Alshareef et al.

Int J Retin Vitr (2017) 3:25

DOI 10.1186/s40942-017-0078-7 International Journal
of Retina and Vitreous

ORIGINAL ARTICLE Open Access

Segmentation errors in macular

ganglion cell analysis as determined by optical
coherence tomography in eyes with macular
pathology
Rayan A. Alshareef1,2, Abhilash Goud3, Mikel Mikhail1, Hady Saheb1, Hari Kumar Peguda3, Sunila Dumpala3,
Shruthi Rapole3 and Jay Chhablani3*

Abstract
Background: To evaluate artifacts in macular ganglion cell inner plexiform layer (GCIPL) thickness measurement in
eyes with retinal pathology using spectral-domain optical coherence tomography (SD OCT).
Methods: Retrospective analysis of color-coded maps, infrared images and 128 horizontal B-scans (acquired in the
macular ganglion cell inner plexiform layer scans), using the Cirrus HD-OCT (Carl Zeiss Meditec, Dublin, CA). The study
population included 105 eyes with various macular conditions compared to 30 eyes of 30 age-matched healthy
volunteers. The overall frequency of image artifacts and the relative frequency of artifacts were stratified by macular
disease.
Results: Scan errors and artifacts were found in 55.1% of the 13,440 B-scans in eyes with macular pathology and
26.8% of the 3840 scans in normal eyes. Segmentation errors were the most common scan error in both groups, with
more common involvement of both segmentation borders in diseased eyes and anterior segmentation border in
normal eyes.
Conclusion: Segmentation errors and artifacts in SD OCT GCA are common in conditions involving the macula.
These findings should be considered when assessing macular GCIPL thickness and careful assessment of scans is
suggested.
Keywords: Artifacts, Errors, Ganglion cell algorithm, Ganglion cell inner plexiform layer, Spectral domain optical
coherence tomography

Background visualization of distinct retinal layers has become fea-

Since the introduction of optical coherence tomogra- sible in vivo and has been shown to correlate well with
phy (OCT), OCT has become an integral part of the histology [1–3]. This has numerous advantages; fore-
ophthalmological equipment, and in particular, for most, quantitative assessment of retinal layers over time
monitoring purposes. The selection of various imaging facilitates longitudinal assessment of pathological pro-
modes enables a broad variety of clinical applications. cesses and may become the standard for the assessment
With the recent introduction of high-resolution spectral of effects of novel therapeutic substances. Several OCT
domain (SD)-OCT and the adaptation of this technique, scan modes are used to evaluate different aspects of the
retinal pathology.

­ irrus® OCT (Carl Zeiss Meditec, Inc., Dub-

In recent years, the ganglion cell analysis (GCA) algo-
*Correspondence: [email protected] rithm on C
3
Smt. Kanuri Santhamma Retina Vitreous Centre, L. V. Prasad Eye Institute, lin, CA) has received particular attention and is gaining
Kallam Anji Reddy Campus, Banjara Hills, Hyderabad 500 034, India
Full list of author information is available at the end of the article momentum as an important diagnostic tool in conditions

© The Author(s) 2017. This article is distributed under the terms of the Creative Commons Attribution 4.0 International License
(http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium,
provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license,
and indicate if changes were made. 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.
Alshareef et al. Int J Retin Vitr (2017) 3:25 Page 2 of 8

­Cirrus® OCT GCA algorithm mode uses segmentation

involving the optic nerve and macula. The onboard included. Inclusion criteria for this group were dis-
eased eyes (as previously mentioned) with good quality
software for automated detection of the outer bound- scans (defined as a signal strength of more than 6), and
ary of the retinal nerve fiber layer (RNFL) and the outer a refractive error between −6.00 diopters (D) and +3.00
boundary of the inner plexiform layer (IPL) and pro- D spherical equivalent. The control group included 30
vides measurements of ganglion cell inner plexiform eyes of 30 healthy patients without any retinal or vitreo-
layer (GCIPL) thickness, thereby allowing in vivo quan- retinal interface abnormalities, with a good quality scan
titative measurement of inner retinal layer thickness. It (signal strength of more than 6) and a refractive error
has emerged as a useful method for research and clini- between −6D and +3D [19]. One eye of each patient was
cal practice, with applications spanning the evaluation included in the study. Subjects were excluded if they had
of glaucomatous damage and progression, to indicating coexisting ocular disease (e.g. uveitis, glaucoma or non
retinal neurodegeneration and being a biomarker for glaucomatous optic neuropathy), a history of intraocular
structural changes overtime [4–8]. Numerous reports surgery, myopia >−6.00 D, hyperopia >+3.00 D or signif-
have evaluated how retinal boundary artifacts can over- icant media opacities.
estimate or underestimate macular thickness [2, 3, 9–18].
The occurrence of these artifacts may result in erroneous OCT image acquisition and processing

