Pediatr Radiol (2013) 43:269–284

DOI 10.1007/s00247-012-2468-1

MINISYMPOSIUM

Technical developments in postprocessing of paediatric

airway imaging

Savvas Andronikou & Benjamin Irving &

Linda Tebogo Hlabangana & Tanyia Pillay & Paul Taylor &
Pierre Goussard & Robert Gie

Received: 15 May 2012 / Accepted: 9 July 2012

# Springer-Verlag 2012

Abstract CT postprocessing allows more scan information to Keywords Minimum intensity projections . Multiplanar
be viewed at one time allowing an accurate diagnosis to be reconstruction . Volume rendering . Computer aided
made more efficiently, and is particularly important in paediat- diagnosis . Paediatric
ric practice where invasive clinical diagnostic tools can be
replaced or at least assisted by modern postprocessing techni-
Introduction
ques. Four visualization techniques in clinical use are described
in this paper including the advantages and disadvantages of
Diagnostic methods for suspected airway obstruction in chil-
each: multiplanar reformation, maximum and minimum inten-
dren include imaging in the form of chest radiographs and CT
sity projections, shaded surface display and volume rendering.
scanning as well as fibre-optic tracheobronchoscopy (FTB).
Volume-rendered internal visualization in the form of virtual
FTB is an excellent way of evaluating stridor caused by airway
endoscopy is also discussed. In addition, the clinical usefulness
compression and for visualizing the dynamic changes in airway
in paediatric practice of demonstrating airway compression and
calibre [1]. It is usually performed by a pulmonologist and
its causes are discussed. Advanced postprocessing techniques
represents the gold standard for airway investigation, but it
that must still find their way from the biomedical research
has numerous limitations. Even though the procedure is con-
environment into clinical use are introduced with specific ref-
sidered a safe examination if performed by an experienced user,
erence to computer-aided diagnosis.
it is invasive and requires general anaesthesia, which carries
further risks [1]. Complications such as hypoxaemia, hyper-
carbia, cardiac arrhythmia and subglottic oedema have resulted
in the search for noninvasive diagnostic techniques [1]. Imag-
S. Andronikou : L. T. Hlabangana : T. Pillay
ing is one of these solutions, but it must meet the aims of FTB.
Radiology Department, Faculty of Health Sciences,
University of the Witwatersrand, The main aims of FTB are to confirm airway narrowing,
Johannesburg, South Africa evaluate the degree of stenosis, aspirate endobronchial con-
tents and perform transbronchial biopsy [1]. Imaging should
B. Irving : P. Taylor
CHIME, Division of Population Health,
also provide additional information acceptable as an alterna-
University College London, tive, particularly where FTB has limitations such as in dem-
London, UK onstrating the cause of a stenosis, demonstrating relationships
with surrounding structures and navigating airways that are
P. Goussard : R. Gie
difficult to access by FTB (such as the lingual and left lower
Department of Pediatrics and Child Health,
Faculty of Health Sciences, University of Stellenbosch, lobe bronchi) and the airways beyond a tight stenosis [1].
Stellenbosch, South Africa CT has established its place in the evaluation of paediat-
ric stridor as it is easy to perform and requires no sedation
S. Andronikou (*)
due to the speed of the procedure. It does, however, carry a
700 Harbouredge, 10 Hospital Str, Greenpoint,
Cape Town 8005, South Africa radiation burden, which must be considered. Simple solu-
e-mail: [email protected] tions such as reducing the kV and mAs can reduce the dose
270 Pediatr Radiol (2013) 43:269–284

by up to 65% [1]. CT volumes have traditionally been

viewed as individual axial slices, but the availability of
multidetector row CT (MDCT) has led to an increase in
the clinical use of images in other planes. The advances in
postprocessing have expanded the role of CT to a degree
where it challenges other diagnostic techniques as the gold
standard for diagnosis [2].
Postprocessing allows more scan information to be
viewed at one time allowing an accurate diagnosis to
be made more efficiently. The speed at which this can
be done may be considered as a real-time interactive
modification process. Therefore, understanding the need
for routine thin-section axial MDCT to allow 3-D applications
is critical for making clinical use of postprocessing [3]. Use of
postprocessing including 3-D visualization has become a ne-
cessity [4] that is particularly important in paediatric practice
where invasive clinical diagnostic tools can be replaced or at
least assisted by modern postprocessing techniques.
This review describes the MDCT postprocessing
tools in routine clinical use and identifies indications,
advantages and disadvantages of each in relation to
imaging the paediatric airways. It also explores more
advanced postprocessing, more familiar to biomedical
researchers, that will soon find their way into clinical
practice, in particular in relation to computer-aided diagnostic
tools.

