Ultrasound intima-media thickness measurement of the carotid artery using ant

colony optimization combined with a curvelet-based orientation-selective filter

Hao Li, Shijie Zhang, Rui Ma, Huiren Chen, Shui Xi, Jue Zhang, and Jing Fang
Citation: Medical Physics 43, 1795 (2016); doi: 10.1118/1.4943567
View online: http://dx.doi.org/10.1118/1.4943567
View Table of Contents: http://scitation.aip.org/content/aapm/journal/medphys/43/4?ver=pdfcov
Published by the American Association of Physicists in Medicine
Articles you may be interested in
Semiautomatic segmentation of atherosclerotic carotid artery wall volume using 3D ultrasound imaging
Med. Phys. 42, 2029 (2015); 10.1118/1.4915925
Optic disc detection using ant colony optimization
AIP Conf. Proc. 1479, 798 (2012); 10.1063/1.4756258
Nonrigid registration of three-dimensional ultrasound and magnetic resonance images of the carotid arteries
Med. Phys. 36, 373 (2009); 10.1118/1.3056458
Segmentation of carotid artery in ultrasound images: Method development and evaluation technique
Med. Phys. 27, 1961 (2000); 10.1118/1.1287111
Accuracy and variability assessment of a semiautomatic technique for segmentation of the carotid arteries
from three-dimensional ultrasound images
Med. Phys. 27, 1333 (2000); 10.1118/1.599014

Ultrasound intima-media thickness measurement of the carotid

artery using ant colony optimization combined with a curvelet-based
orientation-selective filter
Hao Li and Shijie Zhang
Academy of Advanced Interdisciplinary Studies, Peking University, Beijing 100871, China

Rui Ma, Huiren Chen, and Shui Xi

VINNO Technology Co., Ltd., Suzhou 215123, China

Jue Zhanga) and Jing Fang

Academy of Advanced Interdisciplinary Studies, Peking University, Beijing 100871, China
and College of Engineering, Peking University, Beijing 100871, China

(Received 6 September 2015; revised 4 February 2016; accepted for publication 26 February 2016;
published 22 March 2016)
Purpose: Automatic measurement of the intima-media thickness (IMT) from ultrasound carotid
images is an important task in clinical diagnosis. Many computer-based techniques for IMT measurement have been proposed to overcome the limits of manual segmentation. However, the robustness
of the algorithms would be influenced by the inherent speckle noise of ultrasound image. This paper
proposed a curvelet guided ant colony optimization (CGACO) strategy that could achieve satisfied
accuracy for IMT measurement with improved robustness to noise.
Methods: The curvelet-based orientation-selective (CBOS) filter was first introduced for speckle
removal and edge enhancement. Different from conventional methods, CBOS filter processes the
curvelet coefficients by orientations rather than by magnitude. Then, a specially designed two-leg
ant colony optimization technique, combined with Otsu thresholding and Sobel edge detector, was
proposed as a novel segmentation method to extract the media-adventitia (MA) and the lumen-intima
(LI) boundaries. Finally, a coupled snake model was employed to further smooth the contours of MA
and LI.
Results: In addition to 224 carotid artery images acquired from 34 participants, simulated speckled
images with nine levels of noise were also included in the database. The mean absolute distance
errors of CGACO for LI interface tracings, MA interface tracings, and IMT measurements were
0.030 0.027, 0.039 0.036, and 0.041 0.036 mm, respectively. Besides, CGACO had a correlation
coefficient as high as 0.992 and a bias as low as 0.008. All these measures were comparable
to or better than a previous technique and the manual segmentation. On the other hand, CGACO
had the highest success rate of 98.7% in the segmentation of real data. It also maintained a
much higher success rate in the segmentation of simulated images with different levels of speckle
noise.
Conclusions: The proposed technique showed accurate IMT measurement results. Furthermore,
benefiting from the CBOS filter, the robustness to noise of the algorithm was substantially
improved. Therefore, CGACO could provide a reliable way to segment the carotid artery
from ultrasound images and could be used in clinical practice of IMT measurement, particularly in early atherosclerotic stages. C 2016 American Association of Physicists in Medicine.
[http://dx.doi.org/10.1118/1.4943567]
Key words: intima-media thickness, curvelet transform, ant colony optimization, image
segmentation, ultrasound imaging