­Cirrus® HD-OCT after pupillary dilation. The Macular

measurements of inner retinal structures. Recently we Spectral domain OCT scans were obtained by using the
reported a 26.8% prevalence rate of artifacts in a normal
population [19]. The prevalence, features and associated Cube 512 × 128 scan protocol was used for all subjects.
factors of artifacts involving the GCIPL in patients with The protocol performs 128 B-scans and 512 A-scans
macular involved retinal disease has not been previously per B-scan over 1024 samplings within a cube measur-
evaluated, and no study has examined whether there is an ing 6 × 6 × 2 mm centered on the fovea. Images with
alteration in the frequency of artifacts as the architecture signal strength <6 were considered of poor quality and
of the retina changes in various macular diseases. It is discarded.
thus beneficial to identify those scenarios in which arti- As described in our previous publications, the GCA
facts and errors may occur, endangering accurate quan- algorithm was applied to the Macular Cube scans 9, 10
tification and diagnosis in patients with conditions that Briefly, the GCA algorithm identifies the outer bound-
impact the architecture of the macula. ary of the RNFL and the outer boundary of the IPL and
The aim of this study was to assess the prevalence and provides measurements of GCIPL thickness. The average,
magnitude of artifacts and errors in GCIPL segmenta- minimum (lowest GCIPL thickness over a single merid-
tion in various pathologic entities of the macula and ian crossing the annulus), and sectoral (superotemporal,
to compare those with scans obtained from a healthy superior, superonasal, inferonasal, inferior, inferotem-
population. poral) GCIPL thicknesses were measured in an ellipti-
cal annulus around the fovea (dimensions: vertical inner
Methods and outer radius of 0.5 and 2.0 mm, horizontal inner
Spectral-domain OCT (SD-OCT) data sets were gath- and outer radius of 0.6 and 2.4 mm, respectively). The
ered retrospectively from the patient cohort imaged GCA algorithm measures the mean GCIPL thickness for
with the GCA algorithm at L. V. Prasad Eye Institute, each sector, compares them to the internal normative
Hyderabad, India between January 2013 and July 2015. data- base of the device, and generates a thickness map,
The local ethics committee approved the study and all a deviation map, and a color-coded significance map.
patients gave informed consent prior to obtaining imag- Measurements were displayed in green for normal range
ing. The macular scans were obtained from consecutive (P = 5%–95%), in yellow for borderline (1% < P < 5%),
patients representing a spectrum of diseases involving and in red for outside the normal range (P < 1%).
the macula with no prior history of treatment. The study

SD-OCT scans were obtained by using the ­Cirrus® HD-

was conducted in accordance with the ethical standards Assessment of image artifacts
stated in the 1964 Declaration of Helsinki.
One hundred and five eyes of 105 subjects with various OCT (Carl Zeiss Meditech, Dublin, CA) after pupillary
macular conditions, including 15 eyes each presenting dilation. The Macular Cube 512 × 128 scan protocol was
with diabetic macular edema (DME) secondary to type used for all subjects. The protocol performs 512 horizon-
II diabetes, central serous chorioretinopathy (CSCR), tal B-scans comprising 200 A-scan per B-scan over 1024
idiopathic epiretinal membrane (ERM), dry age-related samplings within a cube measuring 6 × 6 × 2 mm cen-
macular degeneration (AMD) and wet AMD, as well as tered on the fovea. Images with signal strength <6 were
an additional 30 eyes of retinitis pigmentosa (RP), were considered of poor quality and discarded.
Alshareef et al. Int J Retin Vitr (2017) 3:25 Page 3 of 8