Visualization techniques

Four visualization techniques are in clinical use on clinical

Fig. 1 Multiplanar reconstruction. Using a cross-hair on (a) the coro-
workstations: multiplanar reformation (MPR) including nal image, it is possible to confirm the location and calcification of
curved plane reformation (CPR), maximum and minimum right hilar lymphadenopathy (thick arrow) on (b) the axial image (thin
intensity projections (MIP/MinIP), shaded surface display arrow)
(SSD) and volume rendering (VR). The first two techniques
are limited to external visualization while SSD and VR over the structure of interest as viewed on transverse
can be used for both external and internal visualization sections, MPRs, MIPs, SSDs or VRs. The points are
[4]. connected to form a 3-D curve that is then extruded
through the volume perpendicular to the desired view to
Multiplanar reconstruction and curved plane reformation create the CPR [4]. CPRs are very useful for displaying
the interior of tubular structures such as blood vessels,
MPR is the process of using the data from axial CT airways and bowel. They are also useful for visualizing
scans to create non-axial 2-D images. MPR images are structures immediately adjacent to these lumina, such as
coronal, sagittal, oblique or curved plane images gener- mural thrombus [4]. Unlike SSDs or VRs, CPR images
ated from only one voxel thickness transecting a set or display the cross-sectional profile of a vessel along its
stack of axial images (Fig. 1). MPR images are in length, facilitating characterization of stenosis or other
routine use, including for the evaluation of airways. intraluminal abnormalities [3].
Similar to MPR, CPR is a tomogram one voxel thick,
but it is capable of demonstrating an uninterrupted lon- Limitations
gitudinal cross-section because the display plane curves
along the structure of interest. A CPR can be created to One major limitation of traditional MPR is that structures
include an entire structure on a single image [3]. CPRs must lie in a plane, and almost all structures for which 3-
are created along points that are manually positioned D visualization is desired do not lie within a single plane.
Pediatr Radiol (2013) 43:269–284 271

MPR therefore cannot be used to demonstrate an entire

structure at one time and pseudostenoses may appear on
MPR images. One solution to this problem is to use CPR
[4]. An important limitation of CPRs is that they are
highly dependent on the accuracy of the curve. Inaccu-
rately positioned or insufficient numbers of points can
result in the curve slipping-off the structure of interest,
again leading to pseudostenoses [3, 4]. A single curve also
cannot accurately display eccentric lesions. Therefore, two
orthogonal curves should always be created to provide a
more complete depiction of eccentric lesions, particularly
in stenosis [4]. Manual derivation of the curved plane is
also time-consuming [3].

Maximum intensity projection

Multiplanar images can be thickened into slabs using

projectional methods such as MIP, MinIP and average Fig. 2 Maximum intensity projection. When viewed parasagitally, the
intensity projections/raysum (sum of all pixel values en- MIP image clearly demonstrates the contrast-enhanced blood vessels
countered by each ray to provide an image similar to a that are displaced and encased by a mass lesion (arrow) in the neck.
Extension of the mass into the thorax is clearly demonstrated
radiograph) [3, 4]. MIP is a widely used rendering tool
particularly for evaluation and display in CT angiography,
and has been shown to be more accurate than surface
rendering for evaluating the vasculature. However, an but this is mainly a problem when evaluating atherosclerotic
understanding of how the MIP algorithm produces render- vessels in adults where there are large calcifications within
ings and of the limitations of MIP is essential for the the vascular lumen. The 2-D MIP display can also depict the
correct interpretation of MIP images [2]. 3-D relationships of structures inaccurately, causing misrep-
MIPs are created when a specific projection is selected resentation, especially when there are overlapping struc-
(e.g. anteroposterior) and rays are cast perpendicular to tures, because there are no visual cues for depth perception
the view through the volume data MIP with only the [3]. This limitation can partially be overcome by thinner
brightest voxel being projected into the resultant image. sliding slab MIP reconstructions [6, 7].
The algorithm achieves this by selecting the voxels with
the highest attenuation along lines projected through the
volume dataset [4, 5]. The result is a 2-D image of 3-D Table 1 Comparison of advantages and disadvantages of VR and
information because the entire volume is collapsed with MinIP/MIP
only the brightest structures being visible. Thick-slab
Volume Rendering (VR) MinIP/MIP
MIPs can be applied to angiography data to include long
segments of a vessel [3] (Fig. 2). There tends to be much Accurate 3-D relationships Inaccurate 3-D relations
less variability in MIP image reconstruction than in VR demonstrated
because fewer parameters are factored into the MIP al- Shows additional structures Only shows airways or vessels
(soft tissue, bone) in detail
gorithm [4] (Table 1). The application of MIP reconstruc-
Colour display helps interpret Grey-scale windowing only
tions for the lung has been shown to increase nodule complex relationships
detection and can help differentiate between small nod- Calcified vessels not a major Calcified vessels result in
ules and vessels [6]. This is not an ideal method for problem for depicting lumen size inaccurate representation of
demonstrating the airway itself, but it is useful for eval- vessel size
uating the surrounding structures to determine a cause of Does not require editing Requires editing
airway compression. Does require adjustment of Simplicity makes it easy to
parameters, and interactive nature master and does not require
makes it subject to interobserver much interactivity
Limitations variability (speed depends on the
user)
MIP images usually contain 10% or less of the original data Advantages in ability to rotate a Some advantages in
structure demonstrating smaller vessels,
[3]. The presence of high-attenuation voxels, other than especially collaterals
those of interest, may obscure evaluation of the vasculature,
272 Pediatr Radiol (2013) 43:269–284

