IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 16, NO. 5, SEPTEMBER 2012 823
A State-of-the-Art Review on Segmentation
Algorithms in Intravascular Ultrasound
(IVUS) Images
Amin Katouzian, Elsa D. Angelini, Member, IEEE,St
´
ephane G. Carlier, Jasjit S. Suri, Nassir Navab,
and Andrew F. Laine, Fellow, IEEE
Abstract—Over the past two decades, intravascular ultrasound
(IVUS) image segmentation has remained a challenge for re-
searchers while the use of this imaging modality is rapidly growing
in catheterization procedures and in research studies. IVUS pro-
vides cross-sectional grayscale images of the arterial wall and the
extent of atherosclerotic plaques with high spatial resolution in real
time. In this paper, we review recently developed image processing
methods for the detection of media–adventitia and luminal borders
in IVUS images acquired with different transducers operating at
frequencies ranging from 20 to 45 MHz. We discuss methodological
challenges, lack of diversity in reported datasets, and weaknesses of
quantification metrics that make IVUS segmentation still an open
problem despite all efforts. In conclusion, we call for a common ref-
erence database, validation metrics, and ground-truth definition
with which new and existing algorithms could be benchmarked.
Index Terms—Intravascular ultrasound (IVUS), lumen, media–
adventitia (MA), segmentation.
I. INTRODUCTION
A. Medical Background
F
OR more than 30 years after its introduction by Andreas
Gr
¨
untzig in 1977, percutaneous coronary interventions
(PCI) remain the most widely used methods by interventional
cardiologists to treat coronary artery disease. Lumenology was
initially used for guidance of the interventions based on X-
ray angiography [1], which had been accidently discovered by
Sones and later deployed for coronary catheterization [2]. Inven-
tion and refinement of intravascular ultrasound (IVUS) imaging
has introduced in vivo “histological” assessment of coronary
atherosclerosis and plaques. As an alternative, invasive coronary
angiography depicts planar projections of the contrast-filled lu-
men. Important quantitative information such as cross-sectional
Manuscript received September 30, 2011; revised January 18, 2012; accepted
February 19, 2012. Date of publication February 28, 2012; date of current ver-
sion September 20, 2012.
A. Katouzian is with the Technical University of Munich, Munich 80333,
Germany, and also with Heffner Biomedical Engineering Imaging Laboratory,
Columbia University, NY 10027 USA (e-mail: amin.katouzian@cs.tum.edu).
E. D. Angelini is with the Institut Telecom, Telecom ParisTech, CNRS LTCI
Paris 75013, France (e-mail: elsa.angelini@telecom-paristech.fr).
S. G. Carlier is with the Department of Cardiology, Universitair Ziekenhuis,
Brussel 1090, Belgium (e-mail: sgcarlier@hotmail.com).
J. S. Suri is with Global Biomedical Technologies, Inc., CA 95661 USA, and
also with Idaho State University, ID 83209 USA (e-mail: jsuri@comcast.net).
N. Navab is with the Technical University of Munich, Munich 80333,
Germany (e-mail: navab@cs.tum.edu).
A. F. Laine is with the Departments of Biomedical and Radiology, Columbia
University, New York, NY 10027 USA (e-mail: laine@columbia.edu).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TITB.2012.2189408
lumen area, thickness, distribution, and composition of plaques
and remodeling of the vessel wall are only revealed by IVUS
that can improve and guide PCI. Moreover, it enables the mon-
itoring of regression and progression of plaques by measuring
changes overtime of the atheroma volumewithin the vessel wall,
especially when evaluating new pharmacological compounds.
Besides, for chronic disease such as atherosclerosis that may re-
occur after balloon angioplasty, atherectomy, stenting, or bypass
surgery, the accurate diagnosis of vulnerable plaques [3]–[7] is
critical. In brief, what makes atherosclerosis one of the dead-
liest disease is not the stenoses alone but failure in detection
and proper treatment of the vulnerable plaques that will lead to
myocardial infarction. This point has motivated researchers to
develop novel imaging modalities such as IVUS, optical coher-
ence tomography (OCT) [8], or near-infrared signals (NIR) [9]
to characterize atherosclerotic plaque components and identify
vulnerable ones.
B. Specificities of IVUS Images
Angiography is the only system routinely used in all PCI
procedures. It provides immediate visualization of stenoses and
guides interventional cardiologists to advance and deploy bal-
loons and stents. However, it suffers from the lack of adequate
geometrical and pathological information on plaque burden size
and composition. So far, IVUS remains the most favorable imag-
ing modality for coronary plaques for the following reasons.