All 128 horizontal OCT B-scans acquired in the macu- Each error was further described according to its loca-
lar cube 512 × 128 protocol were examined by one evalu- tion on the scan. Each scan was divided into central
ator (AG) with an intraclass correlation of 0.93 for the 1000 μm, temporal 2500 μm and nasal 2500 μm. Seg-
presence of artifacts. mentation error affecting the central part of the scan was
Analysis of scans was performed as described in specifically noted, as thickness measures from this sub-
our previous publication [19]. Analysis of the images field are commonly used to guide retreatment in patients
included infrared image, color-coded map and all indi- with nAMD and DME. In addition, the central area of
vidual 128 scans in an eye. Each B scan of the volume was line scans may have a higher frequency of segmentation
reviewed to identify the potential cause of an artifact pre- errors than more peripheral areas, as most patients might
sent in that B scan. Any form of segmentation error or have fovea involving lesions with more disruption of the
artifacts was noted. Artifacts were identified and classi- retinal ganglion cell RGC complex in this central zone.
fied, and the number of B-scans containing artifacts was Furthermore, the total scan area was divided in three
noted for each artifact type. zones: upper zone included 1st to 41st scan, central zone
included 42nd to 84th scan and lower zone included 85th
Boundary line errors to 128th scan. Overall occurrence of the artifacts was
Inner/outer segmentation line misidentification described in these zones.
The GCA algorithm software locates the outer boundary Motion artifact was detected as misalignment of reti-
of the retinal nerve fiber layer (RNFL) presented as a solid nal vessels on rendered fundus infrared image. Figure 1
purple line and the outer boundary of the inner plexiform shows various artifacts in macular conditions. Other arti-
layer (IPL) presented as a solid yellow line. Subsequently, facts including defocus, out of register, shadowing, mir-
all images acquired within each scan pattern (for exam- ror artifact, and blink artifact were also evaluated.
ple, 128 raster scans in Cirrus macular cube 512 × 128)
were reviewed to check for segmentation breakdown in Statistical analysis
the inner and outer segmentation lines. The presence of a Descriptive statistics included mean and standard devia-
macular GCIPL segmentation error was defined as when tion. Statistical analyses were performed using commer-
these 2 lines were not located in the proper positions cial software (Stata data analysis and statistical software,
between the retinal layers in at least 1 cross- sectional version 12.1, StataCorp, College Station, TX). A P value
image. If segmentation algorithm abnormality was noted of <0.05 was considered statistically significant.
in the outer boundary of the RNFL, it was recorded as
inner segmentation line misidentification. If a segmen- Results
tation abnormality was noted in the outer boundary of Eyes with macular related pathologies
the inner plexiform layer (IPL), it was recorded as outer The various retinal pathologies included were as follows:
segmentation line misidentification. If a segmentation 30 eyes with RP, 15 eyes with wet AMD, 15 eyes with idi-
abnormality occurred in both lines in a single B scan, it opathic ERM, 15 eyes with DME secondary to type II dia-
was only recorded once as a segmentation error involving betes, 15 eyes with dry AMD, 15 eyes with CSCR. The
both borders of the GCIPL. mean age of the subjects was 49.54 ± 16.32 years (range
15–85). Gender distribution was 31 females and 74
Missing segmentation line males. Thirteen eyes were pseudophakic.
Inner segmentation line absence was recorded in cases
where a visible absence of the outer boundary of the Overall frequency of scan line artifacts
RNFL line existed. In cases where an outer boundary of A total of 13,440 scans from 105 OCT macular cube
the inner plexiform layer (IPL) was missing, an outer seg- scans of 105 eyes were reviewed. Artifacts were noted
mentation line absence was recorded. Absence of both in 7410 scans of 13,440 scans, equivalent to a frequency
lines bordering the GCIPL was recorded as well. of 55.1%. Artifacts were mostly observed in B scans of
patients with RP 95.5% followed by neovascular AMD
Severity parameters related to boundary errors 64.47% and least in patients with ERM 19.7%.
To grade the severity of segmentation error, each scan
was noted for segmentation deviation (inner or outer or Frequency of artifacts by type of artifacts
both). Deviation of the segmentation line was classified For ganglion cell algorithm scans, misidentification of
into mild (<10 μm), moderate (10–50 μm) and severe both lines was the most common segmentation line
(more than 50 μm). Each deviation, if present, was noted error (50%) observed. Segmentation line absence was
as upward or downward deviation. the second most common segmentation line error (7.8%)
Alshareef et al. Int J Retin Vitr (2017) 3:25 Page 4 of 8