Minimum intensity projection

This is a particularly useful technique for imaging the major

airways [4, 5]. MinIP images are a variation of the MIP
approach where multiplanar slab images are produced by
projecting only the lowest attenuation value when a ray is cast
perpendicular to the view through the volume of data (Fig. 3)
highlighting air-filled structures (as these have the lowest
Hounsfield unit value). Most MinIP and MIP methods use only
windowing parameters (window width and centre, specified in
Hounsfield units) and not colour (Fig. 4). An advantage of
MinIPs over MPRs is that structures that do not lie in a single
plane, such as the bronchial tree, are visible in their entirety
[5]. This allows determination of the length of airway stric-
tures on one slice (Fig. 5). Thin-slab MinIP images (with
section thickness less than 10 mm) viewed in sequence
(much like scrolling through single-pixel MPR images),
also known as a sliding slab, may provide more useful Fig. 4 Thick-slab coronal MinIP image using a soft tissue window
demonstrates the full length of the major airways as well as lymph-
diagnostic information, as small structures are better adenopathy, which is causing subtle compression of the left main
detected [3] (Fig. 6). bronchus (arrows)

Limitations allow several threshold ranges to be defined and displayed

using different colours. In this setting, different tissue types or
Limitations are the same as for MIP especially with regard structures are coded as different colours to allow differentia-
to lack of appreciation of depth relationships [3]. tion of adjacent structures, but for each classification, data
segmentation is required, typically by using a threshold and
Shaded surface display editing, which dramatically increases postprocessing time.
However, the main advantage of threshold-based rendering
SSD is termed a threshold-based classification technique. It is its processing speed, since a comparatively small amount of
provides 3-D views of the surface of an object by using grey- computational power is needed [4].
scale to demonstrate surface reflections from an imaginary
source of light. SSD displays a single surface generated from Limitations
thresholds selected by the user [4] (Fig. 7). The CT data are
reduced to binary data to achieve this, defining each pixel as Much like MIP, SSD is flawed because it uses less than 10% of
either in or out of the threshold range [3, 4]. Some versions the acquired data [3]. The reduction to binary data limits

Fig. 3 Minimum Intensity

projection. a The thickness of
the coronal slab is decided on
the sagittal image as a vertical
thick slab selection (white
arrows). b This yields a coronal
MinIP image that includes the
full thickness of all major
branches of the
tracheobronchial tree, thereby
reliably demonstrating
compression of the trachea and
the right bronchus (black
arrows) in this patient with situs
inversus
Pediatr Radiol (2013) 43:269–284 273

Fig. 5 Coronal MinIP images

utilizing a lung window. The
major airways are demonstrated
on a single image, providing
information on the degree of
compression of the bronchus
intermedius (a white arrow) as
well as allowing measurement
of the length of the compression
(b black arrow), which has
resulted from right hilar and
subcarinal lymphadenopathy

flexibility and makes this postprocessing technique prone to Volume rendering

artefacts. Regardless of the number of tissue groups or classes
assigned, the selection of the threshold range that defines the VR is the most advanced of the rendering techniques and
groups is arbitrary and this limits accuracy [4]. Voxels that is the technique that allows the majority of clinically
represent mixed tissue interfaces cannot be accurately classi- useful postprocessing. It refers to a 3-D volume recon-
fied. This makes the system incompatible with volume aver- struction method that allows every voxel in the volume
aging and results in incorrect classification of voxels that data to contribute to the reconstructed image [7]. The
contain volume averages [3]. The thresholding makes the technique is termed a percentage classification (semitrans-
technique susceptible to noise introduced during scanning. A parent volume-based/continuous technique). Simplistical-
small amount of noise can modify attenuation values and create ly, all voxel values are assigned an opacity level that
the appearance of soft tissue in a voxel that actually represents varies from total transparency to total opacity. This can
mostly bone [2]. All of these disadvantages add up and many be applied to voxel values as a whole or to regions of
artefacts are described on the end image: “holes in structures, the histogram that are classified as specific tissue types
contours that represent voxel boundaries rather than true tissue [4]. In percentage classification, it is assumed that a
interfaces, fragments of structures floating in space, and ab- voxel may represent one or more tissue types and that
sence or exaggeration of details such as bone fractures” [2]. the amount of each tissue as a percentage of the entire

Fig. 6 Sequential coronal oblique thin-slice sliding slab MinIP images patient, there is complete occlusion of the right upper lobe bronchus
using soft-tissue window. Thin slab images do not demonstrate the and moderate narrowing of the bronchus intermedius resulting from
airway on one single slab but rather demonstrate (a) the right major tuberculous lymphadenopathy (L). There is also extensive parenchy-
airways (white arrow) on an anterior slice and (b) the left airways mal disease distal to the right upper lobe obstruction with necrosis and
(black arrow) on contiguous thin slabs. This is similar to scrolling breakdown
through a coronal MPR but having a choice of slice thickness. In this
274 Pediatr Radiol (2013) 43:269–284

viewing perspectives by rotating the 3-D reconstruction

through a preset number of views or by manually cus-
tomizing a desired view in real time (Fig. 9). VR
images do not have many of the computer artefacts found
with SSD. Most importantly, it has been shown that average
reading time using VR images is significantly shorter than
with MIP images [7].