1) It provides real-time cross-sectional grayscale images of
the arterial wall, including morphological and pathologi-
cal structures. Image resolution and signal penetration are
sufficient to allow precise tomographic assessment of the
coronaries.
2) IVUS grayscale images combined with the processing of
radiofrequency backscattered signals can be employed for
further characterization of plaques and the identification
of vulnerable ones [10].
3) Interventional cardiologists can make therapeutic deci-
sions from IVUS images, such as:
a) the need for further treatment (angioplasty, stent im-
plantation, and bypass);
b) the exact spatial location for angioplasty and stent
implantation;
c) the evaluation of the outcome of an angioplasty or
stenting procedure;
d) the need for aggressive management of risk factors
prior to onset of symptoms and advanced disease
state;
e) the predictors of transplant coronary artery disease.
1089-7771/$31.00 © 2012 IEEE
824 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 16, NO. 5, SEPTEMBER 2012
Fig. 1. Schematic of four different types of imaging catheters. (a) Catheter
without guide wire rail. (b) Catheter with a guide wire rail on the side.
(c) Catheter with a guide wire rail at its center. (d) Catheter with a
guide wire rail at its center, an inflatable balloon, and a stent. Reference:
http://ee.isikun.edu.tr/research.asp-page = projects_files.
Comparing with other modalities, the scientific and diagnostic
advantages of IVUS are clear. The American College of Cardi-
ology and the American Heart Association (ACC/AHA) guide-
lines for PCI state that “the limitations of coronary angiography
for diagnosis and interventional procedures can be reduced by
the use of adjunctive technology such as intracoronary ultra-
sound imaging” and that IVUS can improve PCI methods and
outcomes [11]. Additionally, IVUS is the primary screening
choice for validation of novel endovascular coronary imaging
modalities (i.e., OCT and NIR). It has been widely used in
prospective trials to investigate the efficacy of new endovascu-
lar devices or drugs. With respect to the detection of vulnerable
plaques, IVUS could become the most common and trustwor-
thy screening technique if reliable and reproducible image or
signal processing methods are provided for quantitative plaque
characterization. Regarding vulnerable plaques, IVUS could en-
compass a greater predictive value in detecting them with the
combination of morphological features (thin-cap fibro atheroma
(TCFA), lipid core size, and calcification patterns) that are all
detectable and measurable at once, which is not the case for
competitive imaging modalities such as NIR and OCT. Among
the morphological features, TCFA (<65 μm) is better depicted
and measured on OCT images with a higher resolution (∼10
μm) comparing to IVUS (∼120 μm). This limitation may be
resolved in the future, developing ultrahigh-frequency IVUS
transducers that provide images with higher axial resolutions.
C. IVUS Acquisition Systems
The IVUS acquisition system consists of a catheter, a pullback
device, and a scanning console.
1) IVUS Catheter: The IVUS catheter carries an ultrasound
transducer that can be combined with an inflatable balloon,
with or without a stent, for imaging assistance and expansion
of narrowed areas, as illustrated in Fig. 1. The IVUS catheter
is 150 cm long, and has a tip size of 3.2–3.5 F (1.2–1.5 mm)
that can go through 5–6 F guiding catheter. It may be used
to visualize over 15 cm of a coronary artery. The imaging field
goes up to 15–20 mm, well enough for coronary arteries, ranging
from 4 to 5 mm in diameter on average at the level of the left
main artery down to 2 mm in the smallest segment considered
for therapeutic intervention (balloon angioplasty and stenting).
The catheter is typically advanced within the femoral artery
toward coronary arteries and site of occlusion under angiogram
guidance. The catheter is visible in angiographic images and
is advanced along with a guide wire. The guide wire rail is
positioned next to the catheter plastic sheath, as in Fig. 1(b),
Fig. 2. (a) Single-element mechanically rotating focused IVUS transducer and
its beam shape. (b) Multi-element phased-array IVUS transducer and its beam
shape.
or within its center, as in Fig. 1(c). The advantage of the latter
design is that there is no guide wire artifact in the reconstructed
grayscale ultrasound images, but at the cost of usually stiffer,
thicker, and less flexible catheters.
