0 316.9 ± 16.5 0.2013
Nasal Outer 293.4 ± 19.0 292.1 ± 16.5 290.7 ± 13.2 0.7085
Temporal Outer 262.3 ± 16.7 258.4 ± 13.2 260.9 ± 17.3 0.5163
Superior Outer 277.0 ± 17.5 274.5 ± 14.2 275.2 ± 14.3 0.7273
Inferior Outer 264.1 ± 15.1 262.5 ± 14.0 267.5 ± 18.8 0.3284
https://doi.org/10.1371/journal.pone.0218826.t002

When the RNFL is ignored, the present analysis yields the best results. Therefore, presented
results are based on the five of the six innermost layers, RNFL excluded—namely the GCL,
IPL, INL, OPL and ONL layers.
Because the data partition (data split into training and testing sets) is random, and therefore
distinct at each run, results are presented as distributions of performance indicators from 100
consecutive runs (Table 3).
Only four of the features used in the classification process showed a strong correlation with
the thickness of the respective layer and quadrant, with the correlation coefficient ranging
from 0.70 to 0.73. Another 17 features showed a moderate correlation (greater than or equal to
0.50). None of the features presents a strong correlation in more than one of the three groups,
and only two present a strong correlation in one group and a moderate correlation in another
group. As a result, it is possible to state that texture features under analysis convey information

Table 3. Distribution for the sensitivity (SEN) (%) and specificity (SPE) (%) for the healthy control (HC), Alzheimer’s disease (AD) and Parkinson’s disease (PD)
groups, accuracy (ACC) (%), percentage of people with both eyes with the same classification (two eyes), percentage of correct classifications from the pool of people
that received the same classification on both eyes (two eyes correct), and the percentage of eyes with a tie on the classification (unknown), for three k values (k-fold
cross-validation).
== HC == == AD == == PD == Accuracy Two eyes Two eyes correct Unknown
SEN SPE SEN SPE SEN SPE
k=2 Max 90.6 89.2 82.1 94.4 87.0 100.0 82.9 81.7 96.2 4.1
3rd Q 86.8 81.7 74.4 92.5 77.8 97.8 78.8 73.2 92.7 2.1
Median 84.9 80.6 71.8 90.7 74.1 96.7 76.7 69.0 91.4 1.4
1st Q 81.1 78.5 66.7 89.7 70.4 95.1 74.7 66.2 89.3 0.7
Min 67.9 72.0 48.7 84.1 61.1 92.4 69.2 60.6 82.0 0.0
k=5 Max 94.3 88.2 84.6 97.2 87.0 100.0 87.7 83.1 96.6 4.1
3rd Q 88.7 84.9 82.0 93.5 78.7 98.9 82.2 76.1 94.3 2.1
Median 87.7 83.9 79.5 92.5 75.9 97.8 80.8 74.6 92.8 1.4
1st Q 84.9 81.7 74.4 90.6 74.1 96.7 79.5 73.2 91.9 0.7
Min 77.4 76.3 69.2 86.9 66.7 94.6 75.3 64.8 87.0 0.0
k = 10 Max 96.2 88.2 84.6 96.3 85.2 100.0 86.3 81.7 96.6 3.4
3rd Q 90.6 86.0 82.1 93.5 79.6 98.9 83.6 77.5 95.9 2.1
Median 88.7 84.9 79.5 92.5 77.8 97.8 82.2 76.1 94.4 1.4
1st Q 86.7 82.8 78.2 91.6 75.9 97.8 80.8 73.2 92.8 1.0
Min 83.0 79.6 74.4 88.8 66.7 95.7 78.1 67.6 88.9 0.0
https://doi.org/10.1371/journal.pone.0218826.t003

PLOS ONE | https://doi.org/10.1371/journal.pone.0218826 June 21, 2019 8 / 13

 Retinal biomarkers of neurodegenerative disorders

on differences present in the retina that are not conveyed by thickness. For further information
on the correlation values between these 21 features and thickness, refer to S3 Table.
The results obtained, even when considering the most conservative scenario (k = 2), clearly
indicate that these biomarkers are useful in classifying cases into one of the three considered
groups. Moreover, such biomarkers allow distinguishing between AD and PD eyes and
patients. Additionally, these results state that, in just 1 to 2% of the cases, no differences seem
to exist between the health status and either of the two neurodegenerative disorders considered
in this work, nor between the two diseases themselves. Of particular importance is the case
when the two eyes of the same subject received the same classification. In these conditions,
82% to 96% of the cases are correctly classified solely based on the analysis of the texture of the
retina, as imaged by OCT.
Leaving out the RNFL, a total of 15 classification models were employed (three classifica-
tion tests for each of the five remaining inner layers). While global texture metrics were used
in 10 of these 15 models (66.7%), they represent only 16.7% of the total used features.
The analysis of the regional origin of the features that led to the above classification results
shows an uneven contribution of the macular quadrants of the human retina. For the discrimi-
nation between the healthy control group and the AD group, over 65% (17 out of 26) features
come from the superior macular region, seven from the superior-temporal quadrant and ten
from the superior-nasal quadrant. Six and three features come from, respectively, the inferior-
temporal and inferior-nasal quadrants. Additional four features come from the global texture.
Concerning the total number of features discriminating between the healthy control group
and the PD group, there is an equal contribution from the superior and inferior regions of the
retina. Similarly, there is an equal contribution from the temporal and nasal regions. Neverthe-
less, the superior-nasal and inferior-temporal macular regions contribute with only three fea-
tures each (3/23), while the superior-temporal and inferior-nasal macular regions contribute,
respectively, with ten and seven features. Moreover, seven features come from the global tex-
ture, almost twice that of the AD group.

Discussion
Our results indicate that SVM, a supervised machine learning method, may aid in the concom-
itant clinical diagnosis of AD and PD, even in the absence of univariate differences on average
thickness. The technique herewith presented demonstrates to be able not only in discriminat-
ing between eyes of HC and patients but also in distinguishing between the eyes of AD and PD
patients. Furthermore, this method is based on the non-invasive imaging of the ocular fundus
by OCT, a widely available technique, nowadays standard in private clinical practice. Particu-
larly encouraging in terms of the reliability and reproducibility of the approach is the fact that
the vast majority of people receiving the same classification on both eyes do, in fact, have the
correct classification, with median percentages of 91.4 (2-fold cross-validation) up to 94.4
(10-fold cross-validation).
In this work, we used non-linear (radial basis function kernel) SVMs. They allow for the
identification of relevant features that distinguish between any two of the three groups, and for
the use of these features to classify cases into one of the three possible groups.
The six wavelet-based parameters used were computed from the six directionally-selective
detail subband images, which were isolated from the lower-frequency (i.e. less detailed) image
information at the first level of the wavelet decomposition process. Note that each of these sub-
bands captures the original image’s details oriented at one out of six spatial orientations (±15˚,
±45˚, ±75˚). As these parameters correspond to the variance of the magnitude of the complex
wavelet coefficients of the respective subband, they represent a measure of the spread of the

PLOS ONE | https://doi.org/10.1371/journal.pone.0218826 June 21, 2019 9 / 13

 Retinal biomarkers of neurodegenerative disorders

grey-level distribution of that same subband, and ultimately reflect the contrast of the image’s
texture.
Global texture metrics were used in 66.7% of the classification models suggesting that these
differences are spread over the entire macular region. Nevertheless, they represent only 16.7%
of the total used features, which reinforces the need for local analysis to detect actual differ-
ences between health status and neurodegenerative diseases.
In the present work, the methodological research made on OCT data is far from its classical
use, which may explain its discrimination power. While most of the reported works in the field
rely on thickness measurements, texture data (from images computed per each retinal layer)
were the metrics under evaluation here. In other words, our approach captures multivariate
information, which is just not possible with simple thickness-based univariate comparisons.
The inherent limitation of thickness-based approaches is likely a contributing factor to the
inconsistent findings reported for both AD and PD. While several studies do not find thick-
ness differences between HC and AD or PD patients [17, 19, 22], others report significant thin-
ning of different retinal layers [9, 18, 20, 21]. To add to this inconclusiveness, when thickness
differences between each of these two neurodegenerative disorders and the healthy condition
are indeed reported, they typically concern retinal thickness measurements performed on
patients of either disorder’s course at stages other than the initial stage.
Texture conveys information on the regular or irregular distribution of image intensity
and, as such, it carries information on the structural arrangement of the different retinal layers
and how they differ between the health, AD and PD conditions. Currently, the clinical diagno-
sis of these two neurodegenerative disorders is a challenging task, since there are no definite in
vivo biomarkers. Texture analysis of OCT data may thus represent a novel tool in the identifi-
cation of these diseases, providing a simple, inexpensive and non-invasive method of directly
assessing neurodegeneration.

Supporting information
S1 Fig. Relation between patient age and disease duration for the Parkinson’s disease group.
(TIF)
S2 Fig. Diagram of the classification process.
(TIF)
S1 Table. MoCA (Montreal cognitive assessment) score results for the AD patients in the
study group.
(PDF)
S2 Table. MoCA (Montreal cognitive assessment), UPDRS (unified Parkinson’s disease
rating scale)—Motor and H&Y (Hoehn and Yahr) score results for the PD patients in the
study group.
(PDF)
S3 Table. Features showing a moderate (0.5–0.7) and strong (> 0.7; highlighted in yellow)
Pearson Correlation Coefficient (PCC) with thickness, per layer, per Qi—quadrant i (see
Fig 2) for local features, and for each of the subject classes. In the features IMC1 and IMC2,
IMC stands for Informal Measure of Correlation.
(PDF)
S1 File. Full retina thickness, per sector, for all subjects.
(XLSX)

PLOS ONE | https://doi.org/10.1371/journal.pone.0218826 June 21, 2019 10 / 13

 Retinal biomarkers of neurodegenerative disorders

S2 File. Texture metrics for all subjects, quadrants, and layers.

(XLSX)

Acknowledgments
The authors would like to thank Beatriz Santiago and João Lemos for the neurological assess-
ment, Bárbara Regadas and Pedro Serranho for their assistance with the ANOVA performed
on the thickness data, Carolina Alves and Lı́lia Jorge for their support on OCT segmentation
verification and correction, and Hugo Quental for the image acquisition.