Limitations

Inaccuracies in VR stem from mucus secretions, dynamic

airway changes and artefacts such as the stair-step artefact
encountered in up to 3% of images [1] (Fig. 10). VR also
requires more computational power than SSD because each
voxel must be projected into an image, while in SSD only the
surface data need processing. On modern computers, even
large datasets can be manipulated interactively, and the display
can be instantaneously changed from VR to MIP as
Fig. 7 Surface-shaded display image of the chest, anterior view. The required. Interactive adjustment of VR parameters, rota-
image was generated using a threshold setting to preferentially dem-
onstrate dense structures including bones, contrast-enhanced blood tion of the dataset and automated clip plane editing all
vessels and heart can be performed in real time, but this places additional
time constraints on the radiologist, who needs to be
voxel is between 0% and 100% [4, 7]. Once colour and involved interactively.
transparency are assigned to each classified voxel, a 3-D
image is produced by casting simulated rays of light
through the volume. In addition, in VR there are adjust-
Endoluminal visualization/virtual endoscopy
able parameters to change the way the image looks,
including window setting, colour, degree of opacity and
Both SSD and VR allow the production of images of the
shading (Fig. 8). Angle and distance can also be deter-
inner surface of tubular lumens, and the technique has been
mined. VR allows the operator to select a variety of
termed virtual endoscopy. This can be used to examine the
bowel, airways, blood vessels and the urinary tract [4, 5].
The basic technique involves identification of threshold
levels for SSD that exclude pixels of similar attenuation to
the lumen (−900 to −1,000 HU for air-filled volume and 150
to 400 HU for contrast-enhanced volume) or an opacity
curve for VR that results in complete transparency of the
lumen. Rendering creates an image of the interface between
luminal contrast and extraluminal attenuation and not the
mucosa or intimal surface. The two methods widely in use
are orthographic external rendering with cut planes and immer-
sive perspective rendering [4].

Orthographic rendering

This is the most commonly used technique particularly for

external visualization of a lumen from a viewpoint that is
external to the data [3, 4]. It can, however, also be used for
Fig. 8 3-D volume-rendered image generated using density, transpar- internal visualization when combined with cut planes posi-
ency and colour suited to demonstrating bone and blood vessels
enhanced by contrast. As well as demonstrating these structures, the
tioned within the lumen of the structure of interest. This can
depth relationship of each can be appreciated as a possible cause of be compared to cutting a “window into a piece of pipe to
tracheal compression visualize its interior” [4].
Pediatr Radiol (2013) 43:269–284 275

Fig. 9 3-D volume-rendered

images generated using density
and transparency settings suited
to airway and lung visualiza-
tion. Rotation of the dataset
allows the degree and length of
left main bronchus narrowing
(arrows) to be visualized. In
this patient, the left main bron-
chus compression was due to an
anomalous artery

Limitations/disadvantages Immersive/perspective rendering

This method only provides a regional snapshot; it cannot The viewpoint using this technique is from within the lumen
provide a continuous demonstration of all interior surfaces (Fig. 11) mimicking fibre-optic endoscopy, and bypassing the
of a lumen [4]. limitations of the invasive technique that requires access to the
lumen and has a restricted direction of viewing. For the viewer
to have perspective on depth relationships at close range, a
modelling technique is used that functions in a manner similar
to the human visual system. In the same way that light rays are
focused to converge on the retina, the viewer recognizes the
distance of structures depending on their size. A structure
close to the eye appears larger than a structure farther away.
This effect is determined by the field of view of the virtual lens
[4].

Limitations

Opacity and colour selection and the complexity of creating

these visualizations can be daunting. The greatest problem
with immersive visualization, however, is navigation. This is
because there are three spatial degrees of possible position and
three spatial degrees of possible view direction [4]. Without an
external guide of the view position, it is easy to lose track of
location within the lumen. Techniques that automatically cre-
ate a flight path through the centre of a lumen are being
developed to address this [4].

Advantages of CT postprocessing of the airway

in children
Fig. 10 3-D VR image showing the stair-step artefact (arrows) in a
child with mild narrowing of the bronchus intermedius and left main Even though there is a shift away from the use of MDCT in
bronchus children because of the radiation dose, it is important that,
276 Pediatr Radiol (2013) 43:269–284

Fig. 11 The immersive (perspective) rendering technique results in images that mimic fibre-optic tracheobronchoscopy. Navigation tool image (a)
and virtual endoscopy image (b) at the level of the right upper lobe bronchus (white arrow) and bronchus intermedius (black arrow)

having decided to use CT, a full volume of data is obtained demonstrating airways, it is advantageous to depict the full
after administration of intravenous contrast agent. This then tracheobronchial tree at once, the relationships of normal
represents a single and complete MDCT study that can be anatomical structures to the airway, any anomalies associat-
reconstructed as many times as necessary, in as many ways ed and any pathology that may be causing the airway
as desired and for as long as the raw data are stored. compression in a way that clinicians and parents can under-
To be useful, CT postprocessing must result in better stand. This can be achieved using postprocessing of images
anatomical detail and supply additional information relevant to mimic gross pathological specimens or endoscopic views.
to diagnosis and management of the disease than the origi- Thus, without additional imaging or radiation burden to a
nal axial images. A radiologist may, however, see the ad- child, a clinician or a parent may be convinced of the
vantage of providing imaging information that is more easy benefits of further management or even of conservative
to interpret and more easy to communicate to the referring management so that unnecessary interventions are avoided.
clinician or parent even in situations where there is no Reconstructed CT views of the airway in children have been
further diagnostic value than the original axial views. For reported to have significantly reinforced the confidence in