1. INTRODUCTION
Cardiovascular diseases (CVDs) are the leading cause of
morbidity and mortality in industrialized nations.1 The atherosclerotic process, which causes thickening of the artery wall
and reduction of the artery lumen, is the early manifestation
of the possible onset of CVDs.2,3 Studies have shown that the
increase in the intima-media thickness (IMT) of the carotid
artery is one of the prominent clinical signs of an ongoing
1795

Med. Phys. 43 (4), April 2016

atherosclerotic process.4,5 Nowadays, the measurement of the

IMT is widely recognized as a useful tool for evaluation of
cardiovascular risk.6,7
B-mode ultrasound of carotid arteries provides a noninvasive tool for estimating the IMT of carotid arteries.8 For
measuring IMT, the manual tracing method is commonly
used to delineate the boundaries of the intima-media complex.
However, the manual approach requires substantial experience
and can be quite time-consuming and tedious.9 Moreover,

0094-2405/2016/43(4)/1795/13/$30.00

2016 Am. Assoc. Phys. Med.

1795

1796

Li et al.: Intima-media thickness measurement in ultrasound images

the manual segmentation is based on subjective assessment

of operators, which inevitably results in intraobserver and
interobserver variability.10,11
Different techniques have been proposed to perform automatic IMT measurement, including edge detection,1214 dynamic programming,10,15,16 active contour models,1724 Hough
transform,20,25 Nakagami modeling,26 and neural networks.27
However, previous studies mostly focused on the accuracy of
IMT measurement, but gave less consideration to the robustness. Generally, ultrasound images suffer from serious noise
and relatively poor anatomical details, which could produce
images with low signal to noise ratio. Severe noise in the
lumen region, the adventitia layer, and the intima-media interface could all create misleading edges and lead to segmentation errors. In some studies,19,28,29 images selected to test
are required to achieve acceptable quality and clear vessel
wall layers, while images with excessive noise superimposed
would be discarded. Some other studies10,18 require manual
correction for the incorrect segmentation, which would reduce
the operational efficiency.
Therefore, the unsatisfied robustness has become a major
challenge for clinical applications of automatic IMT measurement. The purpose of this study is to develop a more effective strategy, which could achieve satisfied accuracy for IMT
measurement with improved robustness to noise.
Actually, the carotid artery wall roughly goes in a certain
direction in most ultrasound images, which provides us an
alternative perspective in the image enhancement procedure.
In recent years, curvelet transform has received more concern
due to its multidirectional characteristics.30 Principally,
curvelet transform is a multiscale pyramid with many directions and positions at each length scale, and it provides a valuable solution to overcome the weakness of the wavelet transform in sparsely representing lines, curves, and edges. In this
study, based on its high directional sensitivity, curvelet transform was introduced in preprocessing procedure to enhance
edges of the intima-media layers and remove speckle noise, so
as to potentially improve the robustness of IMT measurement.
Distinct from the conventional method based on curvelet transform,31 we proposed a curvelet-based orientation-selective
(CBOS) filter, which processes the curvelet coefficients by
orientations rather than by magnitude.
Subsequently, considering that the intima-media complex
is a double-layered structure with two parallel interfaces, a
specially designed two-leg ant colony optimization (ACO)
technique was utilized to automatically track the lumen-intima
(LI) and the media-adventitia (MA) boundaries. The concept
of ant colony system (ACO), inspired by the behavior of real
ants, was initially introduced by Colorni et al.3234 In ACO,
solutions of the problem are constructed within a stochastic
iterative process, by adding solution components to partial
solutions. In our attempt, after initial threshold based edge
detection, an ant with two legs is designed to move from one
side of the vessel wall to the other side, so as to extract both
the LI and MA boundaries. Finally, the coupled snake model
proposed by Zhou et al.16 was employed to further smooth
the contour of LI and MA and improve the accuracy of IMT
estimation.
Medical Physics, Vol. 43, No. 4, April 2016