Fig. 1 Composite figure shows various artifacts on ganglion cell analysis of ­Cirrus® spectral domain optical coherence tomography. a Correct seg-
mentation of ganglion cell inner plexiform layer in an eyes central serous chorioretinopathy (CSCR); b blink artifact on infrared image (arrow shows
the missing information due to blink artifact); c outer segmentation artifact due to hard exudate (arrow) in diabetic macular edema; d misalignment
of both segmentation lines along with mirror artifact in an eye with retinitis pigmentosa; e downward deviation of both segmentation lines; f outer
segmentation artifact in an eye with diabetic macular edema; misalignment of both segmentation lines along with mirror artifact in an eye with
scarred age-related macular degeneration (AMD) (g); CSCR (h); dry AMD (i)

observed, and it more likely involved the inner segmenta- errors were located in the nasal 2500 μm area of the scan.
tion line (7.8%) Motion artifacts were the most common When reviewing cross-sectional scans with errors, loca-
artifacts noted on IR images. Table 1 summarizes types of tion of errors was not correlated with location of pathol-
artifacts observed. ogy. Upward Deviation of the segmentation line occurred
Percentage of B scans with inner and outer GCIPL more commonly then downward segmentation line devia-
misidentification was calculated for each of the disease tion. Upward deviation occurred more in RP (92.1%) and
states. Highest average percentage of an isolated outer dry AMD 93.9%. Downward deviation occurred more fre-
GCIPL misidentification occurred in DME (20.8%), and quently in DME 26.35% and RP (23.5%). Among all disease
CSCR (12.2%), whereas highest average percentage of categories, ERM had the least amount of deviations. When
an isolated inner retina misidentification occurred in comparing the degree of segmentation line deviation, mild
wet AMD (30%), and dry AMD (15.4%). Furthermore, deviation was more often noted in all disease groups.
the highest average percentage of B scans with com-
bined inner and outer GCIPL boundary misidentification Causes of artifacts in disease categories (Table 2)
occurred in RP (99.5%) and CSCR (86.9%). Retinitis pigmentosa
When comparing the proportion of cross-sectional ret- The majority of errors observed was misidentification
ina scans with errors, nearly all artifact types were more of layers secondary to hyper-reflectivity of the RNFL
common in the center 1000 micron area of the scan in all 92.36%, where the algorithm misinterpreted the highly
disease categories but RP and ERM, where most of the reflective RNFL as the inner border of the GCIPL, other
Alshareef et al. Int J Retin Vitr (2017) 3:25 Page 5 of 8

Table 1 Demographic variables and percentage and characteristics of segmentation errors identified using the ganglion
cell algorithm analysis in macular diseases and healthy eyes
Healthy eyes RP CSCR Dry AMD DME Wet AMD ERM

N (eyes) 30 30 15 15 15 15 15
Age (years) 56.3 ± 4.5 31.1 ± 11.81 37.6 ± 9.74 62.49 ± 11.49 68.0 ± 7.7 58.0 ± 14.2 53.6 ± 15.4
Mean signal strength 6.76 ± 0.81 8.36 8.17 6.60 6.80 6.46 6.86
Mean number of 35.84 122.9 52.6 20.5 61.5 75.8 25.26
scans with error in one eye
Mean overall 26.8 95.5 41.0 20.5 48.0 64.47 19.7
number of scans with errors (%)
Affected zone (%)
Upper (1st–41st) 20 31.27 9.27 6.92 14.16 18.54 6.30
Central (42nd–84th) 30 31.32 16.04 9.01 18.33 25.98 5.20
Lower (85th–125th) 47 33.43 15.78 4.58 15.62 19.94 8.22
Affected scan area (%)
Nasal 2500 μm 11.2 94.27 43.07 39.33 56.39 45.29 88.65
Central 16.8 86.06 79.48 46.15 77.42 85.02 88.15
Temporal 2500 μm 11.4 91.40 75.78 37.52 58.56 42.66 77.00
Layer affected (%)
Inner 62 1.0 0.76 15.48 0 8.64 30.07
Outer 17.2 0.02 12.29 3.55 20.82 6.70 0
Both 20.79 99.59 86.94 80.96 78.95 84.65 69.92
Deviation (%)
Upward 47.91 92.16 85.93 93.90 73.86 84.65 66.49
Downward 33.43 23.56 17.36 6.09 26.35 15.58 0
Both 18.65 2.52 0 0 0 0 33.50
Degree of deviation (%)
Mild 38.6 77.81 88.97 94.41 98.59 94.91 96.04
Moderate 20.9 14.31 11.02 3.80 2.49 4.52 3.95
Severe 40.3 9.49 0 2.02 0 0.56 0
Type (%)
Segmentation error 20.79 95.8 40.0 20.5 48.0 58.0 16
Segmentation loss 2.91 0.0.78 1.1 0.8 1 1.2 2
Defocus 6.70 4.4 0 0 0 0 0
Out-of-register 3.3 10.9 0 0 0 0 0
Shadowing 0 0.4 0 0 0 0 0
Mirror artifact 0 4.5 0 0 0 0 0
Blink artifact 0.09 0.4 0 0 0 0 0
Artifacts on IR image (%)
Motion 53.3 33.33 6.6 20 26 23 6.6
Blink 0.09 6.66 0 0 0 0 0
RP retinitis pigmentosa, CSCR central serous chorioretinopathy, AMD age-related macular degeneration, DME diabetic macular edema, ERM epiretinal membrane