Fig. 12 Thick-slab MinIP in a child with primary lymphobronchial the left main bronchus (LMB). This provides a practical means of
tuberculosis. Axial (a) and coronal (b) views demonstrate airway demonstrating pathology to pulmonologists and thoracic surgeons
compression by lymphadenopathy (L) at multiple sites including the who are contemplating lymph node enucleation for relieving the
right upper lobe bronchus (RULB), the bronchus intermedius (BI) and obstruction
Pediatr Radiol (2013) 43:269–284 277

diagnoses and improved communication with clinicians allow- degree and length and identifying the cause of an airway
ing more comprehensive surgical planning [8–11]. stenosis [1]. The accuracy of detection of airway compression
MPR is as accurate as axial CT, is easy to create and view, by 3-D VR has been reported to 95.7% compared to 91.5% for
and improves perception as it displays information effectively conventional CT [12].
[8, 9]. MinIP allows the detection of low-density structures Virtual bronchoscopy (VB) is a noninvasive endoluminal
and is therefore ideal for improving evaluation of paediatric visualization technique that simulates bronchoscopy [9]. It
airways [9, 10] (Fig. 12). MIP is ideally suited to demonstrat- provides a realistic endoscopic 3-D view of the tracheobron-
ing hyperdense pathologies such as vessel abnormalities, nod- chial tree and is particularly useful in children. VB has addi-
ules, calcifications and foreign bodies that are associated with tional advantages over true bronchoscopy in that it also allows
airway pathology [9]. VR techniques add value to imaging simultaneous visualization of structures around the tracheo-
complex structures and interfaces that cross the traditional bronchial tree, which helps in the identification of the cause of
imaging planes [8, 9]. The 3-D images [9] simplify demon- an obstruction [11]. VB is particularly attractive for imaging
stration of pathology to the referring clinician and increase airway stenosis when traditional bronchoscopy presents a risk
diagnostic confidence for all involved [8]. They are excellent to a child, or when navigation of a bronchoscope is not
for demonstrating a decrease in airway calibre, spatial rela- possible due to a high-grade stricture [8], and VR is used to
tionships of structures, measuring stenoses to determine outline the bronchial lumen beyond this. The technique can

Fig. 13 MinIP images on lung window in children presenting with lous lymphadenopathy (right upper lobe, bronchus intermedius and left
airway symptoms and signs. a Right tracheal displacement due to a left main bronchus). d Isolated narrowing of the left lower lobe bronchus
neck mass (arrow). b Focal left main bronchus compression by tuber- (arrow) due to tuberculous lymphadenopathy
culous lymphadenopathy. c Multifocal airway narrowing by tubercu-
278 Pediatr Radiol (2013) 43:269–284

Fig. 14 3-D VR images. Multiple rotated views demonstrate the advantages of 3-D VR in the diagnosis and evaluation of this double aortic arch
anomaly (arrows) as a cause for stridor and tracheal stenosis

also be used for planning transbronchial biopsy, endobron- defining bronchoceles [8] and when planning a diagnos-
chial laser therapy and stenting [5]. tic lung biopsy [8].
For imaging lymphobronchial tuberculosis (i.e. the in-
volvement of the airway by compressive lymphadenopathy
Postprocessing for airway diseases in children of primary tuberculosis), the relationships between the bron-
chial walls and lymph nodes are extremely well demonstrat-
Identification and better characterization of tracheobronchial ed using reconstruction techniques [8]. The airway is most
stenosis is the main indication for airway postprocessing
techniques. Even though MPR and MinIP are extremely use-
ful tools in evaluating tracheobronchial stenoses (Fig. 13),
which are vertical or slightly oblique [8], it is 3-D VR
that both allows 3-D reformatting of airways and demon-
strates associated adjacent vascular structures [8]. It illus-
trates focal areas of narrowing, the craniocaudal length of
tracheobronchial stenoses and even demonstrates the air-
way beyond the major stenosis. MinIP may demonstrate
the airway equally as well, but VR is more useful when
complex tracheobronchial congenital anomalies with vas-
cular and other anatomical associations are being evaluated
(Fig. 14) [8, 10].
Causes of tracheobronchial stenosis in children in-
clude tracheomalacia often associated with compressive
vascular rings (Fig. 13), tracheooesophageal fistulas, Fig. 15 Coronal MinIP image on lung window. MinIP adequately
hilar or mediastinal tumours, foreign bodies (Fig. 15) demonstrates endobronchial material (arrow) in the right main bron-
chus and bronchus intermedius (in this case granulation tissue from a
[8, 11] and mucoid impaction. In addition, reconstruc- tuberculous lymph node erosion into the airway). Fibre-optic tracheo-
tions can be used for imaging the distal airways [8, 10], bronchoscopy was performed and the material was aspirated acting as
especially in suspected bronchiectasis (Fig. 16), for both a therapeutic and a diagnostic procedure
Pediatr Radiol (2013) 43:269–284 279

Fig. 16 Coronal MinIP images

on lung window. MinIP
demonstrates the distal airways
and can distinguish air
bronchograms related to an air-
space process (white arrows)
(a) from bronchiectasis (black
arrows) (b)

easily compared to a chest radiograph using the MinIP additionally successful in demonstrating a further 51 sites
technique, which is considered to be a useful training tool of stenosis, provided a description of the degree of stenosis
for tuberculosis workers restricted to the use of plain radio- in all patients, and in three of four patients even dem-
graphs (Fig. 13). Du Plessis et al. [1] compared 3-D VR with onstrated stenotic airways beyond a proximal stenosis
FTB in 26 children with lymphobronchial tuberculosis (me- that the bronchoscope could not pass [1]. The readers
dian age 21 months) and demonstrated a 92% sensitivity and also expressed the length of the stenosis and identified
85% specificity for airway compression compared to tradi- the cause of the stenosis in all children. FTB assumed
tional bronchoscopy. However, the VR technique was the cause to be lymphadenopathy in all children, but

Fig. 17 3-D VR images.