Author Contributions
Conceptualization: Rui Bernardes.
Data curation: Ana Nunes, Cristina Duque, Cristina Januário, Isabel Santana.
Formal analysis: Ana Nunes, Gilberto Silva, Miguel Castelo-Branco, Rui Bernardes.
Funding acquisition: António Francisco Ambrósio, Miguel Castelo-Branco.
Investigation: Ana Nunes, Gilberto Silva, Cristina Duque, Cristina Januário, Isabel Santana,
António Francisco Ambrósio, Miguel Castelo-Branco, Rui Bernardes.
Methodology: Ana Nunes, Gilberto Silva, Miguel Castelo-Branco, Rui Bernardes.
Project administration: Rui Bernardes.
Resources: Cristina Duque, Cristina Januário, Isabel Santana, Miguel Castelo-Branco.
Software: Ana Nunes, Gilberto Silva, Rui Bernardes.
Supervision: Rui Bernardes.
Validation: Ana Nunes, Miguel Castelo-Branco, Rui Bernardes.
Visualization: Ana Nunes, Rui Bernardes.
Writing – original draft: Ana Nunes, Rui Bernardes.
Writing – review & editing: Ana Nunes, Isabel Santana, António Francisco Ambrósio, Miguel
Castelo-Branco, Rui Bernardes.

References
1. Ferri CP, Prince M, Brayne C, Brodaty H, Fratiglioni L, Ganguli M, et al. Global prevalence of dementia:
a Delphi consensus study. Lancet. 2005; 366(9503):2112–2117. https://doi.org/10.1016/S0140-6736
(05)67889-0 PMID: 16360788
2. Alzheimer’s Association. 2016 Alzheimer’s disease facts and figures. Alzheimer’s and Dementia. 2016;
12(4):459–509. PMID: 27570871
3. World Health Organization. Neurological disorders—public health challenges. 2006.
4. Shah H, Albanese E, Duggan C, Rudan I, Langa KM, Carrillo MC, et al. Research priorities to reduce
the global burden of dementia by 2025. The Lancet Neurology. 2016; 15(12):1285–1294. https://doi.
org/10.1016/S1474-4422(16)30235-6 PMID: 27751558
5. Masland RH. The neuronal organization of the retina. Neuron. 2012; 76(2):266–280. https://doi.org/10.
1016/j.neuron.2012.10.002 PMID: 23083731
6. London A, Benhar I, Schwartz M. The retina as a window to the brain—from eye research to CNS disor-
ders. Nature Reviews Neurology. 2012. https://doi.org/10.1038/nrneurol.2012.227 PMID: 23165340
7. Svetozarskiy SNN, Kopishinskaya SVV. Retinal Optical Coherence Tomography in Neurodegenerative
Diseases (Review). Sovremennye tehnologii v medicine. 2015; 7(1):116–123. https://doi.org/10.17691/
stm2015.7.1.14

PLOS ONE | https://doi.org/10.1371/journal.pone.0218826 June 21, 2019 11 / 13

 Retinal biomarkers of neurodegenerative disorders

8. Cheung CYl, Ikram MK, Chen C, Wong TY. Imaging retina to study dementia and stroke. Progress in
Retinal and Eye Research. 2017; 57:89–107. https://doi.org/10.1016/j.preteyeres.2017.01.001 PMID:
28057562
9. den Haan J, Verbraak FD, Visser PJ, Bouwman FH. Retinal thickness in Alzheimer’s disease: a system-
atic review and meta-analysis. Alzheimer’s and Dementia: Diagnosis, Assessment and Disease Moni-
toring. 2017; 6:162–170. https://doi.org/10.1016/j.dadm.2016.12.014 PMID: 28275698
10. Hart NJ, Koronyo Y, Black KL, Koronyo-Hamaoui M. Ocular indicators of Alzheimer’s: exploring disease
in the retina. Acta Neuropathologica. 2016; 132:767–787. https://doi.org/10.1007/s00401-016-1613-6
PMID: 27645291
11. Archibald NK, Clarke MP, Mosimann UP, Burn DJ. The retina in Parkinson’s disease. Brain. 2009;
132:1128–1145. https://doi.org/10.1093/brain/awp068 PMID: 19336464
12. Tian T, Zhu XH, Liu YH. Potential role of retina as a progression of Parkinson’s disease. International
Journal of Ophthalmology. 2011; 4(4):433–438. https://doi.org/10.3980/j.issn.2222-3959.2011.04.21
PMID: 22553695
13. Ascaso F, Cruz N, Modrego P, Lopez-Anton R, Santabárbara J, Pascual L, et al. Retinal alterations in
mild cognitive impairment and Alzheimer’s disease: an optical coherence tomography study. Journal of
Neurology. 2014; 261:1522–1530. https://doi.org/10.1007/s00415-014-7374-z PMID: 24846203
14. Cheung CYL, Ong YT, Ikram MK, Ong SY, Li X, Hilal S, et al. Microvascular network alterations in the
retina of patients with Alzheimer’s disease. Alzheimer’s and Dementia. 2014; 10(2):135–142. https://
doi.org/10.1016/j.jalz.2013.06.009 PMID: 24439169
15. Chorostecki J, Seraji-Bozorgzad N, Shah A, Bao F, Bao G, George E, et al. Characterization of retinal
architecture in Parkinson’s disease. Journal of the Neurological Sciences. 2015; 355:44–48. https://doi.
org/10.1016/j.jns.2015.05.007 PMID: 26071887
16. Garcia-Martin E, Larrosa JM, Polo V, Satue M, Marques ML, Alarcia R, et al. Distribution of retinal layer
atrophy in patients with Parkinson disease and association with disease severity and duration. Ameri-
can Journal of Ophthalmology. 2014; 157(2):470–478. https://doi.org/10.1016/j.ajo.2013.09.028 PMID:
24315296
17. Aaker GD, Myung JS, Ehrlich JR, Mohammed M, Henchcliffe C, Kiss S. Detection of retinal changes in
Parkinson’s disease with spectral-domain optical coherence tomography. Clinical Ophthalmology.
2010; 4:1427–1432. https://doi.org/10.2147/OPTH.S15136 PMID: 21188154
18. Altintaş Ö, Işeri P, Özkan B, Çağlar Y. Correlation between retinal morphological and functional findings
and clinical severity in Parkinson’s disease. Documenta Ophthalmologica. 2008; 116:137–146. https://
doi.org/10.1007/s10633-007-9091-8 PMID: 17962989
19. Archibald NK, Clarke MP, Mosimann UP, Burn DJ. Retinal thickness in Parkinson’s disease. Parkinson-
ism and Related Disorders. 2011; 17:431–436. https://doi.org/10.1016/j.parkreldis.2011.03.004 PMID:
21454118
20. Hajee ME, March WF, Lazzaro DR, Wolintz AH, Shrier EM, Glazman S, Bodis-Wollner IG. Inner retinal
layer thinning in Parkinson disease. Archives of Ophthalmology. 2009; 127(6):737–741. https://doi.org/
10.1001/archophthalmol.2009.106 PMID: 19506190
21. Garcia-Martin E, Bambo MP, Marques ML, Satue M, Otin S, Larrosa JM, et al. Ganglion cell layer mea-
surements correlate with disease severity in patients with Alzheimer’s disease. Acta Ophthalmologica.
2016; 94:e454–e459. https://doi.org/10.1111/aos.12977 PMID: 26895692
22. Lad EM, Mukherjee D, Stinnett SS, Cousins SW, Potter GG, Burke JR, et al. Evaluation of inner retinal
layers as biomarkers in mild cognitive impairment to moderate Alzheimer’s disease. PLoS ONE. 2018;
13(2):e0192646. https://doi.org/10.1371/journal.pone.0192646 PMID: 29420642
23. Bernardes R, Silva G, Chiquita S, Serranho P, Ambrósio AF. Retinal biomarkers of Alzheimer’s disease:
insights from transgenic mouse models. In: 14th International Conference on Image Analysis and Rec-
ognition (ICIAR); 2017.
24. Nunes A, Ambrósio AF, Castelo-Branco M, Bernardes R. Texture biomarkers of Alzheimer’s disease
and disease progression in the mouse retina. 18th International Conference on Bioinformatics and Bio-
engineering (BIBE). 2018;.
25. Anantrasirichai N, Achim A, Morgan JE, Erchova I, Nicholson L. SVM-based texture classification in
optical coherence tomography. IEEE 10th International Symposium on Biomedical Imaging: From
Nano to Macro. 2013.
26. Mohammad S. Textural measurements for retinal image analysis. University of Manchester. 2014.
Available from: https://www.research.manchester.ac.uk/portal/files/54570132/FULL_TEXT.PDF.
27. Gao W. Improving the quantitative assessment of intraretinal features by determining both structural
and optical properties of the retinal tissue with optical coherence tomography. University of Miami.
2012. Available from: https://scholarlyrepository.miami.edu/oa_dissertations/855.