causes included defocus errors 6.47%, floater 0.46% and ERM

out of register scans 5.10%. Again, high reflectivity of the RNFL was identified as the
cause of artifacts in 15.1% scans. The epiretinal mem-
CSCR brane itself was the cause for artifacts in 3.6% scans.
Hyper-reflectivity of the RNFL was also the most com-
mon identifiable element causing artifacts in 20% of Dry AMD
scans. Furthermore, 12% were caused by subretinal fluid, Pigment epithelial detachment was a cause in 14% of
12% as a result of pigment epithelial detachments (PED) scans. Degraded image was also observed in 3.48%
while 5.7% were other morphological elements. scans.
Alshareef et al. Int J Retin Vitr (2017) 3:25 Page 6 of 8

Table 2 Various identifiable causes of segmentation errors in various macular conditions

RNFL hyper-reflectivity SRF PED Floater Degraded Inner retinal Hard exudates
image thickening

CSCR 20% 12% 5.7% 0 0 0 0

RP 92.36% 0 0 0.46% 0 0 0
ERM 15.10% 0 0 0 0 0 0
Dry AMD 3.95% 0 14.0% 0 3.48% 0 0
DME 10.67% 0 0 0 0 27.86% 4.42%
Wet AMD 12.5% 11.45% 28.69% 0 0 6.14% 0
RP retinitis pigmentosa, CSCR central serous chorioretinopathy, AMD age-related macular degeneration, DME diabetic macular edema, ERM epiretinal membrane, PED
pigment epithelial detachment, SRF subretinal fluid

DME When compared to normal eyes [19], our results indi-

Among DME eyes, inner retinal thickening was the cause cate that the accuracy of macular GCIPL thickness meas-
in 27.86% scans, whereas hyper-reflectivity of the RNFL urements may be largely influenced by the presence and
and hard exudates were morphological elements that severity of macular disorders. In contrast to OCT technol-
resulted in artifacts in 10.67 and 4.42% scans respectively. ogy development which has been a field of active research
since 1991, OCT image segmentation has only been more
Neovascular AMD fully explored during the last decade. Segmentation, how-
Identifiable causes included RNFL hyper reflectivity in ever, remains one of the most difficult and at the same time
12.5% scans, subretinal fluid in 11.45% scans, PED in most commonly required steps in OCT image analysis. No
28.69% scans and inner retinal thickening in 6.14% scans. typical segmentation method exists that can be expected
to work equally well for all tasks [2]. In this paper, we tried
Normal healthy eyes to cite most related works from 1997 to 2012, however,
We reported these results in our recent publication [19]. this is in no way complete. It should also be noticed that
Briefly, A total of 1029 (26.8%) out of total 3840 scans had the number as reported in tables cannot be used for direct
scan errors. Misidentification of the inner GCIPL bound- comparison of the relative performances, since different
ary was most frequent (62%). Upward Deviation of the
­ irrus®, the measurement of inner retinal
settings are utilized in each method.
segmentation line (47.91%) and severe deviation (40.3%) With the Zeiss C
were more often noted. Artifacts were mostly located thickness is determined by inbuilt protocols of measure-
in the central scan area (16.8%). The average number of ments. Although different machines may have a different
scans with errors per eye was 34.3% and was not related number of artifacts, it appears that SD-OCT has reduced
to signal strength on Spearman correlation (P = 0.36)
­ irrus®
errors in automated retinal delineation. It has been
(Table 1). reported that a low frequency of errors on the C
OCT compared to the frequency of errors of other OCT
Discussion instruments when evaluating patients with retinal disease
In ophthalmic imaging, the results of image segmenta- [2, 16]. Errors linked to post-image processing and OCT
tion can assist with locating layer pathologies, measuring instrument software frequently were observed and most
tissue and layer volumes, studying anatomical structure commonly involved misidentification of the retina layers.
and diagnosing several disorders. For example, retinal tis- More convincingly, Hwang et al. evaluated GCA related
sue segmentation for delineation of anatomical structures artifacts in glaucomatous and normal eyes without macular
from OCT images plays an important role in several sce- abnormalities and identified segmentation errors in 9.7%
narios, this is particularly relevant in the evaluation of of eyes. In their study, the most prominent features of seg-
neurodegenerative disorders [20], where one can charac- mentation errors were: (1) they affected both the inner and
terize morphological differences between subjects based outer segmentation lines (2) they were located at the nasal
on volumetric analysis of GCIPL, RNFL and outer reti- quadrant (centered to the fovea), (3) they were diffuse. Fur-
nal layers [3, 9, 11]. However, these are valid only if the thermore, the presence of a segmentation error was asso-
results of image segmentation are correct. Image artifacts ciated with a higher degree of myopia. In contrast to their
such as the presence of motion, retinal layer misidenti- study, and to shed some light on the impact of macular
fication, can cause classification errors in the results of diseases on segmentation errors we strictly included eyes
image segmentation. As a result, image segmentation is with macular disorders. While evaluating the RNFL and
still a challenging task in ophthalmic image processing. GCIPL of several diseased eyes, we noted misidentification
Alshareef et al. Int J Retin Vitr (2017) 3:25 Page 7 of 8