2) IVUS Transducer: Currently, there are two types of IVUS
transducers commercially available regardless of their nominal
center frequencies. The main difference relies in the transmit and
receive modes for monitoring ultrasound signals, which are il-
lustrated in Fig. 2. The first system, illustrated in Fig. 2(a), uses a
single-element mechanically rotating focused IVUS transducer
(e.g., Atlantis, Boston Scientific imaging catheter) that rotates
at approximately 1800 revolutions/min. For a 40-MHz trans-
ducer, the axial and lateral resolutions of the beam are about
80–100 and 200–250 μm, respectively. The transducer sends
an ultrasound pulse and receives the backscattered signals. The
transducer is surrounded by a plastic sheath and a syringe is used
to flush saline water inside the sheath to remove air bubbles and
obtain high-quality IVUS images.
The second system, illustrated in Fig. 2(b), uses a multiele-
ment phased-array transducer (e.g., Eagle Eye Gold, Volcano
imaging catheter). An electronic board controls a subset of el-
ements to send several ultrasound pulses at once and receive
the backscattered signals. These circular array systems use syn-
thetic aperture processing to produce images with higher lateral
resolution than single-element transducers.
3) Catheter Pullback Device: The catheter is first manually
advanced to the distal end of the coronary (typically after the
stenoses location) and is then pulled back, manually or with an
automatic pullback system, at a speed of 0.5–1 mm/s.
4) IVUS Scanning Consoles: A scanning console carries a
computer that is used for postprocessing and storage of recorded
IVUS data. A cable from the end of the pullback device is con-
nected, via a dedicated port, with a computer for data processing.
During the catheterization procedure, the clinician uses a key-
board and functional buttons to enter the patient information,
determine the percentage of stenoses, and apply image pro-
cessing and possibly tissue characterization techniques to better
understand and evaluate atherosclerotic plaques.
D. IVUS Image Formation and Display
IVUS transducers operate at different frequencies, depend-
ing on the manufacturer. Fig. 3 displays the schematic of an
artery, an IVUS catheter, and four distinct IVUS image frames
acquired with transducers with different center frequencies. As
KATOUZIAN et al.: A STATE-OF-THE-ART REVIEW ON SEGMENTATION ALGORITHMS IN INTRAVASCULAR ULTRASOUND (IVUS) IMAGES 825
Fig. 3. (a) Schematic of an artery, catheter, atherosclerotic plaque, and
IVUS image cross section (reference: http://www.bmj.com). (b) Cross-sectional
anatomy of the arterial wall. Four distinct IVUS frames acquired with
(c) 20-MHz, (d) 30-MHz, (e) 40-MHz, and (f) 45-MHz transducers. Green
and red borders represent the vessel wall (MA) and lumen (intima) borders,
respectively. The yellow dashed line depicts the trajectory of transducer scan
lines.
illustrated, at higher center frequency, spatial resolution is im-
proved, at the cost of more scattering from red blood cells inside
the lumen. It is worth mentioning that the axial and lateral res-
olutions depend on the transducer center frequency and beam
width, respectively.
During acquisition, IVUS backscattered radiofrequency (RF)
signals that are continuous-time real-valued and band-limited
signals, x(t), are digitized x(nT
s
)=x
n
at periodic time inter-
vals of T
s
= f
−1
s
and stored in the hard disk of a computer. f
s
is the sampling rate of the digitizing board and may vary from
one system to another. For example, in the Boston Scientific
(Fremont, CA) Galaxy or iLab imaging systems, the acquisi-
tion boards sample IVUS signals at the rate of f
s
= 400 MHz
whereas the sampling rate for the Volcano s5
TM
imaging sys-
tem is f
s
= 200 MHz. Once the IVUS backscattered signals are
digitized, numbers of steps need to be taken in order to convert
digitized RF signals into typical IVUS grayscale images. First,
the envelope of each RF signal (A-line) is computed to generate
a corresponding analytical signal [12]. This is followed by dec-
imation and interpolation along the axial and lateral directions,
respectively. Log compression is also used to enhance image
quality followed by a quantization (e.g., 8 bit).
As depicted in Fig. 3(a), the transducer has a spiral trajectory
(yellow dashed line) while acquiring cross-sectional grayscale
images. The original domain of acquisition is polar (r, θ) and
the resulting grayscale image is transformed to (x, y) Carte-
sian coordinates to reconstruct a typical IVUS frame. Planar
cuts through stack of cross-sectional images provide longitudi-
nal views of an artery. Interventional cardiologists can assess
the length of an artery and the distribution of atherosclerotic
plaques within this pullback direction. Fig. 4 illustrates an IVUS
grayscale image in polar and Cartesian coordinates along with
an example of a cut in the longitudinal pullback direction.
E. IVUS Image Artifacts
IVUS images may suffer from severe acquisition artifacts.