1796

We named the proposed algorithm as curvelet guided ant

colony optimization (CGACO). It was validated by direct
comparison with the ground truth (GT) results, which refer
to the average of three manual tracings performed by one
operator. Besides, typical ultrasound carotid images added
with different levels of simulated noise were used to evaluate
the robustness to noise. CGACO was also systematically
compared on an equal footing with several other approaches,
including the same ACO based algorithm with CBOS filter
removed, another promising method proposed by Zhou et al.16
and the manual tracings performed by another operator.
2. METHOD
Figure 1 presents the flowchart of the proposed method.
First, the CBOS filter and scale multiplication procedures
are applied for obtaining an edge map that highlight LI and
MA from the original ultrasound images. Then, discontinuous
initial edges are obtained by edge detection procedure. Subsequently, edge tracking based on ACO is used to connect the
discontinuous edges. After the contour refinement based on
snake model, the IMT can be finally measured.
2.A. Orientation-selective filter based
on curvelet transform

The second-generation discrete curvelet transform was

applied in this paper, which was implemented in four steps:
1.
2.
3.
4.

Apply the 2D fast Fourier transform to the image;

obtain the product of the angle and scale windows;
wrap the product around the origin; and
apply the 2D inverse fast Fourier transform.

Unlike traditional wavelets, curvelets are localized not

only in position (the spatial domain) and scale (the frequency
domain) but also in orientation. Figure 2(b) shows the curvelet
decomposition of the ultrasound image shown as Fig. 2(a). The
curvelet coefficients are grouped according to orientations and
scales. The concentric coronae represent the different scales,
starting from the lowest scale in the center.
As shown in Fig. 2(c), the anatomical information of
intima-media layers mainly distributed in the sub-bands which
correspond to the orientations vertical to the vessel, whereas
the parallel sub-bands contain little tissue information except
for some scattered points [Fig. 2(d)]. This distribution enables
us to set different gain factors for the sub-bands of different
orientations. In our proposed CBOS filter, coefficients in the
sub-bands vertical to the vessel are multiplied by a large gain
factor, while the sub-bands parallel to the vessel are multiplied
by a small gain factor. In this way, the edges of the intimamedia layers, which distribute in the vertical sub-bands, can
be enhanced, and speckle noise can be suppressed.
The steps of CBOS filter are as follows:
2.A.1. Determining the interesting region
and orientation

Similar with the method used by Cheng et al.,17 we need to

manually choose the starting point and end point, Ps and Pe ,

1797

Li et al.: Intima-media thickness measurement in ultrasound images

1797

F. 1. The flowchart of the proposed approach.

which enclose the ROI for IMT measurement. The two points
need to be located on or near the intima. Let Ps
= (x s , ys ) and Pe = (x e , ye ). Thus an ROI can be determined by
WR = Rec[(x S , ymin c),(x e , ymax + c)], where Rec represents
a rectangular area defined by the left-top and right-bottom
coordinates, c is a constant to ensure that the ROI includes
all of the possible vessel wall layers, and ymin and ymax are
the minimum and of maximum of ys and ye , respectively.
Besides, the orientation of carotid artery can be determined by
calculating the slope k = ( ye ys )/(x e x s ). Here, we chose
c = 40 pixels, which refers to 2.5 mm in this study.

coefficients Cm, n (i, j) are multiplied by the gain factors G L and

G H separately,
C1,1 (i, j) = C1,1 (i, j) G L ,
Cm, n (i, j) = Cm, n (i, j) G H (m,n), m, n > 1,

(1)
(2)

where G L is a constant, G H is given by a function associated

with scale index m, and orientation index n,
G H (m,n) = ae(

n AB 2
C

),

(3)

where
2.A.2. Decomposition in the curvelet domain

The ROI is decomposed in the curvelet domain to obtain

low-frequency sub-band and different scales of high-frequency
sub-bands. In our study, the number of scales including the
coarsest wavelet level is set to 6, and the number of angles at
the second coarsest level is set to 16.
2.A.3. Gain function