artifacts involving the inner and outer GCIPL segmenta- be related to the macular involvement, which may cause
tion line. Overall, misidentification of both boundary lines a higher frequency of artifacts.
of the GCIPL were most common (50%), followed by outer The vast majority of patients exhibited inner segmenta-
boundary misidentification, which occured in 2.9% of eyes. tion error, and this was common across all pathologies.
Recent studies have investigated the macular altera- Similarly, deviation was mostly in the upward direction
tions associated with both dry and wet AMD at the level rather than in the downward direction. Outer border
of the inner retinal layers using OCT devices [21–23]. misidentification is likely to be higher in eyes presenting
Zucchiatti et al. [22] reported that macular GCIPL in pathology in the outer retinal layer, such as subreti-

types of AMD using ­Cirrus® OCT. They also theorized

involvement displayed by thinning is present in both nal fluid in CSCR or wet AMD, as well as drusen in dry
AMD. These can have significant clinical implications,
that inner retinal morphology alteration might be help- particularly in the follow-up of those patients. The cur-
ful when considering treatment options. For this reason, rent GCA algorithm does now allow manual correction
artifacts and errors might be of additional value to inner of inner and outer border misidentifications.
retinal topographic evaluation and decision-making The presence of morphologic features, such as fluid
strategies. Ho et al. compared various OCT devices in between the retina and the RPE or fluid within the retina,

­Cirrus® HD-OCT has the lowest percentages of any type

patients with various retinal conditions and showed that may confound the segmentation algorithms that attempt
to identify the inner retinal boundaries. In contrast, eyes
of artifacts for patients with wet AMD [2]. In our study, with a diagnosis of dry AMD with no morphological ele-
more scan artifacts were observed in eyes with wet AMD ments were less likely to have errors in retinal segmenta-

Cirrus® OCT, Ho et al. [2] examined 15

compared to dry AMD, 64.4 and 20.5%, respectively. tion. This observation increases confidence in the use of
Using the ­ quantitative OCT data in studies of dry AMD, but raises
patients with ERM and reported improper central thick- concerns regarding its use in the examination of neovas-
ness measurements (10%) secondary to artifacts. Using cular AMD, DME and RP. Furthermore, the impact of
the GCA algorithm in 15 eyes with idiopathic ERM in these conditions on GCIPL measurement should be con-
our study, errors occurred in 19.7%. Additionally, in their sidered when interpreting the GCA in glaucoma patients
study, eight patients had DME and diabetic retinopathy with macular co-morbidities.
of which 25% of B scans were with outer retina misiden- Recently published OCT segmentation reports focus on
tification and 23% of B scans with inner retina misiden- improving precision and accuracy, reducing the required
tification [2]. This is higher than what we observed in processing time and increasing the understanding of the
our study where none of the B scans had isolated inner various errors associated with these segmentation algo-
GCIPL segmentation line misidentification (boundary rithms [13, 21, 24–26]. Current imaging modalities are
line artifact), and 20.8% of B scans showed isolated outer moving research in the direction of layer and volume seg-
GCIPL misidentification. However, B scans with inner mentation, as well as 3-D rendering and visualization. It
and outer GCIPL boundary misidentification were seen is therefore important to develop robust methods that are
in 78.9% of B scans. The higher number of patients in our capable of dealing with pathologic cases in OCT imaging.
study could explain this. The frequency of GCIPL bound- The limitations of this study include a small sample
ary artifacts in eyes with DME was 48% in our study. Fur- size in each group. The inaccuracy of the macular GCIPL
thermore, the segmentation algorithm misinterpreted thickness could not be assessed because the software
the hyper-reflective intraretinal areas representing hard does not permit manual correction when an artifact
exudates to correspond to the outer IPL boundary, miss- is detected. We did not longitudinally assess if the fre-
ing correct alignment of the outer IPL border. This led to quency of these artifacts change with time or in response
the wrong estimation of the GCIPL thickness, as calcu- to treatment. Although this study examined eyes across
lated by the software. a broad spectrum of macular disorders, the results
Our results show that RP patients exhibited the great- obtained from this study can only be generalized to this
est mean number of errors across all types of pathology, subset of patients undergoing this specific imaging pro-
despite a better mean signal strength (8.36) when com- tocol. We did not evaluate effect of vision, fixation, or the
pared to other conditions. This could be related to fixa- grade of cataract on artifacts. Because of the small sam-