PLOS ONE | https://doi.org/10.1371/journal.pone.0218826 June 21, 2019 12 / 13

 Retinal biomarkers of neurodegenerative disorders

28. González A, Remeseiro B, Ortega M, Penedo MG, Charlón P. Automatic cyst detection in OCT retinal
images combining region flooding and texture analysis. IEEE International Symposium on Computer-
Based Medical Systems. 2013.
29. Kassner A, Thornhill RE. Texture analysis: A review of neurologic MR imaging applications. American
Journal of Neuroradiology. 2010; 31(5):809–816. https://doi.org/10.3174/ajnr.A2061 PMID: 20395383
30. Haralick RM, Shanmugam K, Dinstein I. Texture features for image classification. IEEE Transactions
on Systems, Man and Cybernetics. 1973; SMC-3(6):610–621. https://doi.org/10.1109/TSMC.1973.
4309314
31. Tomita F, Tsuji S. Computer analysis of visual textures. Springer. 1990.
32. World Medical Association. Declaration of Helsinki—ethical principles for medical research involving
human subjects. Journal of the American Medical Association. 2013; 310(20):2191–2194. https://doi.
org/10.1001/jama.2013.281053 PMID: 24141714
33. McKhann GM, Knopman DS, Chertkow H, Bradley HT, Clifford JJR, Kawas CH, et al. The diagnosis of
dementia due to Alzheimer’s disease: Recommendations from the National Institute on Aging—Alzhei-
mer’s Association workgroups on diagnostic guidelines for Alzheimer’s disease. Alzheimer’s & Demen-
tia. 2011; 7(3):263–269. https://doi.org/10.1016/j.jalz.2011.03.005
34. Nasreddine ZS, Phillips NA, Béridian V, Charbonneau S, Whitehead V, Collin I, et al. The Montreal cog-
nitive assessment, MoCA: a brief screening tool for mild cognitive impairment. Journal of the American
Geriatrics Society. 2005; 53:695–699. https://doi.org/10.1111/j.1532-5415.2005.53221.x PMID:
15817019
35. Berg L. Clinical Dementia Rating (CDR). Psychopharmacology Bulletin. 1988; 24(4):367–369.
36. Leuzy A, Chiotis K, Hasselbalch SG, Rinne JO, De Mendonça A, Otto M, et al. Pittsburgh compound B
imaging and cerebrospinal fluid amyloid-beta in a multicentre European memory clinic study. Brain.
2016; 139(9):2540–2553. https://doi.org/10.1093/brain/aww160 PMID: 27401520
37. Hughes AJ, Daniel SE, Kilford L, Lees AJ. Accuracy of clinical diagnosis of idiopathic Parkinson’s dis-
ease: A clinico-pathological study of 100 cases. Journal of Neurology Neurosurgery and Psychiatry.
1992; 55:181–184. https://doi.org/10.1136/jnnp.55.3.181
38. Guimarães P, Rodrigues P, Lobo C, Leal S, Figueira J, Serranho P, et al. Ocular fundus reference
images from optical coherence tomography. Computerized Medical Imaging and Graphics. 2014;
38:381–389. https://doi.org/10.1016/j.compmedimag.2014.02.003 PMID: 24631317
39. Yazdi M, Gheysari K. A new approach for the fingerprint classification based on gray-level co-occur-
rence matrix. World Academy of Science, Engineering and Technology. 2008; 47.
40. Clausi DA. An analysis of co-occurrence texture statistics as a function of grey level quantization. Cana-
dian Journal of Remote Sensing. 2002; 28(1):45–62. https://doi.org/10.5589/m02-004
41. Maheshwari S, Pachori RB, Acharya UR. Automated diagnosis of glaucoma using empirical wavelet
transform and correntropy features extracted from fundus images. IEEE Journal of Biomedical and
Health Informatics. 2017; 21(3). https://doi.org/10.1109/JBHI.2016.2544961 PMID: 28113877
42. Häfner M, Kwitt R, Uhl A, Gangl A, Wrba F, Vécsei A. Feature extraction from multi-directional multi-res-
olution image transformations for the classification of zoom-endoscopy images. Pattern Analysis and
Applications. 2009; 12:407–413. https://doi.org/10.1007/s10044-008-0136-8
43. Wimmer G, Tamaki T, Tischendorf JJW, Häfner M, Yoshida S, Tanaka S, et al. Directional wavelet
based features for colonic polyp classification. Medical Image Analysis. 2016; 31:16–36. https://doi.
org/10.1016/j.media.2016.02.001 PMID: 26948110
44. Etehadtavakol M, Ng EYK, Chandran V, Rabbani H. Separable and non-separable discrete wavelet
transform based texture features and image classification of breast thermograms. Infrared Physics &
Technology. 2013; 61:274–286. https://doi.org/10.1016/j.infrared.2013.08.009
45. Jian W, Sun X, Luo S. Computer-aided diagnosis of breast microcalcifications based on dual-tree com-
plex wavelet transform. BioMedical Engineering Online. 2012; 11(96). https://doi.org/10.1186/1475-
925X-11-96 PMID: 23253202
46. Selesnick IWW, Baraniuk RGG, Kingsbury NCC. The dual-tree complex wavelet transform. IEEE Signal
Processing Magazine. 2005; p. 123–151. https://doi.org/10.1109/MSP.2005.1550194
47. Celik T, Tjahjadi T. Multiscale texture classification using dual-tree complex wavelet transform. Pattern
Recognition Letters. 2009; 30:331–339. https://doi.org/10.1016/j.patrec.2008.10.006
48. Wang S, Lu S, Dong Z, Yang J, Yang M, Zhang Y. Dual-tree complex wavelet transform and twin sup-
port vector machine for pathological brain detection. Applied Sciences. 2016; 6(169).
49. Duda RO, Hart PE, Stork DG. Pattern classification. 2nd ed. Wiley-Interscience. 2000.
50. Chang CC, Lin CJ. LIBSVM: a library for support vector machines. ACM Transactions on Intelligent Sys-
tems and Technology. 2011; 2(3(27)).

PLOS ONE | https://doi.org/10.1371/journal.pone.0218826 June 21, 2019 13 / 13

Lemmens et al. Alzheimer's Research & Therapy (2020) 12:144
https://doi.org/10.1186/s13195-020-00715-1

RESEARCH Open Access

Combination of snapshot hyperspectral

retinal imaging and optical coherence
tomography to identify Alzheimer’s disease
patients
Sophie Lemmens1,2,3*† , Toon Van Craenendonck3†, Jan Van Eijgen1,2,3, Lies De Groef4, Rose Bruffaerts5,6,
Danilo Andrade de Jesus2, Wouter Charle7, Murali Jayapala7, Gordana Sunaric-Mégevand8, Arnout Standaert3,
Jan Theunis3, Karel Van Keer1,2, Mathieu Vandenbulcke9, Lieve Moons4, Rik Vandenberghe5,6,10,
Patrick De Boever3,11,12 and Ingeborg Stalmans1,2

Abstract
Introduction: The eye offers potential for the diagnosis of Alzheimer’s disease (AD) with retinal imaging techniques
being explored to quantify amyloid accumulation and aspects of neurodegeneration. To assess these changes, this
proof-of-concept study combined hyperspectral imaging and optical coherence tomography to build a
classification model to differentiate between AD patients and controls.
Methods: In a memory clinic setting, patients with a diagnosis of clinically probable AD (n = 10) or biomarker-
proven AD (n = 7) and controls (n = 22) underwent non-invasive retinal imaging with an easy-to-use hyperspectral
snapshot camera that collects information from 16 spectral bands (460–620 nm, 10-nm bandwidth) in one capture.
The individuals were also imaged using optical coherence tomography for assessing retinal nerve fiber layer
thickness (RNFL). Dedicated image preprocessing analysis was followed by machine learning to discriminate
between both groups.
(Continued on next page)

* Correspondence: [email protected]
†
Sophie Lemmens and Toon Van Craenendonck contributed equally to this
work.
1
Department of Ophthalmology, University Hospitals UZ Leuven, Herestraat
49, 3000 Leuven, Belgium
2
Department of Neurosciences, Research Group Ophthalmology, KU Leuven,
Biomedical Sciences Group, Herestraat 49, 3000 Leuven, Belgium
Full list of author information is available at the end of the article

© The Author(s). 2020 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License,
which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give
appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if
changes were made. The images or other third party material in this article are included in the article's Creative Commons
licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons
licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain
permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the
data made available in this article, unless otherwise stated in a credit line to the data.
Lemmens et al. Alzheimer's Research & Therapy (2020) 12:144 Page 2 of 13

(Continued from previous page)

Results: Hyperspectral data and retinal nerve fiber layer thickness data were used in a linear discriminant
classification model to discriminate between AD patients and controls. Nested leave-one-out cross-validation
resulted in a fair accuracy, providing an area under the receiver operating characteristic curve of 0.74 (95%
confidence interval [0.60–0.89]). Inner loop results showed that the inclusion of the RNFL features resulted in an
improvement of the area under the receiver operating characteristic curve: for the most informative region
assessed, the average area under the receiver operating characteristic curve was 0.70 (95% confidence interval [0.55,
0.86]) and 0.79 (95% confidence interval [0.65, 0.93]), respectively. The robust statistics used in this study reduces the
risk of overfitting and partly compensates for the limited sample size.
Conclusions: This study in a memory-clinic-based cohort supports the potential of hyperspectral imaging and
suggests an added value of combining retinal imaging modalities. Standardization and longitudinal data on fully
amyloid-phenotyped cohorts are required to elucidate the relationship between retinal structure and cognitive
function and to evaluate the robustness of the classification model.
Keywords: Retina, Brain, Neurodegeneration, Cognitive impairment, Alzheimer’s disease, Amyloid-beta (Aβ),
Hyperspectral imaging, Machine learning, Biomarker

Background retina holds potential to play a major role in early diag-

Diagnosing Alzheimer’s disease (AD) is a challenging nosis of AD, as also suggested by Alber et al. [9].
task. In recent years, the “ATN” categorization, which is Optical coherence tomography (OCT) is a non-
a framework for defining AD based on biomarker prox- invasive, high-resolution diagnostic tool capable of gen-
ies of pathology, where A stands for amyloid-beta (Aβ) erating cross-sectional coupes of the retina and choroid.
status, “T” for tau, and “N” for neurodegeneration Studies using this imaging modality have demonstrated
biomarkers, has gained attention [1, 2]. The highest the thinning of the RNFL, mostly in the superior and in-
diagnostic accuracy can likely be achieved by combining ferior quadrants, and the macular ganglion cell complex
several ATN biomarkers [3]. However, given the associ- [19]. This reflects the loss of the retinal ganglion cell
ated cost, invasiveness, and/or potential side effects, complex and thereby corroborates the findings from
amyloid-PET and cerebrospinal fluid (CSF) biomarker postmortem histological studies [13, 20–25]. Although
analyses are not recommended for screening [1, 4, 5]. RNFL and macular ganglion cell complex thickness have
These limitations warrant the identification of bio- been inversely correlated with disease duration and se-
markers using affordable and non-invasive diagnostic verity [26–28], longitudinal data to support the signifi-
tools [6]. cance of OCT imaging are not available, and the
Because of a shared ontogenesis, the retina displays diagnostic accuracy of RNFL changes alone is probably
similarities to the brain and spinal cord in terms of anat- insufficient due to its low specificity. The newer gener-
omy, functionality, response to insult, and immunology. ation of spectral domain OCT devices offers a markedly
Hence, the eye provides a unique window to the central improved signal-to-noise ratio. Nonetheless, imaging the
nervous system without the need for expensive, invasive, elderly poses an additional challenge due to the possibil-
and/or potentially harmful examinations [7–9]. One line ity of media opacities such as cataracts [29] and an im-
of investigation is focused on retinal changes occurring paired ability to focus properly. Most commercially
in patients with AD. There is increasing evidence point- available OCT devices offer an image quality indicator
ing to neuroretinal thinning and ganglion cell degener- to assess scan quality.
ation, abnormal electrical responses, reduced retinal Building on the unique biochemical properties of Aβ,
perfusion, and microvascular changes, as well as elevated different imaging techniques have been developed with
retinal levels of Aβ40/Aβ42 peptides and pTau [10–13]. the aim to detect changes caused by the presence of ret-
Retinal nerve fiber layer (RNFL) and macular thinning inal Aβ in vivo. One such imaging technique is based on
and loss of the melanopsin-immunopositive subtype of the use of the fluorochrome curcumin that binds to Aβ.
ganglion cells have been documented in early AD pa- This approach has shown promising results, with a
tients. Although research on the identification of patho- retinal Aβ index that correlates well with cerebral Aβ
logical Aβ accumulation in the human retina is limited plaques [14, 30–32]. Another such imaging modality,
and inconsistent, retinal Aβ accumulation and retinal Aβ hyperspectral retinal imaging (HSRI), which was recently
plaques were detected before their cerebral counterparts reviewed [33], is a label-free imaging technique. This
in both in vivo and ex vivo transgenic mouse models technique allows one to quantify a decrease in the spec-
[13–18]. Collectively, these findings suggest that the tral reflectance of retinal and cerebral tissue of AD
Lemmens et al. Alzheimer's Research & Therapy (2020) 12:144 Page 3 of 13