After the curvelet transform decomposition, low-frequency

sub-band coefficients C1,1 (i, j) and high-frequency sub-bands
Medical Physics, Vol. 43, No. 4, April 2016

2 2 +2
tan1 k,

m
m

2 2 + 0.5, 1 n 32 2

m
m
m
B=
52 2 + 0.5, 32 2 < n 72 2 ,

92 m2 + 0.5, 72 m2 < n 82 m2

m
C = b2 2 ,
A=

(4)

(5)

(6)

where m/2 denotes the round-up of m/2. a is a parameter

controlling the amplitude of the peaks and b is a parameter

1798

Li et al.: Intima-media thickness measurement in ultrasound images

1798

F. 2. Curvelet coefficients intensity of the test image. (a) The test carotid ultrasound image. (b) The curvelet decomposition of Fig. 2(a) into four frequency
scales with 16 orientations at the second coarsest level. (c) The sub-band of orientation index 4 at the fourth coarsest level (solid border), which corresponds to
the orientation vertical to the vessel. (d) The sub-band of orientation index 13 at the fourth coarsest level (dashed border), which corresponds to the orientation
parallel to the vessel.

controlling the widths of the peaks. Figure 3 demonstrates the

gain factors when typical scale index m = 3 and the slope k = 0.

where
G1 (x, y) =

2.A.4. Reconstruction

The processed coefficients were reconstructed to obtain the

enhanced image.
2.B. Scale multiplication

G i
max 0,
I (x, y) ,
y
i=1

x 2+y 2
2 12

, G2 ( y) =

y 2

e 222 ,

22

(8)

where G1 (x, y) is a 2D Gaussian function with a small scale

1 and G2 ( y) is a 1D Gaussian function with a large scale
2.
2.C. Initial edge detection

To further improve the border detection accuracy, an edge

map was computed to suppress the background information
and indicate the location of the intima-media layers.16 The
edge map is defined by the product of the responses from the
two filters with different scales of Gaussian density kernels. To
reduce the interference of noise, we only select positive step
edges with respect to the y-axis in the edge map,
(x, y) =

1
e
212

(7)

As described above, an edge map with clear LI and MA

boundaries could be obtained by the previous steps [Fig. 4(a)].
Before the ACO based boundary tracking, a coarse edge detection procedure is designed to acquire several initial line segments of the LI and MA boundaries to improve the efficiency
of ACO.
As demonstrated in Fig. 4, Otsu thresholding is first used to
acquire a binary image from the edge map [Fig. 4(b)]. Then,
the vertical Sobel detector is applied to extract the upper and
lower edge lines of LI and MA interfaces [Fig. 4(c)]. The
expected results of this step should have two pairs of contours
(four lines) corresponding to the LI and the MA interfaces.
However, the edge lines achieved by the edge detector always
contain redundant misleading lines or breakages. Therefore,
those parts that have more or less than two pairs of lines (four
lines) along the vertical axis will be discarded [Fig. 4(d)].
The remaining parts are incorporated into two lines by taking
the midpoints of the upper and lower edge lines of each pair
[Fig. 4(e)]. In the next step, the incorporated line segments are
defined as mandatory paths of the ants, while the removed parts
would be filled up by the ACO algorithm.
2.D. Boundary tracking based on ant colony
optimization

F. 3. The gain factors of different orientation indexes when m = 3 and

k = 0.
Medical Physics, Vol. 43, No. 4, April 2016

Generally, the LI and the MA interfaces can be regarded

as two parallel lines. In order to acquire the boundaries of
both LI and MA interfaces, two ants are set to move from

1799

Li et al.: Intima-media thickness measurement in ultrasound images

1799

F. 4. Results of each steps of initial edge detection. (a) The edge map obtained by previous procedures. (b) Binary image acquired by Otsu thresholding. (c)
Edge lines acquired by vertical Sobel detector (white lines for the upper edges of LI and MA interfaces and gray lines for the lower edges). (d) The selected
qualified edge lines. (e) Final edge lines acquired by taking the midpoints of the upper and lower edge lines.