We can cite five main artifacts: presence of the guide wire,
Fig. 4. IVUS grayscale image in (a) polar (r, θ)and(b)(x, y) Cartesian
domains. (c) Longitudinal display along an arbitrary planar cut identified as the
yellow line in (b).
ring-down, nonuniform rotational distortion (NURD), reverber-
ation, and discontinuity at 0
◦
in the Cartesian domain. When a
guide wire rail is designed along with a plastic sheath of the
catheter, it obstructs the propagation of ultrasound signals, re-
sulting in shadowing behind the guide wire, as illustrated in
Fig. 5(a). The second artifact arises from the repetitive reflec-
tions of the ultrasound signal from the surface of the transducer
because of impedance mismatch, as illustrated in Fig. 5(b). The
NURD artifact is due to a mechanical glitch in the driving shaft
or the binding of the catheter in curved arteries, as illustrated
in Fig. 5(d). The fourth artifact, known as reverberation, cor-
responds to oscillations of the ultrasound signals between the
transducers and the arc of calcified plaques, which causes mul-
tiple appearances of calcified arcs, as illustrated in Fig. 5(c).
The last artifact corresponds to a discontinuity of tissue appear-
ance at 0
◦
in the Cartesian domain due to the spiral trajectory
of the transducer as well as severe catheter or heart motions, as
illustrated in Fig. 5(a).
F. In Vivo Data Collection
Generally, an IVUS catheter is advanced into the left or
right coronary artery and possibly in some side branches on
a guide wire coming out of a guiding catheter inserted in the
femoral artery. Acquisition of cross-sectional ultrasound images
of the right coronary arteries (RCA), left anterior descending
(LAD), and left circumflex (LCX) coronary arteries can be per-
formed with a rotating single-element transducer or a phased-
array transducer. The catheter pullback speed varies between
0.5 and 1 mm/s and the frame rate can be set to 30–60 frames/s.
The IVUS RF data and images are acquired as described in
Section I-D.
G. Image Processing Challenges
During a catheterization procedure, hundreds to thousands
of IVUS images are recorded. Therefore, automatic detection
of the vessel wall [media–adventitia (MA)] and luminal bor-
ders is required to quantify the degree of stenoses and measure
826 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 16, NO. 5, SEPTEMBER 2012
Fig. 5. IVUS image artifacts. (a) Guide wire artifact and discontinuity artifact at 0
◦
. (b) Ring-down artifact. (c) Ring-down artifact and reverberation artifact.
(d) NURD artifact. Calibration markers (small white squares) in (b) are used for measurements.
Fig. 6. IVUS grayscale image (a) without TGC adjustment and (b) with TGC
adjustment.
Fig. 7. IVUS grayscale image with (a) manually traced vessel wall (green)
and lumen (red) borders, (b) zoomed-in region with anatomical structures of
the artery such as lumen, intima (I) or plaque, media (M), and adventitia (A),
(c) histology image of artery anatomical structures.
the luminal area in which blood flows. The lumen border is at
the innermost surface of atherosclerotic plaques. Since ultra-
sound signals are progressively weakened with depth, time gain
compensation (TGC) can be applied to compensate for this, as
illustrated in Fig. 6. The vessel wall border, also called the ex-
ternal elastic membrane (EEM) border, is a contour drawn at
the interface between the media and the adventitia. Made of
smooth muscle cells, the media does not reflect the ultrasound
signal and appears as a dark ring. Adventitia is the outer layer of
an artery, formed of sheets layers that are hyperechogenic and
appear as a bright region. Fig. 7 illustrates the borders, the cor-
responding anatomical structures in an IVUS grayscale image
and a histology image of an artery.
Generally speaking, detection of vessel wall borders is less
difficult than that of lumen borders since the vessel media con-
sists of smooth muscle cells and does not reflect ultrasound
signals. It appears as a dark region on IVUS images, which can
be used as a marker to detect the vessel wall. In contrast, due
to high scattering from red blood cells inside the lumen, the
detection of the luminal border is more challenging, especially
when a high-frequency transducer is used. Comparing IVUS
ultrasound probes, the lumen border is more easily detected in
images acquired with a 64-element phased-array 20-MHz trans-
ducer than with a single-element mechanically rotating 45-MHz
transducer, as illustrated in Fig. 3.