subjects at wavelengths between 460 and 570 nm. This provided in Additional file 1: Fig. A1. The study adhered
spectrum may be indicative of increased Rayleigh to the principles of the European Union Directive on Clin-
scattering due to the presence of Aβ [34]. Postmortem ical Trials (2001/20/EC) and all requirements for the pro-
studies in both animal and human retinas, and in vivo visions of the Declaration of Helsinki (World Medical
studies in rodents, have shown that HSRI can detect Association, Edinburgh, 2000). Approval was issued by the
spectral changes that could potentially be caused by the Ethics Committee of the University Hospitals Leuven be-
presence of retinal Aβ aggregates provided that there are fore the study commenced (reference number S59048).
such aggregates in the human retina [34, 35]. It has to
be noted, however, that HSRI does not directly visualize Participant recruitment
retinal Aβ deposition, but records a spectral shift that Participants were consecutively recruited from an aca-
could be explained by the presence of retinal protein de- demic memory-clinic-based cohort. The clinical diagnos-
posits in certain stages of aggregation, given the relation- tic workup in 7 of the 17 participants included either
ship between particle size and different types of light Fujirebio ELISA for Aβ42, total tau, and 181phosphotau
scattering. We cannot exclude that factors other than or EuroImmun ELISA of Aβ42, Aβ40, and total tau in
amyloid deposition may underly a spectral shift in AD CSF. The cut-offs used were based on Adamczuk et al.
versus controls. It has recently been shown that machine [37]. One out of 7 also underwent [11C]-Pittsburgh com-
learning methods using HSRI data are capable of distin- pound B amyloid-PET. In all 7, this led, together with
guishing between amyloid-PET-positive cases and con- the clinical evaluation, neuropsychological assessment,
trols in a clinical setting [36], which is in line with the and imaging investigations, to a diagnosis of biomarker-
assumption that these spectral differences are due to the proven AD in the dementia phase.
retinal accumulation of Aβ in amyloid-PET-positive In the ten remaining participants, a diagnosis of clinic-
cases. However, it has to be born in mind that the ally probable AD according to the National Institute on
pathological correlates of these recently reported spec- Aging-Alzheimer’s Association (NIA-AA) criteria [38]
tral changes in AD patients’ retinas have not yet been was made based on a clinical evaluation by a cognitive
identified, and as such, alternative explanations (e.g., tau neurologist (RV), blood examination, detailed neuro-
accumulation, neuro-inflammation) cannot be excluded. psychological assessment (performed in 8 out of 10) re-
This proof-of-concept clinical study investigates vealing a cognitive profile characteristic of AD, MRI
whether bimodal retinal image analysis, using both HSRI (performed in 9 out of 10) or CT (in one) for the exclu-
and OCT, can differentiate between AD patients (cases) sion of cerebrovascular disease that could explain the
and controls. The current study combines two elements cognitive decline, and in selected cases [18F] fluorodeox-
of the “ATN” categorization framework: a snapshot HSRI yglucose PET (FDG PET) (performed in 3 out of 10)
setup for in vivo detection of a spectral retinal shift pre- demonstrating a pattern characteristic of AD. All cases
sumably related to Aβ presence (“A”) is deployed, and the underwent six-monthly neurological visits for several
neurodegeneration pillar (“N”) is assessed by quantifying years (ranging from 1 to 10.5 years of follow-up). The
changes in RNFL thickness using OCT. disease course following the diagnosis was in agreement
This study examined the diagnostic performance of a with AD in all cases, including gradually progressive
set of ophthalmological measures in a clinical cohort. In cognitive decline with relative preservation of personality
such a cohort, patients who have received a prior diag- as well as the absence of clinical neurological signs be-
nosis based on standardized clinical diagnostic criteria yond the cognitive changes. Ten out of 11 in whom
are consecutively recruited. Technology assessment in a Apolipoprotein E (APOE) status was available were ε4
clinical cohort may be hampered by potentially lower carriers, and one was ε3 homozygous carrier. All cases
diagnostic accuracy compared to a research cohort. were in an early or moderate dementia stage at the time
However, this disadvantage is at least partly counterba- of study inclusion, with AD diagnosis based on thorough
lanced by the fact that a clinical cohort may be more clinical workup and follow-up, without amyloid bio-
representative for the population where this novel tech- marker confirmation.
nology will be implemented. CSF was collected, stored, and analyzed as described
by Adamczuk et al. [37]. Lumbar punctures were carried
Materials and methods out at the L4/5 level in the morning (10 a.m.–2 p.m.)
Study design and collected in polypropylene tubes (total volume 15
This single-center cross-sectional academic memory- ml, Greiner Bio-one Cellstar; VWR, Leuven, Belgium),
clinic-based study was executed during June to September discarding 1 ml to avoid traumatic blood contamination.
2019 at University Hospitals UZ Leuven, Department of Samples were centrifuged within 30 min after collection
Ophthalmology (Leuven, Belgium). An overview of the (2600 rpm, 10 min, 4 °C). After centrifugation, superna-
study with the most important steps in data analysis is tants were transferred into polypropylene tubes and
Lemmens et al. Alzheimer's Research & Therapy (2020) 12:144 Page 4 of 13

from there aliquoted in 1.5 ml polypropylene tubes (1 ml acquisition during an acquisition window of 40–70 min
volume CSF/tube; Kartell, Noviglio, Italy). In samples post-injection. A low-dose computed tomography scan
collected prior to 2018, the CSF AD biomarker assay was performed for attenuation correction, prior to the
used was the Fujirebio ELISA for Aβ42, total tau, and PET scan. Results are based on visual reads by an accre-
181
phosphotau; thereafter, the Euroimmun ELISA for dited nuclear medicine physician with special expertise
Aβ42, Aβ40, and total tau was used. Tests were per- in amyloid imaging. Diagnostic information for all 17
formed at the Laboratory Medicine Department of UZ AD cases is presented in Table 1.
Leuven, Belgium, in a ISO-15189 and Joint Commission Non-amyloid-phenotyped controls (Mini-Mental State
International-accredited environment by an expert Examination (MMSE) scores 29–30) were recruited from
technician. Cut-offs are based on Adamczuk et al. [37]: the family and/or caretakers accompanying the AD pa-
Fujirebio: Abeta42 853 pg/ml, total tau 352 pg/ml, phos- tients as well as the Department of Ophthalmology UZ
photau 86 pg/ml, Abeta42/total tau 2.258; Euroimmun Leuven. Subjects were recruited only if they were able to
Abeta42 745 pg/ml, Abeta42/Abeta40 0.096, total tau provide written informed consent. Exclusion criteria in-
436 pg/ml, Abeta42/total tau 2.006 [37]. MRI was per- cluded an age of under 55 or above 85 years, a visual
formed on a 3-T clinical MRI scanner. [18F]-FDG PET acuity worse than 20/40, presence of glaucoma or
scans were acquired using a HiRez PET-CT camera (Sie- occludable anterior chamber angle, an insufficient clarity
mens) operated in 3-dimensional mode. 18F-FDG (150 of optical media to allow retinal imaging, a personal
MBq) was injected intravenously under standard condi- medical history of retinal neovascularization or retinal
tions, that is, subjects lying supine in a dimly lit, quiet dystrophy, or the presence of retinal drusen, as well as
room, with ears and eyes open. Thirty minutes after 18F- the presence of neurological comorbidities. A total of 41
FDG injection, a dynamic scan of 30 min (6 frames of 5 subjects met the above criteria (18 AD, 23 controls).
min each) was started. During the acquisition, the One AD subject and one control subject were excluded
subject’s head was immobilized by means of a vacuum from further analysis due to the poor quality of their
pillow. 18F-FDG images were reconstructed using itera- hyperspectral images.
tive ordered-subset expectation maximization (4 itera-
tions, 4 subsets). Visual readings were based on Z map Patient examination and imaging procedures
renderings in line with current guidelines [39]. [11C]- General history and general ocular examination
Pittsburgh compound B amyloid-PET images were ac- Subjects filled in a questionnaire about their general and
quired on a TruePoint Siemens PET scanner using static ocular health history. Visual acuity (VA (logMAR)) was
Table 1 Available diagnostic information for all AD cases
AD subject Age (years) MMSE Neuropsychol. assessment Structural MRI FDG PET Amyloid-PET CSF Duration of follow-up (years)
1 82 15 + +/− − − − 10.5
2 69 18 + − + − − 1
3 63 22 + + − − − 2
4 67 27 + + − − − 4.5
5 62 ≤ 8* + +/− − + + 3.5
6 73 17 + + − − − 1.5
7 74 10 – + + − − 2
8 76 14 – + − − − 1
9 71 15 + + + − + 7
10 81 17 + + − − − 6
11 79 22 + + + − + 1
12 77 20 + +/− +/− − + 4
13 72 22 + + + − − 3.5
14 70 14 – + + − + 1.5
15 73 24 + + − − − 1.5
16 75 24 + + + − + 5
17 58 10 – + + − + 1.5
MMSE Mini-Mental State Examination score at the time of testing, Neuropsychol. assessment neuropsychological assessment as part of the diagnostic workup, +
performed and in accordance with an AD diagnosis, − not done, +/− performed but not contributive. APOE genotypes are not provided for confidentiality reasons
*MMSE no longer possible at the time of ocular imaging; noted score is the latest available one
Lemmens et al. Alzheimer's Research & Therapy (2020) 12:144 Page 5 of 13

determined in both eyes, and the better eye was included Basic statistics
for further examination and imaging. In case of symmet- Statistical analyses on patient characteristics and OCT
ric visual acuity, one eye was randomly chosen. A gen- parameters were performed using SPSS 26.0 for Win-
eral ophthalmological examination of the eye being dows (SPSS Inc., Chicago, USA). After evaluation of the
studied was performed, including biomicroscopy, kerato- distribution of the results for normality, differences were
metry, and intraocular pressure (IOP) measurement by analyzed using an independent sample t test for continu-
pneumotonometry using Tonoref II (Nidek Co Ltd., ous parameters. Ordinal and dichotomous parameters
Aichi, Japan); dilated fundoscopy (tropicamide 0.5% and were compared using the Mann-Whitney and chi-square
phenylephrine 2.5%), stereoscopic optic disc photog- testing, respectively. Analysis of RNFL parameters has
raphy, and macula-centered fundus photography using been corrected for age, gender, and scan quality using
the Visucam PRO NM (Carl Zeiss Meditec AG, Jena, multivariate linear regression. Statistical significance was
Germany); and ultra-widefield scanning laser ophthal- determined based on two-sided P values of < 0.05. A
moscopy (UWF-SLO) imaging using Optomap Daytona Bonferroni correction for multiple comparisons has been
P200C UWF-SLO (Optos Plc, Dunfermline, UK). applied to the RNFL parameters.