the left side of the vessel wall to the right sides. The two
ants, one fixed on top of the other, would move step by step
synchronously, and thus can be seen as the two legs of one
ant. The line segments obtained by initial edge detection are
defined as mandatory paths in which the ants are required to
move along. The gaps between the segments need to be filled
up by the ACO algorithm described as follows.
The ant moves from left to right, starting from Ps (x s , ys )
that we selected previously. At time t, the leg l (l = 1,2) of
antk moves from the pixel (x,i) to its neighboring pixel (x + 1, j)
according to a transition probability that is defined as

The global updating rule is implemented as follows. Once

all ants have built their tours, pheromone is updated on all
edges according to
(x,i) (t + 1) = (x,i) (t) + (x,i),

(10)

k,l
(t)
p(x,i,
j)

 


(x+1, j) (t) (x+1, j) (t)

k,l
k,l

 
 , j N(x,i),s N(x,i),

(t)
(t)

=
(x+1, s)
(x+1, s)

0,
otherwise,

(9)
where is the pheromone and is the intensity of pixels in the
edge map. and are two parameters determining the relative
influence of the pheromone trail and heuristic information.
k,l
N(x,i)
represents the set of positions that is allowed for
the leg l of antk in the next step. It corresponds to the right
neighborhood pixels of pixel (x,i). It should be pointed out
that any neighborhood pixel cannot be shared by the two legs.
Therefore, the shared right neighborhood needs to be excluded
k,l
from N(x,i)
during evolution. In this way, the upper leg is
determined to be on top of the lower leg and the final contours
will not crossover or overlap.
k,l
Figure 5 shows typical position sets N(x,i)
during two-leg
ant exploration. The ant locates in the abscissa of x and is moving to x + 1. The blue and green blocks denote the permitted
positions. The red block is a forbidden position because it
exists in both the right neighborhoods of the positions of upper
and lower legs.
Medical Physics, Vol. 43, No. 4, April 2016

F. 5. An example of possible positions that the upper and lower legs can
go to. The light grey blocks denote the permitted positions, and the dark grey
block refers to the forbidden position.

1800

Li et al.: Intima-media thickness measurement in ultrasound images

where (x,i) (t) is the amount of pheromone on a given pixel

(x,i) at iteration t and (x,i) (t + 1) is the amount of pheromone
on the pixel (x,i) at next iteration. is a decay constant used to
simulate the evaporation of pheromone. (x,i) is the amount
of deposited pheromone, which is given by
m

Q
,
(x,i) =
C (k)
k=1

(11)

where Q is a given constant and m is the number of ants. C (k)

is the total cost of the path solution of antk , given by
n
1
(1 I N (x i , yi )) + D(k),(x i , yi )
C (k) =
n i=1
tour done by antk ,

(12)

where I N (x, y)(0 I N (x i , yi ) 1) is the normalized grayscale

value of pixel (x i , yi ), and n is the step number of the path. We
define 1 I N (x, y) as the darkness of pixel (x,i) and calculate
the average darkness of the pixels in the path of antk . D(k)
is the distance between the end point of antk and the point
Pe (x e , ye ) that we select in the previous procedure. is the
penalty factor, which denotes a penalty weight for the error
distance D(k).
The steps of ACO are as follows:
1. Initialize the pheromone matrix;
2. place one ant near the starting point Ps (x s , ys ) (two
pixels above for the upper leg and two pixels below for
the lower leg);
(a) the ant moves from Ps (x s , ys ). The number of steps
equals to x e x s ;
(b) calculate the pheromone of pixels on the ants path;
3. repeat step 2 until all ants have finished their tours;
4. evaporate and update the pheromone matrix;
5. repeat steps 25 for i iterations;
6. select the ant path which has the minimum cost C (k) as
the detected boundaries of LI and MA interfaces.
2.E. Curve refinement based on snake model