In any case, clinical applications of automated segmentation
methods have seen limited success due to several intrinsic arti-
facts (presence of the guide wire, presence of calcified plaques,
presence of side branches, motion of the catheter and the heart)
and extrinsic parameters (such as manual setting of TGC). For
example, the presence of the guide wire, calcified plaques, and
side branches significantly affects an algorithm performance,
particularly when deformable models are employed. On the
other hand, variability among system specifications or changes
of acquisition parameters by an expert would lead to incon-
sistency among datasets so that supervised techniques or those
that rely on statistical properties of gray level intensities may
not perform efficiently.
In this paper, we review state-of-the-art segmentation algo-
rithms that detect either both borders simultaneously or one of
them. These methods can be categorized based on their clinical
application, domain of analysis, transducer center frequency, di-
mensionality, and underlying image processing framework, as
summarized in Table I. To tackle the segmentation problem, re-
searchers have developed several algorithms, employing differ-
ent techniques such as graph searching, gradient-driven methods
with dynamic programming, deformable models in combina-
tion with statistical properties of grayscale values correspond-
ing to blood and nonblood regions, statistical shape models,
probabilistic approaches, edge-enhancement frameworks along
with active contours, or multiscale techniques, for example.
We distinguished three main families of approaches: 1) direct
detection of border(s); 2) blood speckle reduction (i.e., spa-
tiotemporal filtering) as a preprocessing step prior to border
detection; and 3) supervised classification [e.g., support vector
machine (SVM)] of blood versus nonblood regions by extract-
ing appropriate spatial/temporal/spectral features. For each fam-
ily of methods, we describe the main principles, performance,
KATOUZIAN et al.: A STATE-OF-THE-ART REVIEW ON SEGMENTATION ALGORITHMS IN INTRAVASCULAR ULTRASOUND (IVUS) IMAGES 827
TABLE I
L
IST OF SOME OF EXISTING SEGMENTATION ALGORITHMS IN IVUS IMAGES AND THEIR SPECIFICATIONS
advantages, and limitations. Particular attention was also paid to
the level of automation of each technique. Some methodologies,
lack complete automation and require user interaction in order
to perform the segmentation tasks.
II. D
IRECT DETECTION OF BORDER(S)
A. Edge-Tracking and Gradient-Based Techniques
The interactions of IVUS signal with the blood–tissue inter-
face and smooth muscle cells in media give rise to typical edge
patterns that could be used to distinguish lumen and MA con-
tours, respectively. In practice, these patterns seldom embody
clean borders due to scattering effects within the lumen, dis-
continuity in intensity values, drops in edge reflections, noise.
Hence, further refinement (e.g., smoothing for noise reduction)
and hybrid algorithms were designed to assemble edge fea-
tures into desirable target boundaries. The IVUS segmentation
techniques that deploy such image descriptors usually require
precise initialization and rely on an energy minimization frame-
work. The very first work on IVUS border detection from Her-
rington et al. [13] developed a semiautomated algorithm based
on such principle. Later, Sonka et al. [14] introduced one of the
earliest comprehensive works on the detection of internal and
external elastic laminae borders as well as lumen borders. The
internal and external elastic laminae borders refer to the inner
and outer layers of the media, which consist of smooth muscle
cells. Normally, the MA border can be drawn anywhere be-
tween these two borders (within the corresponding hypoechoic
region). After removing the calibration markers, illustrated in
Fig. 5(b), regions of interest (ROIs) were interactively selected
and Sobel-like edge detectors were applied on subimages to
construct laminae and lumen border graphs. A heuristic graph
search technique [15], [16] was then performed deploying two
distinct cost functions to detect the borders. The key point for
precise identification of borders was to define appropriate cost
functions for each border by incorporating aprioriknowledge
such as shape models and edge patterns. The results demon-
strated good correlation between manual and automated lumen
borders (r = 0.96), plaques (r = 0.95), and stenoses areas (r
= 0.93). Although the presented technique required some user
interaction and was only applied on in vitro images using cir-
culating saline water, where there was not much scattering in
the lumen area compared to in vivo images, the results were en-
couraging and this study raised attentions toward this particular
problem. An extended version of this approach using a different
cost function and fully 3-D graph search has been presented in
[17].
With a similar type of approach, the authors in [18] and [19]
presented a semi-automated methodology using dynamic pro-
gramming to find the optimal path within the vessel and detect
both MA and lumen borders in polar coordinates from delineated
contours. The minimal path search was performed between
end points interactively selected in reconstructed longitudinal
images at the intersection of two perpendicular cut planes. The
results were validated using both tubular phantom data and
in vivo images acquired with a 30-MHz transducer [18]. They
studied inter- and intraobserver variability and showed high
consistency of the method. They later evaluated the performance