Hyperspectral image analysis

Hyperspectral retinal imaging
Definition of regions of interest
HSRI was performed with a XIMEA SNm4x4 VIS hyper-
Raw reflectances were first converted to relative reflec-
spectral snapshot camera (Ximea, Münster, Germany;
tances using the preprocessing steps outlined in
XIMEA CamTool software version 4.11) connected with a
Additional file 2: Preprocessing. Subsequently, pixels cor-
C-mount to a TL-230T relay lens (Topcon Corporation,
responding to blood vessels were identified as described in
Japan), installed on a Topcon TRC-50DX fundus camera
Additional file 2: Removal of retinal blood vessels and dis-
(Topcon Corporation, Japan) (see Additional file 1: Fig. A1).
carded for further analysis. Four rectangular regions of
The XIMEA snapshot camera contains a hyperspectral
interest (ROIs) as determined relative to the line going
sensor from IMEC. This mosaic pattern of pixel-size
through the center of the optic disc (OD) and fovea were
spectral filters is integrated on top of a standard comple-
defined for standardization of analysis between subjects
mentary metal oxide semiconductor (CMOS) sensor
(Fig. 1). This approach is comparable to the one described
(1088 × 2048 pixels). This allows acquiring spatial and
by Hadoux et al. [36] The a priori selection of ROIs limits
spectral information (460–620 nm, 10-nm bandwidth) in
the risk of diluting a possibly weak Aβ signal when consid-
one capture (272 × 512 pixels) without the need for
ering the entire retina and of detecting random effects
wavelength or spatial scanning by combining 4 × 4 im-
when considering a greater number of regions. In this
aging pixels into hyperspectral pixels with 16 spectral
study, locations of the center of the OD and the fovea
bands [40, 41]. Settings for the image acquisition con-
were marked manually. Each of the ROIs has a height of
sisted of an exposure time of 0.2 ms, 50-degree field of
40 pixels and a width equal to 35% of the distance be-
view, and no background illumination. Macula-centered
tween the center of the OD and the fovea. The width of
images were recorded. The first image of each study eye
the ROIs was defined relative to the OD-fovea distance to
was captured with a flash intensity of 50 Ws, which was
guarantee that the ROIs did not overlap. The range of the
subsequently increased or decreased to capture an image
widths of the ROIs varied between 32 and 52 pixels. The
with maximum light intensity while avoiding saturation
relative reflectance values of the spectrum were averaged
outside the optic nerve head (ONH). For the patient, the
in the four individual ROIs and standardized using the
acquisition of one hyperspectral snapshot image implies
procedure described in Additional file 2: Standardization,
exposure to one flash of low to moderate intensity dur-
resulting in a 14 × 1 vector for each ROI.
ing an acquisition time of 0.2 ms, while focusing on an
external fixation light.
Classification model’s training and evaluation
Linear discriminant analysis (LDA) classifiers were
Optical coherence tomography trained to distinguish AD subjects from controls. LDA
OCT measurements were performed and analyzed using was chosen as it is a linear classifier that does not re-
the RT-vue XR Avanti (Optovue, Fremont, CA, USA; quire hyperparameter tuning, which makes it suitable for
software version 2015.1.1.98). RNFL thickness and our use case as we do not have a large enough sample
vertical cup-to-disc ratio (vCDR) were recorded from size to train more complex models and do associated
the ONH report. RNFL thickness was measured over hyperparameter tuning. Models were trained using
360° (RNFLAVG, average) and per quadrant (RNFLSUP, scikit-learn library (version 0.21.3) in a Python program-
superior; RNFLNAS, nasal; RNFLINF, inferior; RNFLTEM, ming language environment [42]. In the model selection
temporal). procedure, performances obtained for the predefined
Lemmens et al. Alzheimer's Research & Therapy (2020) 12:144 Page 6 of 13

Fig. 1 Illustration of the positioning of the 4 rectangular regions of interest. Regions are indicated by superior 1 (S1), superior 2 (S2), inferior 1 (I1),
and inferior 2 (I2). The green zones refer to the parts in the image that were used in the analysis after removing the retinal blood vessels. OD
refers to the optic disc

ROIs and their combination with RNFL features were visual acuity (BCVA), prevalence of pseudophakia, and
compared. ROIs were included in the model selection RNFL thickness (average and inferior) were statistically
procedure if the standard error of the mean (SEM) inter- significantly lower in the AD group. An overview of de-
vals of the average spectrum for AD patients and con- mographical and clinical characteristics is given in
trols did not overlap for at least one wavelength. For Table 2.
each ROI, two input configurations for the classifier
were considered: one that consisted of normalized
hyperspectral data, and one that combined normalized Results of multimodal image analysis
hyperspectral data and 5 RNFL features, one for each Normalized mean reflectance spectra are shown in
OCT quadrant and the averaged value over the 4 OCT Fig. 2a. For ROIs, S1, and I2, the SEM values did not
quadrants. The performance of the selected ROIs, with overlap between AD and controls for at least one wave-
and without combining the spectra with RNFL features, length, indicating that the population mean of AD sub-
was compared using nested leave-one-out cross- jects is different from that of controls. These ROIs were
validation (LOOCV). Note that we limit the number of selected for the model selection procedure. Hence, four
regions that we consider and the associated feature com- configurations were considered as input to the machine
binations (i.e., only consider spectra with or without learning model: S1 spectra, I2 spectra, S1 spectra +
RNFL features, without doing further extensive feature RNFL thickness, and I2 spectra + RNFL thickness. The
selection) with the aim of reducing the risk of overfitting first two configurations consist of 14 input features, one
in the model selection procedure. We refer to Additional for each wavelength. The last two configurations consist
file 2: Nested leave-one-out cross-validation for a brief of 19 input features, one for each wavelength and 5 add-
description of nested LOOCV and to Varma et al. [43] itional ones that represent RNFL thickness values.
for a more detailed description. The inner LOOCV loop results from the LDA allow
comparing the different model configurations. Figure 2b
Results shows the average ROC curves for the inner cross-
Patient characteristics validation (CV) runs. For the S1 region, the average area
The cohort consisted of 17 participants with AD and 22 under the curve (AUC) is 0.67 (95% CI [0.51, 0.83]) with
cognitively intact controls, as described under the only spectra as input to the model and 0.72 (95% CI
“Participant recruitment” section. There were no signifi- [0.57, 0.88]) with spectra and RNFL features as input.
cant differences in age and sex distribution between the For the I2 region, the average AUC is 0.70 (95% CI
AD and the control groups. MMSE, the best corrected [0.55, 0.86]) and 0.79 (95% CI [0.65, 0.93]), respectively.
Lemmens et al. Alzheimer's Research & Therapy (2020) 12:144 Page 7 of 13

Table 2 Demographical and clinical characteristics

Parameter Alzheimer’s disease (AD) patients (n = 17) Controls (n = 22) P
Time since AD diagnosis (years) 2.7 ± 2.6 NA –
Age (years) 71.9 ± 6.6 68.6 ± 8.4 0.193*
Sex (male/female) 7/10 12/9 0.267‡
2
Body mass index (kg/m ) 24.9 ± 2.9 26.0 ± 4.4 0.412*
Eye (right/left) 10/7 10/12 0.408‡
MMSE 17.6 ± 5.5 29.3 ± 0.9 < 0.001†
BCVA (logMAR) 0.14 ± 0.11 0.06 ± 0.08 0.027*
IOP (mmHg) 14 ± 3 15 ± 4 0.359*
Phakic (yes/no) 15/2 12/10 0.024‡
Vertical cup/disc ratio 0.52 ± 0.15 0.51 ± 0.21 0.878*
RNFLAVG (μm) 84.8 ± 7.5 92.1 ± 7.3 0.005§
RNFLSUP (μm) 104.2 ± 8.9 109.8 ± 12.4 0.019§
RNFLINF (μm) 104.3 ± 11.1 115.6 ± 11.4 0.009§
RNFLTEM (μm) 63.3 ± 8.1 70.4 ± 6.7 0.069§
RNFLNAS (μm) 66.3 ± 11.7 72.7 ± 8.6 0.012§
MMSE Mini-Mental State Examination, BCVA best corrected visual acuity, RNFL retinal nerve fiber layer. Data are presented as mean ± standard deviation
*Independent samples t test
†
Mann-Whitney U test
‡
Chi-square test
§
Multivariate linear regression corrected for age, gender, and image quality

Inclusion of the RNFL features resulted in an improve- underlining the added value of the bimodal imaging ap-
ment of the AUC in both the S1 and I2 regions. proach. Within the field of glaucoma, the most prevalent
The results of the inner LOOCV runs consistently ocular neurodegenerative disorder, peripapillary RNFL
provided the I2 region combined with RNFL thickness measurements are the most standardized across the vari-
values to be selected for validation in the outer loop. ous available OCT devices [44], partly because this is
The I2 region and RNFL thickness values were selected one clearly defined anatomical layer. Thinning of the
in 38 out of 39 inner runs. Figure 3a shows the final peripapillary RNFL is directly associated with the struc-
ROC curve generated for predictions in the outer tural loss of ganglion cell axons in the retina of glau-
LOOCV loop. An AUC of 0.74 with a 95% confidence coma patients [45]. Significant thinning of the RNFLAVG
interval of [0.60–0.89] was obtained. The AUC gener- was observed in AD patients, most pronounced in the
ated in this nested LOOCV is an unbiased estimate inferior quadrant. Previous cross-sectional studies using
according to Varma and Simon [43]. Fifteen out of the OCT have demonstrated that RNFL thinning in AD pa-
22 controls had a probability of having AD close to 0, tients is not uniform and most pronounced in the super-
and 9 out of 17 AD patients had a score near 1. Figure 3b ior and inferior quadrants. Of note, anatomically, the
shows the distribution of the AD probabilities that were RNFL fibers converge in the superior and inferior quad-
produced in the outer LOOCV loop. There were no sig- rants, giving rise to the characteristic “double-hump”
nificant differences in non-retinal parameters between pattern of RNFL thickness and rendering those quad-
AD patients with high and low probability scores. Of rants more discriminatory for changes in thickness [16,
note, comparison of spectral properties between 19]. From a meta-analysis by den Haan et al., it has been
biomarker-proven and non-biomarker-proven AD sub- shown that mean peripapillary RNFL thickness is de-
jects did not reveal any significant difference. creased by 9.70 μm in AD versus control, with a larger
effect for time domain OCT compared to spectral do-
Discussion main OCT. [19] Chan et al. performed a meta-analysis
This clinical study reports a proof-of-concept for a bi- limited to studies using spectral domain OCT and re-
modal imaging approach using hyperspectral and OCT ported a decrease in mean peripapillary RNFL of
imaging to detect retinal changes related to Aβ presence. 5.99 μm [46]. These findings are in line with our study,
All retinal data were fed to a dedicated analysis pipeline which was performed with spectral domain OCT, and
to discriminate AD patients from controls. The perform- where a decrease in mean peripapillary RNFL of 7.7 μm
ance of the current machine learning model improved (P = 0.008) was noted. Both longitudinal studies that
with the addition of peripapillary RNFL data as input, have been published so far consistently indicate that
Lemmens et al. Alzheimer's Research & Therapy (2020) 12:144 Page 8 of 13