Multiple views demonstrate
tracheal (a white arrow) and left
main bronchus (a black arrow)
narrowing caused by large
mediastinal and subcarinal
lymphadenopathy (a–d L) in a
child with primary pulmonary
tuberculosis
280 Pediatr Radiol (2013) 43:269–284

Fig. 18 Immersive virtual

endoscopy in the same child as
in Fig. 17 shows tracheal
narrowing (a white arrows) and
left main bronchus narrowing
(b black arrows) adjacent to the
respective navigation coronal
images that help orientate the
operator

VR demonstrated that 14% of the airway stenoses were postprocessing can apply learning algorithms to develop
from other causes [1]. models from entire datasets of CT scans. These models
Figure 17 demonstrates the usefulness of the 3-D VR can then be used to automatically identify each anatom-
technique in demonstrating tracheal and left main bronchus ical airway branch and then detect normal and patho-
compression while also demonstrating the causative lymph- logical variations related to a specific disease. Most
adenopathy and relationships to blood vessels. Figure 18 postprocessing methods have been developed for and
demonstrates the navigation and endoscopic views in the tested on adult airways, and limited work has focused
same patient, which may assist in communicating the loca- on paediatric airways. Paediatric airway analysis
tion and severity of stenosis to a bronchoscopist planning to
perform a transbronchial biopsy.

Research tools: advanced postprocessing currently

for research and CAD

Automatic segmentation of high-resolution CT images

can extract and analyse the structure of the tracheobron-
chial tree and distinguish airway regions from surround-
ing tissue, including the lung. Improvements to these
methods mean that smaller branches can be visualized.
Postprocessing can be used to perform various measure-
ments including identification of branching topology as
well as branch dimensions and cross-section measure-
ments. This technique can generate automated airway Fig. 19 Segmentation with bone rendering of the airway in a 20-
statistics at the click of a button. More sophisticated month-old patient
Pediatr Radiol (2013) 43:269–284 281

introduces additional challenges to postprocessing.

Smaller patient size and radiation dose considerations
mean that the airways are extracted at a lower resolu-
tion and fewer branches are extracted [10]. In this
section, we present general methods that can be applied
to adults and children as well as specific methods and exam-
ples from our own work in paediatric airway analysis.

Segmentation

There are many airway segmentation algorithms that use

a variety of mathematical techniques, but these methods
Fig. 21 Branch measurement by projection of points onto the airway
usually comprise common steps. A typical airway seg- surface from the medial line
mentation algorithm would consist of the following steps:
(1) identification of an initialization point within the air-
ways, (2) application of filtering techniques to enhance Very few of the methods have been applied to paedi-
the airways, (3) detection of connected airway regions atric CT volumes. Irving et al. [18, 20] developed a
from the initialization point, and (4) detection and re- method targeted at paediatric airway volumes in which a
moval of non-airway regions that have been mistakenly threshold is applied to extract air-filled regions and there-
classified. fore approximate the lung regions. The trachea is identi-
The main airways can be segmented more easily than fied in the axial slices above the lung region and used to
smaller bronchi because they have a well-defined wall and initialize the segmentation. The trachea is air-filled and
therefore there is greater contrast between wall and lumen. therefore has a lower HU value than the surrounding
Simple thresholding methods are often used to extract the tissue, and can be identified in cross-section. However,
larger airways while more sophisticated methods are used for other objects within the same HU range may also be extracted
the smaller bronchi. This saves processing time and allows the so that a number of features, including the cross-sectional
algorithm to be tailored more specifically to the smaller air- area, location and circularity, are used to identify the trachea
ways [13]. Airway segmentation methods include: adaptive [18]. A morphological filtering method is then applied to the
thresholding [14], fuzzy connectivity [15], rule-based segmen- CT volume in the axial, coronal and sagittal directions. This
tation [13], morphology-based connection cost [16], and mor- filter uses morphological closing and reconstruction to en-
phological filtering [17, 18]. The evaluation of these methods hance regions with a circular appearance [21] and therefore
is a challenge because there is no common gold standard enhances cross-sections of the airway in the three directions.
airway segmentation dataset. However, a number of state-of- Once a range of morphological filters have been applied to
the-art airway segmentation methods were evaluated as part of enhance airways of various sizes, a threshold value is applied
the EXACT'09 airway segmentation competition [19]. to extract likely airway regions. Finally, a region-growing
method is applied to extract just the connected airway. This
region-growing method starts at the seed point located in the
trachea and adds neighbouring voxels with 26-connectivity
until no other voxels can be added [18].
Segmentation of the airway can be used to improve the
rendering of the airway as well as to provide automated
analysis of the airway that cannot be achieved with just
rendering. An example of the resulting segmentation of the
airways in a 20-month-old patient is shown in Fig. 19. Note
also that some small regions have been misclassified as airway
and there is a limit to the number of branches that are seg-
mented. Each improvement in paediatric CT imaging and
segmentation will lead to more accurate airway extraction.
The limitation of many segmentation algorithms is that in
order to extract just the airways without additional regions
Fig. 20 Airway centre line and bifurcation point detection from a being mistakenly included, the airway needs to be complete-
paediatric chest CT scan ly connected, i.e. branches are not obstructed. To include
282 Pediatr Radiol (2013) 43:269–284