­ irrus® Cube 512 × 128 and

tion difficulties and pan-layer architecture changes in the ple size in each disease, valuable correlation between ret-
retina. Ho et al. used the C inal thickness, signal strength and GCIPL thickness could
reported that Stargardt disease caused the greatest per- not be studied. We did not evaluate the effect of repeat
centage of significant OCT examination error. Although scans on artifacts. Nevertheless, to the best of our knowl-
the pathophysiology of RP is different from Stargardt dis- edge, this is the only study reporting the frequency, type
ease, the progressive loss of vision in both diseases may and distribution of artifacts, as well as the morphological
Alshareef et al. Int J Retin Vitr (2017) 3:25 Page 8 of 8

elements that may lead to artifacts using the GCA algo- 2. Ho J, et al. Assessment of artifacts and reproducibility across spectral- and
time-domain optical coherence tomography devices. Ophthalmology.
rithm. Finally, we have not evaluated the effect of repeat 2009;116(10):1960–70.
scans on artifacts. 3. Alshareef RA, et al. In vivo evaluation of retinal ganglion cells degen-
In conclusion, the GCA algorithm provides a signifi- eration in eyes with branch retinal vein occlusion. Br J Ophthalmol.
2016;100:1506–10.
cant advance in our ability to image and assess the inner 4. Hwang YH, et al. Macular ganglion cell analysis for early detection of
retina. Nevertheless, there are multiple errors associated glaucoma. Ophthalmology. 2014;121(8):1508–15.
with the segmentation and identification of retinal lay- 5. Mwanza JC, et al. Macular ganglion cell-inner plexiform layer: auto-
mated detection and thickness reproducibility with spectral domain-
ers. In addition, multiple artifacts can be noted both in optical coherence tomography in glaucoma. Invest Ophthalmol Vis Sci.
pathological and normal eyes. Clinicians must pay atten- 2011;52(11):8323–9.
tion to these errors and factor them into their decision 6. Lee HJ, et al. Thickness of the macula, retinal nerve fiber layer, and
ganglion cell layer in the epiretinal membrane: the repeatability
making process. study of optical coherence tomography. Invest Ophthalmol Vis Sci.
2015;56(8):4554–9.
7. Chen Y, et al. Diabetic macular morphology changes may occur in the
Abbreviations early stage of diabetes. BMC Ophthalmol. 2016;16:12.
GCA: ganglion cell analysis; GCIPL: ganglion cell inner plexiform layer; RP: 8. Lim HB, et al. Prediction of retinal ischemia in branch retinal vein occlu-
retinitis pigmentosa; CSCR: central serous chorioretinopathy; AMD: age- sion: spectral-domain optical coherence tomography study. Invest
related macular degeneration; DME: diabetic macular edema; ERM: epiretinal Ophthalmol Vis Sci. 2015;56(11):6622–9.
membrane. 9. Chhablani J, et al. Retinal ganglion cells thinning in eyes with nonprolif-
erative idiopathic macular telangiectasia type 2A. Invest Ophthalmol Vis
Sci. 2015;56(2):1416–22.
Author details 10. Kupersmith MJ, et al. Retinal ganglion cell layer thinning within one
1
Department of Ophthalmology, McGill University, Montreal, QC, Canada. month of presentation for optic neuritis. Mult Scler. 2016;22(5):641–8.
2
Department of Ophthalmology, King Abdulaziz University, Jeddah, Kingdom 11. Chhablani J, et al. Neurodegeneration in type 2 diabetes: evidence from
of Saudi Arabia. 3 Smt. Kanuri Santhamma Retina Vitreous Centre, L. V. Prasad spectral-domain optical coherence tomography. Invest Ophthalmol Vis
Eye Institute, Kallam Anji Reddy Campus, Banjara Hills, Hyderabad 500 034, Sci. 2015;56(11):6333–8.
India. 12. Ray R, Stinnett SS, Jaffe GJ. Evaluation of image artifact produced by