Fig. 2 a Mean spectra in the 4 ROIs after normalization. Shaded areas indicate the mean ± the standard error of the mean. S1 and S2 refer to the
superior regions, and I1 and I2 refer to the inferior regions (cfr. Fig. 2). b Average receiving operating characteristic (ROC) curves over all inner
loop cross-validation runs for all configurations. S1 and I2 refer to models taking only spectra as input, and S1+RNFL and I2+RNFL refer to models
with both spectra and retinal nerve fiber layer (RNFL) thickness as input

specifically in the inferior quadrant RNFL thinning is as- AUC of 0.87 (CI 0.69–1.0) on a separate validation co-
sociated with progression of cognitive decline in AD pa- hort of 4 AD patients. While Hadoux et al. used only
tients [47, 48]. hyperspectral data in their machine learning pipeline,
In the current study, an AUC of 0.74 (CI 0.60–0.89) the current study reports the training of a multimodal
was obtained using a combination of hyperspectral and model and a validation using nested LOOCV. But there
RNFL data, whereas Hadoux et al. [36] report an AUC are also important differences in hardware characteris-
of 0.82 (CI 0.67–0.97) on their principal validation set tics: whereas the off-the-shelf available snapshot camera
consisting of fellow eyes of training subjects, and an used in the current study relies on a mosaic pattern of
Lemmens et al. Alzheimer's Research & Therapy (2020) 12:144 Page 9 of 13

Fig. 3 a Receiver operating characteristic (ROC) curve generated through nested leave-one-out cross-validation (LOOCV). Area under the curve
(AUC) given with 95% confidence interval. b Distribution of AD probabilities. Probabilities predicted by the models in the outer LOOCV loop for
AD patients (top) and cognitively intact elderly (CIE) subjects (bottom)

pixel-size spectral filters integrated on top of a standard resolution of the snapshot camera, at the expense of
CMOS sensor, Hadoux et al. made use of a wavelength longer acquisition times and higher hardware cost. This
scanning HSRI technique (metabolic hyperspectral ret- probably accounts for the difference in the AUC esti-
inal camera (MHRC), Optina Diagnostics, Montreal, mate between this study and the study by Hadoux et al.,
Canada). The latter outweighs the spectral and spatial but it should be pointed out that both CIs entirely
Lemmens et al. Alzheimer's Research & Therapy (2020) 12:144 Page 10 of 13

overlap and thus cannot be considered significantly dif- presence of retinal Aβ, More et al. [34] have previously
ferent. Most importantly, the current study shows that substantiated the hypothesis that the observed spectral
with a cheaper and faster HSRI technique, boosted with effects observed in these wavelengths were caused by the
data from the already widely available OCT technique, presence of soluble Aβ in the retina. This hypothesis is
comparable results can be achieved. More et al. [35] pro- based on a simulation of the light paths through the dif-
vide a biostatistical analysis of the differences in optical ferent retinal layers and the proposition that the accu-
density between AD subjects and controls, but they did mulation of soluble Aβ aggregates in the retina causes
not develop a classification model and consequently did additional Rayleigh scattering over time, which leads to a
not report performance results that allow for a direct reduction in measured light at shorter wavelengths
comparison with the present study. Sharafi et al. [49] within a recording aperture [41]. The issue of retinal ac-
also trained a classifier to distinguish AD subjects from cumulation of Aβ in AD remains controversial, with di-
controls based on hyperspectral images. They extracted vergent results across research groups and studies,
vessel tortuosity and diameter as well as several spatial- which might at least partly be explained by the signifi-
spectral texture measures in different retinal anatomical cant heterogeneity in techniques for staining and tissue
regions. Their best model obtained a classification accur- preparation, and in study design [9]. Opposite each of
acy of 85%. Sharafi et al. [49] used single-level CV to the handful of studies that support the hypothesis of
perform both model selection and evaluation, which retinal Aβ in AD [13, 14, 34–36, 51, 52], one can be put
may have resulted in over-optimistic performance esti- that could not confirm the presence of Aβ plaques in
mates [43]. Of note, in the present study, a nested the human AD retina using immunohistochemistry
LOOCV was used to obtain an unbiased estimate of per- [53–56]. Nevertheless, further standardization of
formance. This method is more appropriate for smaller ex vivo and in vivo methods is crucial to evolve towards
sample sizes and probably reflects the true accuracy bet- clearance of this controversy. Such studies, investigat-
ter than single-level CV. ing pathological processes in the retina, such as protein
In the present study, four ROIs positioned relative to depositions of Aβ or tau, but also neuro-inflammation,
the OD-fovea line were selected to ensure consistent are required to assess what underlies the spectral
sampling locations. Similar to Hadoux et al. [36], the lar- changes in AD retinas.
gest HSRI differences were observed in the S1 region
(cfr. Fig. 2a). The subsequent model selection procedure, Limitations
however, identified the I2 region as the most informative The present study should be interpreted within the con-
one to discriminate between AD and controls, both text of its strengths and limitations. First, the lack of bio-
when considering HSRI results only and when based on marker confirmation in the majority of subjects is a
a bimodal approach combining HSRI and OCT. Con- limitation. Nevertheless, although only 7 out of 17 AD
cerning the spectral shifts measured by HSRI, one could subjects had a biomarker-proven AD diagnosis, they all
hypothesize that the I2 region shows relative differences fulfilled the widely used NIA-AA criteria for the diagno-
compared to the other regions regarding blood flow, ret- sis of probable AD [38]. To assess the diagnostic per-
inal vascular reactivity, tissue composition and perme- formance of the ophthalmological markers, the study
ability, and light stimulation, which could make it more participants were recruited consecutively from an aca-
susceptible to deposition/less susceptible to clearance of demic memory-clinic-based cohort of patients who had
proteins, such as amyloid. Within the field of glaucoma, received a prior diagnosis of either biomarker-proven
the most prevalent ocular neurodegenerative disorder, AD or clinically probable AD in an early or moderate
the inferior temporal peripapillary neuronal tissue is dementia stage. The diagnostic investigations which led
most prone to glaucomatous damage resulting in thin- to a diagnosis of AD were thorough, but CSF or
ning of the retinal neuronal tissue. It has been postulated amyloid-PET biomarker tests were only available for
that differences in vascular hemodynamics (less respon- those patients in whom this was considered clinically in-
sive to vasodilation and more responsive to vasoconstric- dicated. Although the value of the current results ob-
tion) might contribute to this finding [50]. Such tained in a clinical cohort is considerable, future
differences could equally contribute to regional differ- research building on the results of this pilot study should
ences in protein deposition and clearance, affecting focus on data collection in fully Aβ-phenotyped cohorts.
HSRI results. The largest difference in relative reflect- Second, the HSRI setup used here relies on snapshot im-
ance spectra in the current study was observed at aging. This is both a strength and a limitation. Previous
shorter wavelengths (< 550 nm), which is consistent with studies [35, 42] have used the MHRC, which offers su-
the observations by More et al. [35] and Hadoux et al. perior spatial and spectral resolution, but requires longer
[36]. While further research is needed to ensure that the acquisition times. On the other hand, More et al. [34,
observed effects are not due to other factors than the 35] have developed a custom imaging system that
Lemmens et al. Alzheimer's Research & Therapy (2020) 12:144 Page 11 of 13

simultaneously captures a conventional two-dimensional Abbreviations

retinal image and a spectral image along one dimension. Aβ: Amyloid-beta; AD: Alzheimer’s disease; ATN: Amyloid-beta, tau,
neurodegeneration; AUC: Area under the curve; BCVA: Best corrected visual
While this setup provides short acquisition times, it only acuity; CI: Confidence interval; CMOS: Complementary metal oxide
provides spectral information along a single horizontal semiconductor; CSF: Cerebrospinal fluid; CV: Cross-validation; [18F]-FDG
line. The XIMEA snapshot camera in the current setup PET: [18F] fluorodeoxyglucose PET; IOP: Intraocular pressure;
HSRI: Hyperspectral retinal imaging; LDA: Linear discriminant analysis;
overcomes several of these issues. Spatial and spectral in- LOOCV: Leave-one-out cross-validation; MHRC: Metabolic hyperspectral
formation is obtained in one take, thus enabling real-time retinal camera; MMSE: Mini-Mental State Examination; OCT: Optical
data acquisition which is crucial to avoid eye movements coherence tomography; OD: Optic disc; ONH: Optic nerve head;
RNFL: Retinal nerve fiber layer; ROI: Region of interest; SEM: Standard area of
in retinal imaging, although at the cost of spatial and spec- the mean; VA: Visual acuity; vCDR: Vertical cup-to-disc ratio
tral resolution [40, 41]. Third, the current study included
only peripapillary OCT data. Further research into the use Acknowledgements
The authors express their gratitude to Ximea for loaning the XIMEA SNm4x4
of macular OCT parameters, such as the ganglion cell VIS hyperspectral snapshot camera for the duration of this project.
complex and the ganglion cell-inner plexiform layer com- Furthermore, we thank Elsa Lauwers and the Mission Lucidity initiative for
plex, within the proposed model for bimodal retinal im- their support in setting up this study. To conclude, we would like to express
our thanks to Marie-Isaline Billen, Freya Cachet, Astrid Margot, Sarah Spileers,
aging in AD is a path that should be explored in the Axelle Stockmans, Jana Van Laeken, and Aurélie Van Schoote for their assist-
future. Finally, the robust statistics using the LOOCV ap- ance in the data collection.
proach in this study partly compensate for the limited
Authors’ contributions
sample size. The limited sample size was also the motiv- SL designed the study, performed the experiments, analyzed the data, and
ation to select only the regions for which the SEM did not wrote the manuscript; TVC analyzed the data and wrote the manuscript; JVE
overlap for at least one wavelength in the model selection performed the experiments; LDG, KVK, and LM made substantial
contributions to the conception of the work and revised the manuscript; AS
procedure. While this selection could introduce a selec- and JT analyzed the data and revised the manuscript; DADJ, WC, and MJ
tion bias, it also mitigates the risk of overfitting by redu- contributed to the technical study design and revised the manuscript; GSM
cing the number of configurations to consider. contributed to the data interpretation and revised the manuscript; RB and
MV recruited patients and revised the manuscript; PDB designed the study,
The results of the current study provide potential for supervised the data analysis, and revised the manuscript; RV and IS designed
future research. First, these findings should be confirmed the study, recruited the patients, supervised the experiments, and revised the
in a larger, fully Aβ-phenotyped cohort with an assess- manuscript. The authors read and approved the final manuscript.
ment of the association between AD probability scores
Funding
and Aβ status/load. Second, additional parameters could Sophie Lemmens and Jan Van Eijgen are holders of a joint VITO-UZ Leuven
be integrated in the multimodal retinal imaging model PhD grant. Part of this research work has been funded in the context of the
to investigate their potential to further improve the dis- VITO-UZ Leuven HERALD project that was granted by the ATTRACT consor-
tium, which received funding from the European Union’s Horizon 2020 Re-
criminatory performance. OCT-angiography [57–60], search and Innovation Programme (2014–2020). Lies De Groef and Rose
Doppler OCT [61, 62], and/or systemic variables such as Bruffaerts are postdoctoral fellows of the Research Foundation Flanders.
age, sex, APOE status, or blood pressure could provide
Availability of data and materials
added value in this respect. The datasets used and/or analyzed during the current study are available
from the corresponding author on reasonable request.
Conclusion
Ethics approval and consent to participate
Retinal imaging offers a fast and straightforward method Subjects were recruited only if they were able to provide written informed
to examine the central nervous system, allowing direct as- consent. The study adhered to the principles of the European Union
sessment of neurodegeneration, possibly reflecting Ab de- Directive on Clinical Trials (2001/20/EC) and all requirements for the
provisions of the Declaration of Helsinki (World Medical Association,
position or other neurodegenerative-related changes. This Edinburgh, 2000). Approval was issued by the Ethics Committee of the
study supports the idea that hyperspectral imaging and University Hospitals Leuven before the study commenced (reference number
OCT, combined with a machine learning approach, can S59048).
contribute to a classification model for the detection of
Consent for publication
AD. It further supports the idea that this can be achieved Not applicable.
with a low-cost, compact, and easy-to-use snapshot cam-
era mounted on top of a standard fundus camera. Competing interests
The authors declare that they have no competing interest.