obstructed branches, our algorithm performs a shape analy- calculated from the skeleton length and branch volume is
sis of each section of the airway to identify obstructed approximated as the number of voxels in the branch multi-
branches and then searches for additional airway compo- plied by the voxel dimensions [22]. From this, the approx-
nents beyond the obstruction [20]. imate mean radius of the branch can be determined. Cross-
sections can also be measured along each branch by projec-
Skeletonization ting vectors orthogonal to the centre line to the branch
surface [15]. This allows cross-sectional area and radius to
Once the airways have been segmented, centre-line extrac- be calculated, and can be extended to identify and measure
tion is an important intermediate step before more advanced branch local minima and maxima [23]. Figure 21 shows the
airway processing can be performed. projection of orthogonal vectors and calculation of the in-
Airway centre-line extraction algorithms produce a centre- tersection of these vectors with the surface of the airways to
line one voxel thick that bifurcates to form the centre line measure a branch.
for each child branch (Fig. 20). The topological structure—
including branch start and end points and the relationships Branch labelling
between parent and child branches—can be extracted. The
extracted medial line can also be used to direct VB and The topological structure of the airway can be used to match
extract airway cross-sections. One method uses iterative each branch to its anatomical label. This can be used to
“thinning” of the segmentation by removing voxels that provide annotated visualizations of the airway, present data
do not affect the airway structure until only a one-voxel about a branch of interest, and automatically compare
thick branching centre line remains [22]. The centre line branches in a dataset of airways. However, the labelling is
extracted using this method in a child is shown in Fig. 20. made difficult by the variation in the shape and orientation
of the branches, and the possibility of anomalous branches.
Branch measurements Common methods use a template of the airway branching
structure and match this to individual airway trees. Branch
Once the branching structure has been extracted, the start, measurements such as branch length and orientation, angles
endpoint and centre line of the branch are defined. This between branches, and relationships between parent and
allows automatic measurement of the length, curvature and child branches are used to match the template to the airways
cross-section of the branch. The voxel size is used as a and a function is used to optimize these labels over the
scaling factor for the volume. Branch length can be whole airway tree [24, 25].

Fig. 22 Principal component

analysis variation along one
example mode of the trachea/
right main bronchus/left main
bronchus (Trachea-LMB-RMB)
and one example mode of the
right main bronchus/right upper
lobe/bronchus intermedius
(RMB-RUL-BI) regions
Pediatr Radiol (2013) 43:269–284 283