optical coherence tomography of retinal pathology. Am J Ophthalmol.
Authors’ contributions 2005;139(1):18–29.
Conception and design: RA, JC, AG, HS; Data collection: AG, HKP, SD, SR; Data 13. Sadda SR, et al. Errors in retinal thickness measurements obtained by
Analysis: RA, HKP, AG, SD, SR; Writing: RA, HKP, AG, SD, SR, MM, HS; Critical optical coherence tomography. Ophthalmology. 2006;113(2):285–93.
Review of Manuscript: MM, RA, HS, JC. All authors read and approved the final 14. Domalpally A, et al. Quality issues in interpretation of optical coherence
manuscript. tomograms in macular diseases. Retina. 2009;29(6):775–81.
15. Patel PJ, et al. Segmentation error in Stratus optical coherence tomogra-
Acknowledgements phy for neovascular age-related macular degeneration. Invest Ophthal-
None. mol Vis Sci. 2009;50(1):399–404.
16. Giani A, et al. Artifacts in automatic retinal segmentation using different
Competing interests optical coherence tomography instruments. Retina. 2010;30(4):607–16.
The authors declare that they have no competing interests. 17. Han IC, Jaffe GJ. Evaluation of artifacts associated with macular
spectral-domain optical coherence tomography. Ophthalmology.
Availability of data and materials 2010;117(6):1177–89.
Available data and materials on request of the original data. Data of the con- 18. Song Y, et al. Overcoming segmentation errors in measurements of
trol group could be found at: “Prevalence and Distribution of Segmentation macular thickness made by spectral-domain optical coherence tomogra-
Errors in Macular Ganglion Cell Analysis of Healthy Eyes Using Cirrus HD-OCT.” phy. Retina. 2012;32(3):569–80.
https://www.ncbi.nlm.nih.gov/pubmed/27191396. 19. Alshareef RA, et al. Prevalence and distribution of segmentation errors in
macular ganglion cell analysis of healthy eyes using cirrus HD-OCT. PLoS
Consent for publication ONE. 2016;11(5):e0155319.
We consent to publish this paper in case of acceptance. 20. Kupersmith MJ, et al. Retinal ganglion cell layer thinning within one
month of presentation for optic neuritis. Mult Scler J. 2015;22:641–8.
Ethics approval and consent to participate 21. Hwang YH, Kim MK. Segmentation errors in macular ganglion cell analy-
Obtained and approved by the LV PRASAD IRB. All patients consented to this sis as determined by optical coherence tomography. Ophthalmology.
study. 2016;123(5):950–8.
22. Zucchiatti I, et al. Macular ganglion cell complex and retinal nerve fiber
layer comparison in different stages of age-related macular degenera-
Publisher’s note tion. Am J Ophthalmol. 2015;160(3):602–7.
Springer Nature remains neutral with regard to jurisdictional claims in pub-
23. Savastano MC, et al. Differential vulnerability of retinal layers to early
lished maps and institutional affiliations.
age-related macular degeneration: evidence by SD-OCT segmentation
analysispostreceptoral abnormalities in early AMD. Invest Ophthalmol Vis
Received: 14 January 2017 Accepted: 2 May 2017
Sci. 2014;55(1):560–6.
24. Chiu SJ, et al. Automatic segmentation of seven retinal layers in SDOCT
images congruent with expert manual segmentation. Opt Express.
2010;18:19413–28.
25. Ghorbel I, et al. Automated segmentationof macular layers in OCT
References images and quantitative evaluation of performances. Pattern Recognit.
1. Toth CA, et al. A comparison of retinal morphology viewed by optical 2011;44:1590–603.
coherence tomography and by light microscopy. Arch Ophthalmol. 26. Kafieh R, et al. A review of algorithms for segmentation of optical coher-
1997;115(11):1425–8. ence tomography from retina. J Med Signals Sens. 2013;3(1):45–60.