Author details
Supplementary Information 1
Department of Ophthalmology, University Hospitals UZ Leuven, Herestraat
The online version contains supplementary material available at https://doi.
49, 3000 Leuven, Belgium. 2Department of Neurosciences, Research Group
org/10.1186/s13195-020-00715-1.
Ophthalmology, KU Leuven, Biomedical Sciences Group, Herestraat 49, 3000
Leuven, Belgium. 3VITO (Flemish Institute for Technological Research), Health
Additional file 1. Study set-up depicting the multimodal retinal Unit, Boeretang 200, 2400 Mol, Belgium. 4Neural Circuit Development and
imaging set-up, image processing and analysis. Regeneration Research Group, Department of Biology, KU Leuven,
Additional file 2. Details on hyperspectral image analysis. Naamsestraat 61, 3000 Leuven, Belgium. 5Laboratory for Cognitive
Neurology, Department of Neurosciences, KU Leuven, Herestraat 49, 3000
Lemmens et al. Alzheimer's Research & Therapy (2020) 12:144 Page 12 of 13

Leuven, Belgium. 6Department of Neurology, University Hospitals UZ Leuven, 18. Garcia-Martin ES, Rojas B, Ramirez AI, De Hoz R, Salazar JJ, Yubero R, et al.
Herestraat 49, 3000 Leuven, Belgium. 7Imec, Kapeldreef 75, 3001 Leuven, Macular thickness as a potential biomarker of mild Alzheimer’s disease.
Belgium. 8Clinical Research Center, Mémorial A. de Rothschild, 22 Chemin Ophthalmology. 2014. https://doi.org/10.1016/j.ophtha.2013.12.023.
Beau Soleil, 1208 Geneva, Switzerland. 9Division of Psychiatry, University 19. den Haan J, Verbraak FD, Visser PJ, Bouwman FH. Retinal thickness in
Hospitals Leuven, Herestraat 49, 3000 Leuven, Belgium. 10Alzheimer Research Alzheimer’s disease: a systematic review and meta-analysis. Alzheimer’s
Center KU Leuven, Leuven Brain Institute, Herestraat 49, 3000 Leuven, Dement Diagnosis, Assess Dis Monit 2017. doi:https://doi.org/10.1016/j.
Belgium. 11Hasselt University, Center of Environmental Sciences, Agoralaan, dadm.2016.12.014.
3590 Diepenbeek, Belgium. 12Department of Biology, University of Antwerp, 20. Blanks JC, Hinton DR, Sadun AA, Miller CA. Retinal ganglion cell
Universiteitsplein 1, 2610 Wilrijk, Belgium. degeneration in Alzheimer’s disease. Brain Res. 1989. https://doi.org/10.
1016/0006-8993(89)90653-7.
Received: 6 August 2020 Accepted: 22 October 2020 21. Hinton DR, Sadun AA, Blanks JC, Miller CA. Optic-nerve degeneration in
Alzheimer’s disease. N Engl J Med. 1986. https://doi.org/10.1056/
NEJM198608213150804.
22. Blanks JC, Schmidt SY, Torigoe Y, Porrello KV, Hinton DR, Blanks RH. Retinal
References
pathology in Alzheimer’s disease. II. Regional neuron loss and glial changes
1. Jack CR, Bennett DA, Blennow K, Carrillo MC, Dunn B, Haeberlein SB, et al.
in GCL. Neurobiol Aging. 1996.
NIA-AA Research Framework: toward a biological definition of Alzheimer’s
23. Blanks JC, Torigoe Y, Hinton DR, Blanks RHI. Retinal pathology in Alzheimer’s
disease. Alzheimers Dement. 2018. https://doi.org/10.1016/j.jalz.2018.02.018.
disease. I. Ganglion cell loss in foveal/parafoveal retina. Neurobiol Aging
2. Jack CR, Bennett DA, Blennow K, Carrillo MC, Feldman HH, Frisoni GB, et al.
1996. doi:https://doi.org/10.1016/0197-4580(96)00010-3.
A/T/N: an unbiased descriptive classification scheme for Alzheimer disease
24. Curcio CA, Drucker DN. Retinal ganglion cells in Alzheimer’s disease and
biomarkers. Neurology. 2016. https://doi.org/10.1212/WNL.
aging. Ann Neurol. 1993. https://doi.org/10.1002/ana.410330305.
0000000000002923.
25. Sadun AA, Bassi CJ. Optic nerve damage in Alzheimer’s disease.
3. Frisoni GB, Boccardi M, Barkhof F, Blennow K, Cappa S, Chiotis K, et al.
Ophthalmology. 1990. https://doi.org/10.1016/S0161-6420(90)32621-0.
Strategic roadmap for an early diagnosis of Alzheimer’s disease based on
biomarkers. Lancet Neurol. 2017. https://doi.org/10.1016/S1474- 26. Bambo MP, Garcia-Martin E, Otin S, Pinilla J, Larrosa JM, Polo V, et al. Visual
4422(17)30159-X. function and retinal nerve fibre layer degeneration in patients with
4. Blennow K, Dubois B, Fagan AM, Lewczuk P, De Leon MJ, Hampel H. Clinical Alzheimer disease: correlations with severity of dementia. Acta Ophthalmol.
utility of cerebrospinal fluid biomarkers in the diagnosis of early Alzheimer’s 2015. https://doi.org/10.1111/aos.12635.
disease. Alzheimers Dement. 2015. https://doi.org/10.1016/j.jalz.2014.02.004. 27. Iseri PK, Altinaş Ö, Tokay T, Yüksel N. Relationship between cognitive
5. Grossman I, Lutz MW, Crenshaw DG, Saunders AM, Burns DK, Roses AD. impairment and retinal morphological and visual functional abnormalities in
Alzheimer’s disease: diagnostics, prognostics and the road to prevention. Alzheimer disease. J Neuro-Ophthalmology. 2006. https://doi.org/10.1097/
EPMA J. 2010. https://doi.org/10.1007/s13167-010-0024-3. 01.wno.0000204645.56873.26.
6. Cummings J, Aisen PS, Dubois B, Frölich L, Jack CR, Jones RW, et al. Drug 28. Garcia-Martin E, Bambo MP, Marques ML, Satue M, Otin S, Larrosa JM, et al.
development in Alzheimer’s disease: the path to 2025. Alzheimers Res Ther. Ganglion cell layer measurements correlate with disease severity in patients
2016. https://doi.org/10.1186/s13195-016-0207-9. with Alzheimer’s disease. Acta Ophthalmol. 2016. https://doi.org/10.1111/
7. London A, Benhar I, Schwartz M. The retina as a window to the brain - from aos.12977.
eye research to CNS disorders. Nat Rev Neurol. 2013. https://doi.org/10. 29. El-Ashry M, Appaswamy S, Deokule S, Pagliarini S. The effect of
1038/nrneurol.2012.227. phacoemulsification cataract surgery on the measurement of retinal nerve
8. De Groef L, Cordeiro MF. Is the eye an extension of the brain in central fiber layer thickness using optical coherence tomography. Curr Eye Res.
nervous system disease? J Ocul Pharmacol Ther. 2018. https://doi.org/10. 2006. https://doi.org/10.1080/02713680600646882.
1089/jop.2016.0180. 30. Koronyo Y, Salumbides BC, Black KL, Koronyo-Hamaoui M. Alzheimer’s
9. Alber J, Goldfarb D, Thompson LI, Arthur E, Hernandez K, Cheng D, et al. disease in the retina: imaging retinal Aβ plaques for early diagnosis and
Developing retinal biomarkers for the earliest stages of Alzheimer’s disease: therapy assessment. Neurodegener Dis. 2012. https://doi.org/10.1159/
what we know, what we don’t, and how to move forward. Alzheimers 000335154.
Dement. 2020. https://doi.org/10.1002/alz.12006. 31. Shah TM, Gupta SM, Chatterjee P, Campbell M, Martins RN. Beta-amyloid
10. Hart NJ, Koronyo Y, Black KL, Koronyo-Hamaoui M. Ocular indicators of sequelae in the eye: a critical review on its diagnostic significance and
Alzheimer’s: exploring disease in the retina. Acta Neuropathol. 2016. https:// clinical relevance in Alzheimer’s disease. Mol Psychiatry. 2017. https://doi.
doi.org/10.1007/s00401-016-1613-6. org/10.1038/mp.2016.251.
11. Kusne Y, Wolf AB, Townley K, Conway M, Peyman GA. Visual system 32. Kayabasi U. Retinal examination for the diagnosis of Alzheimer’s disease. Int
manifestations of Alzheimer’s disease. Acta Ophthalmol. 2017. https://doi. J Ophthalmic Pathol. 2014. https://doi.org/10.4172/2324-8599.1000145.
org/10.1111/aos.13319. 33. Lemmens S, Devulder A, Van Keer K, Bierkens J, De Boever P, Stalmans I.
12. Lim JKH, Li QX, He Z, Vingrys AJ, Wong VHY, Currier N, et al. The eye as a Systematic review on fractal dimension of the retinal vasculature in
biomarker for Alzheimer’s disease. Front Neurosci. 2016. https://doi.org/10. neurodegeneration and stroke: assessment of a potential biomarker. Front
3389/fnins.2016.00536. Neurosci. 2020. https://doi.org/10.3389/fnins.2020.00016.
13. La Morgia C, Ross-Cisneros FN, Koronyo Y, Hannibal J, Gallassi R, Cantalupo 34. More SS, Beach JM, Vince R. Early detection of amyloidopathy in Alzheimer’s
G, et al. Melanopsin retinal ganglion cell loss in Alzheimer disease. Ann mice by hyperspectral endoscopy. Investig Ophthalmol Vis Sci. 2016. https://
Neurol. 2016. https://doi.org/10.1002/ana.24548. doi.org/10.1167/iovs.15-17406.
14. Koronyo-Hamaoui M, Koronyo Y, Ljubimov AV, Miller CA, Ko MHK, Black KL, 35. More SS, Vince R. Hyperspectral imaging signatures detect amyloidopathy
et al. Identification of amyloid plaques in retinas from Alzheimer’s patients in Alzheimers mouse retina well before onset of cognitive decline. ACS
and noninvasive in vivo optical imaging of retinal plaques in a mouse Chem Neurosci. 2015. https://doi.org/10.1021/cn500242z.
model. Neuroimage. 2011. https://doi.org/10.1016/j.neuroimage.2010.06.020. 36. Hadoux X, Hui F, Lim JKH, Masters CL, Pébay A, Chevalier S, et al. Non-
15. Kirbas S, Turkyilmaz K, Anlar O, Tufekci A, Durmus M. Retinal nerve fiber invasive in vivo hyperspectral imaging of the retina for potential biomarker
layer thickness in patients with Alzheimer disease. J Neuro-Ophthalmology. use in Alzheimer’s disease. Nat Commun. 2019. https://doi.org/10.1038/
2013. https://doi.org/10.1097/WNO.0b013e318267fd5f. s41467-019-12242-1.
16. Lu Y, Li Z, Zhang X, Ming B, Jia J, Wang R, et al. Retinal nerve fiber layer 37. Adamczuk K, Schaeverbeke J, Vanderstichele HMJ, Lilja J, Nelissen N, Van
structure abnormalities in early Alzheimer’s disease: evidence in optical Laere K, et al. Diagnostic value of cerebrospinal fluid Aβ ratios in preclinical
coherence tomography. Neurosci Lett. 2010. https://doi.org/10.1016/j.neulet. Alzheimer’s disease. Alzheimers Res Ther. 2015. https://doi.org/10.1186/
2010.06.006. s13195-015-0159-5.
17. Bambo MP, Garcia-Martin E, Pinilla J, Herrero R, Satue M, Otin S, et al. 38. McKhann GM, Knopman DS, Chertkow H, Hyman BT, Jack CR Jr, Kawas CH,
Detection of retinal nerve fiber layer degeneration in patients with Klunk WE, Koroshetz WJ, Manly JJ, Mayeux R, Mohs RC. The diagnosis of
Alzheimer’s disease using optical coherence tomography: searching new dementia due to Alzheimer’s disease: recommendations from the National
biomarkers. Acta Ophthalmol. 2014. https://doi.org/10.1111/aos.12374. Institute on Aging-Alzheimer’s Association workgroups on diagnostic
Lemmens et al. Alzheimer's Research & Therapy (2020) 12:144 Page 13 of 13