Computer-assisted detection of airway pathology supply images that are easily interpreted, specifically for
improved communication with referring clinicians and
Automated segmentation and branch analysis can be used to parents. Postprocessing methods also have the potential
evaluate and classify airways in a paediatric CT volume data- to automate airway visualization, identify airway anato-
set. This procedure is used to differentiate airway deformation my and detect regions of pathology. Improved airway
and stenosis from lymphadenopathy in children with tubercu- segmentation will offer improved visualization of the
losis. However, this can be extended to other airway patholo- smaller bronchi and allow automatic extraction of the
gies and used in conjunction with other features of pathology. airway structure. Branch measurements can then be au-
Once a dataset of airways has been segmented and the tomated and analysis of large datasets of images can be
structure analysed, the dataset can be used to train a classi- performed. These methods would allow clinicians to
fier to distinguish between normal and abnormal airway view the airways with overlays indicating airway prop-
variations. Irving et al. [23] created a point distribution erties or the likelihood of disease based on comparison
model that represents the airways as a set of points each to a dataset of airway images.
corresponding to a point on every other airway in the data-
set. The point distribution model for each airway is auto-
matically generated using a thin plate spline warp and References
localized matching to fit a template airway onto each airway
segmentation. In the model of Irving et al. [23], three branch
1. du Plessis J, Goussard P, Andronikou S et al (2009) Comparing
sections of the airway are represented by a set of about three-dimensional volume-rendered CT images with fibreoptic tra-
1,500 (n) vertices. Each is a point in 3-D space and therefore cheobronchoscopy in the evaluation of airway compression caused
each airway is represented by (3×n) features. Principal by tuberculous lymphadenopathy in children. Pediatr Radiol
component analysis is a technique that can be used to reduce 39:694–702
2. Fishman EK, Ney DR, Heath DG et al (2006) Volume rendering
the dimensionality of this feature space and extract the main
versus maximum intensity projection in CT angiography: what
modes of variation from the covariance of the points within works best, when, and why. Radiographics 26:905–922
the airway dataset. This results in a more tractable number 3. Dalrymple NC, Prasad SR, Frecketon MW et al (2005) Informatics
of features that can be used to classify normal and abnormal in radiology (infoRAD): introduction to the language of three-
dimensional imaging with multidetector CT. Radiographics
variations. Figure 22 shows one mode of variation from the
25:1409–1428
trachea/right main bronchus/left main bronchus region 4. Rubin GD (2003) 3-D imaging with MDCT. Eur J Radiol 45
and one from the right main bronchus/right upper (Suppl 1):S37–S41
lobe/bronchus intermedius region. Irving et al. [23] used 5. Kirchgeorg MA, Prokop M (1998) Increasing spiral CT benefits
with postprocessing applications. Eur J Radiol 28:39–54
ten modes of variation to represent each airway.
6. Ueno J, Murase T, Yoneda K et al (2004) Three-dimensional
This method was evaluated using cross-validation on a imaging of thoracic diseases with multi-detector row CT. J Med
dataset of 61 patients with and without tuberculosis. All the Invest 51:163–170
patients with tuberculosis showed symptoms and signs of 7. Peloschek P, Sailer J, Weber M et al (2007) Pulmonary nodules:
sensitivity of maximum intensity projection versus that of volume
major airway compression. The method was able to distin-
rendering of 3D multidetector CT data. Radiology 243:561–569
guish between patients with and those without tuberculosis 8. Papaioannou G, Young C, Owens CM (2007) Multidetector row
with a sensitivity of 86% and a specificity of 91%. CT for imaging the paediatric tracheobronchial tree. Paediatr
Kiraly et al. [26] reviewed a number of other automat- Radiol 37:515–529
9. Perandini S, Feccioli N, Zaccarella A et al (2010) The diagnostic
ed methods for airway assessment and visualization. Al-
contribution of CT volumetric rendering techniques in routine
though these methods were evaluated on adult datasets, practice. Indian J Radiol Imaging 20:92–97
many methods could also be applied to paediatric airway 10. Siegel M (2003) Multiplanar and three-dimensional multi-detector
analysis. These include methods to determine airway lu- row CT of thoracic vessels and airways in the pediatric population.
Radiology 229:641–650
men size and wall thickness, identify mucus plugs and
11. Sodhi KS, Aiyappan SK, Saxena AK et al (2010) Utility of multi-
determine bronchus–artery ratios. These measurements can detector CT and virtual bronchoscopy in tracheobronchial obstruc-
then be presented as colour-coded airway visualizations tion in children. Acta Paediatr 99:1011–1015
for assessment. 12. Remy-Jardin M, Remy J, Artaud D et al (1998) Volume rendering
of the tracheobronchial tree: clinical evaluation of bronchographic
images. Radiology 208:761–770
13. Sonka M, Park W, Hoffman E (1996) Rule-based detection of
Conclusion intrathoracic airway trees. IEEE Trans Med Imaging 15:314–326
14. Mori K, Hasegawa J, Toriwaki J et al (1996) Recognition of
bronchus in three-dimensional X-ray CT images with application
CT postprocessing techniques permit the demonstration
to virtualized bronchoscopy system. In: Proceedings of the 13th
of fine anatomical detail, provide additional information International Conference on Pattern Recognition, vol 3. Vienna,
relevant to diagnosing and managing airway disease and Austria. IEEE Computer Society, Washington DC, pp 528–532
284 Pediatr Radiol (2013) 43:269–284

15. Tschirren J, Hoffman E, McLennan G et al (2005) Intratho- statistical shape analysis. In: Proceedings of the eighth IEEE
racic airway trees: segmentation and airway morphology anal- international symposium on biomedical imaging: from nano to
ysis from low-dose CT scans. IEEE Trans Med Imaging macro. IEEE, Piscataway, NJ, pp 447–451
24:1529–1539 21. Vincent L (1993) Morphological grayscale reconstruction in image
16. Fetita C, Preteux F, Beigelman-Aubry C et al (2004) Pulmonary analysis: applications and efficient algorithms. IEEE Trans Image
airways: 3-D reconstruction from multislice CT and clinical inves- Process 2:176–201
tigation. IEEE Trans Med Imaging 23:1353–1364 22. Palágyi K, Tschirren J, Hoffman E et al (2006) Quantitative analysis
17. Kiraly A, Higgins W, McLennan G et al (2002) Three-dimensional of pulmonary airway tree structures. Comput Biol Med 36:974–996
human airway segmentation methods for clinical virtual bronchos- 23. Irving B, Goussard P, Gie R et al (2011) Identification of paediatric
copy. Acad Radiol 9:1153–1168 tuberculosis from airway shape features. Med Image Comput
18. Irving B, Taylor P, Todd-Pokropek A (2009) 3D segmentation of the Comput Assist Interv 14:133–140
airway tree using a morphology based method. In: Brown M, de Bruijne 24. Tschirren J, McLennan G, Palágyi K et al (2005) Matching and
M, van Ginneken B et al (eds) Proceedings of the second international anatomical labeling of human airway tree. IEEE Trans Med Imag-
workshop on pulmonary image analysis. London, pp 297–307 ing 24:1540–1547
19. Lo P, van Ginneken B, Reinhardt J et al (2009) Extraction of 25. Mori K, Hasegawa J, Suenaga Y et al (2000) Automated anatomical
airways from CT. In: Brown M, de Bruijne M, van Ginneken B labeling of the bronchial branch and its application to the virtual
et al (eds) Proceedings of the second international workshop on bronchoscopy system. IEEE Trans Med Imaging 19:103–114
pulmonary image analysis. London, pp 175–189 26. Kiraly A, Odry B, Godoy M et al (2008) Computer-aided diagno-
20. Irving B, Goussard P, Gie R et al (2011) Segmentation of sis of the airways: beyond nodule detection. J Thorac Imaging
obstructed airway branches in CT using airway topology and 23:105