guidelines for Alzheimer’s disease. Alzheimers Dement. 2011. https://doi. angiography findings. JAMA Ophthalmol. 2018. https://doi.org/10.1001/
org/10.1016/j.jalz.2011.03.005. jamaophthalmol.2018.3556.
39. Nobili F, Arbizu J, Bouwman F, Drzezga A, Agosta F, Nestor P, et al. 59. Yoon SP, Grewal DS, Thompson AC, Polascik BW, Dunn C, Burke JR, et al.
European Association of Nuclear Medicine and European Academy of Retinal microvascular and neurodegenerative changes in Alzheimer’s
Neurology recommendations for the use of brain 18 F-fluorodeoxyglucose disease and mild cognitive impairment compared with control participants.
positron emission tomography in neurodegenerative cognitive impairment Ophthalmol Retin. 2019. https://doi.org/10.1016/j.oret.2019.02.002.
and dementia: Delphi consensus. Eur J Neurol. 2018. https://doi.org/10. 60. Van De Kreeke JA, Nguyen HT, Konijnenberg E, Tomassen J, Den Braber A,
1111/ene.13728. Ten Kate M, et al. Optical coherence tomography angiography in preclinical
40. Lambrechts A, Gonzalez P, Geelen B, Soussan P, Tack K, Jayapala M. A Alzheimer’s disease. Br J Ophthalmol. 2019. https://doi.org/10.1136/
CMOS-compatible, integrated approach to hyper- and multispectral bjophthalmol-2019-314127.
imaging. In: Technical Digest - International Electron Devices Meeting, IEDM. 61. Berisha F, Feke GT, Trempe CL, McMeel JW, Schepens CL. Retinal
2015 doi:https://doi.org/10.1109/IEDM.2014.7047025. abnormalities in early Alzheimer’s disease. Investig Ophthalmol Vis Sci. 2007.
41. Geelen B, Tack N, Lambrechts A. A compact snapshot multispectral imager https://doi.org/10.1167/iovs.06-1029.
with a monolithically integrated per-pixel filter mosaic. In: Advanced 62. Szegedi S, Dal-Bianco P, Stögmann E, Traub-Weidinger T, Rainer M,
fabrication technologies for micro/nano optics and photonics VII 2014 doi: Masching A, et al. Anatomical and functional changes in the retina in
https://doi.org/10.1117/12.2037607. patients with Alzheimer’s disease and mild cognitive impairment. Acta
42. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, et al. Ophthalmol. 2020. https://doi.org/10.1111/aos.14419.
Scikit-learn: machine learning in Python. J Mach Learn Res. 2011.
43. Varma S, Simon R. Bias in error estimation when using cross-validation for
model selection. BMC Bioinformatics. 2006. https://doi.org/10.1186/1471-
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in
2105-7-91.
published maps and institutional affiliations.
44. Seibold LK, Mandava N, Kahook MY. Comparison of retinal nerve fiber layer
thickness in normal eyes using time-domain and spectral-domain optical
coherence tomography. Am J Ophthalmol. 2010. https://doi.org/10.1016/j.
ajo.2010.06.024.
45. Quigley HA. Glaucoma. In: The Lancet 2011 doi:https://doi.org/10.1016/
S0140-6736(10)61423-7.
46. Chan VTT, Sun Z, Tang S, Chen LJ, Wong A, Tham CC, et al. Spectral-domain
OCT measurements in Alzheimer’s disease: a systematic review and meta-
analysis. Ophthalmology. 2019. https://doi.org/10.1016/j.ophtha.2018.08.009.
47. Shi Z, Wu Y, Wang M, Cao J, Feng W, Cheng Y, et al. Greater attenuation of
retinal nerve fiber layer thickness in Alzheimer’s disease patients. J
Alzheimers Dis. 2014. https://doi.org/10.3233/JAD-131898.
48. Trebbastoni A, D’Antonio F, Bruscolini A, Marcelli M, Cecere M, Campanelli
A, et al. Retinal nerve fibre layer thickness changes in Alzheimer’s disease:
results from a 12-month prospective case series. Neurosci Lett. 2016. https://
doi.org/10.1016/j.neulet.2016.07.006.
49. Sharafi SM, Sylvestre JP, Chevrefils C, Soucy JP, Beaulieu S, Pascoal TA, et al.
Vascular retinal biomarkers improves the detection of the likely cerebral
amyloid status from hyperspectral retinal images. Alzheimer’s Dement Transl
Res Clin Interv. 2019. https://doi.org/10.1016/j.trci.2019.09.006.
50. Chung HS, Harris A, Halter PJ, Kagemann L, Roff EJ, Garzozi HJ et al. Regional
differences in retinal vascular reactivity. Investig Ophthalmol Vis Sci 1999.
51. Koronyo Y, Biggs D, Barron E, Boyer DS, Pearlman JA, Au WJ, et al. Retinal
amyloid pathology and proof-of-concept imaging trial in Alzheimer’s
disease. JCI insight. 2017. https://doi.org/10.1172/jci.insight.93621.
52. Shi H, Koronyo Y, Rentsendorj A, Regis GC, Sheyn J, Fuchs DT, et al.
Identification of early pericyte loss and vascular amyloidosis in
Alzheimer’s disease retina. Acta Neuropathol. 2020. https://doi.org/10.
1007/s00401-020-02134-w.
53. Schön C, Hoffmann NA, Ochs SM, Burgold S, Filser S, Steinbach S, et al.
Long-term in vivo imaging of fibrillar tau in the retina of P301S transgenic
mice. PLoS One. 2012. https://doi.org/10.1371/journal.pone.0053547.
54. Ho CY, Troncoso JC, Knox D, Stark W, Eberhart CG. Beta-amyloid, phospho-
tau and alpha-synuclein deposits similar to those in the brain are not
identified in the eyes of Alzheimer’s and Parkinson’s disease patients. Brain
Pathol. 2014. https://doi.org/10.1111/bpa.12070.
55. Williams EA, McGuone D, Frosch MP, Hyman BT, Laver N, Stemmer-
Rachamimov A. Absence of Alzheimer disease neuropathologic changes in
eyes of subjects with Alzheimer disease. J Neuropathol Exp Neurol. 2017.
https://doi.org/10.1093/jnen/nlx020.
56. den Haan J, Morrema THJ, Verbraak FD, de Boer JF, Scheltens P, Rozemuller
AJ, et al. Amyloid-beta and phosphorylated tau in post-mortem Alzheimer’s
disease retinas. Acta Neuropathol Commun. 2018. https://doi.org/10.1186/
s40478-018-0650-x.
57. Grewal DS, Polascik BW, Hoffmeyer GC, Fekrat S. Assessment of differences
in retinal microvasculature using OCT angiography in Alzheimer’s disease: a
twin discordance report. Ophthalmic Surg Lasers Imaging Retin. 2018.
https://doi.org/10.3928/23258160-20180601-09.
58. O’Bryhim BE, Apte RS, Kung N, Coble D, Van Stavern GP. Association of
preclinical Alzheimer disease with optical coherence tomographic