The contours obtained by ACO are still not desirable

enough due to the unsmooth path explored by two-leg ant. The
coupled snake model, proposed by Zhou et al.,16 was adopted
to smooth the contour. Given the edge map f (x, y) : R+,
the snake model is defined by minimizing
)2
(
2
2

1
yi
E ( y1, y2) =

dx
f (x, yi )dx
2 i=1
x
i=1

smoothing energy

+
2

boundary energy

2
( y2 y1)
dx ,
x

(13)

uniform energy

where y1 (x) and y2 (x) represent the contours of the LI and the
MA interface, respectively. Parameter controls the weight of
the smoothing energy term. The boundary energy refers to the
external force which drives the contours toward the expected
edges. Parameter controls the weight of the uniform energy
Medical Physics, Vol. 43, No. 4, April 2016

1800

term, which connects the two independent contours together

by forcing a relatively uniform distance.

3. EXPERIMENT
3.A. Data acquisition

The image database consisted of 224 2D B-mode longitudinal ultrasound images of the carotid artery. They were
taken from 34 participants (19 male and 15 female, age
46 22): six of them were patients with atherosclerosis, and
the others were healthy participants. Among the 224 B-mode
images, 192 were with normal carotid arteries, 28 had an
increased IMT, and 4 had plaques. In addition to the 224
actual B-mode images, 270 images with different levels of
simulated speckle noise were included as our simulated
database.
All the images were acquired by a Vinno G60 device
[VINNO Technology (Suzhou) Co., Ltd, China] equipped with
a linear probe working in the frequency range 7.314 MHz. All
participants provided written informed consent. These images
were resampled at a standard density of 16 pixels/mm.29 Thus,
the axial resolution was equal to 0.0625 mm. These images
were captured by the ultrasound device and transferred to a
computer (Intel Core i7-3610M, 12GB memory) to perform
the manual and automatic IMT measurement. All the programs
were written in .
One technician and two clinical experts were involved in the
performance evaluation. The technician first chose the starting
and end point on or near the intima of the proximal carotid
artery for each image to determine a rectangular ROI. Then the
IMT measurement procedure was run in the ROI only. Manual
tracings were performed by a previously developed graphical
user interface. For each ROI, the LI and MA boundaries were
manually delineated by two skilled operators. The first operator (OP1) performed three manual tracings and the second
operator (OP2) performed one manual tracing. The GT was
defined as the average profile of OP1s three manual tracings.
OP2s tracing was defined as another manual tracing (MAN)
and was compared with GT.
3.B. Speckled image simulation

Speckled images were simulated to investigate the robustness of the methods under different levels of noise. The speckle
noise model35 can be approximated as
g(x, y) = f (x, y) n(x, y) + a(x, y),

x, y R2,

(14)

where f (x, y) represents the noise-free image, g(x, y) represents the noisy observation of f (x, y), and n(x, y) and a(x, y)
represent the multiplicative and additive noise, respectively,
and are the spatial location indices that belong in the 2-D space
of real numbers.
Since the effect of additive noise is considerably small
compared to the effect of multiplicative noise in ultrasound
images, Eq. (14) can be approximated as
g(x, y) = f (x, y) n(x, y),

x, y R2.

(15)

1801

Li et al.: Intima-media thickness measurement in ultrasound images

Realistic spatially correlated speckle noise in ultrasound

images can be simulated by low-pass filtering a complex
Gaussian random field and taking the magnitude of the filtered
output.36 The low-pass filtering is performed by averaging
the complex values in a sliding window. Such a short-term
correlation was found sufficient to model the realistic images
well.37 In this way, images with different noise levels can be
generated by changing the variance of the complex Gaussian
random field.
In our experiments, 30 images were selected as reference
noise-free images. With the variance varies from 0 to 40,
increases five each time, nine levels of simulated speckle noise
are considered. Finally, we included 270 simulated noisy images as our simulated database.
3.C. Parameter settings

The parameters in the CBOS filter are a = 2, b = 1, and

G L = 0.4. G L is the gain factor of the low-frequency curvelet
coefficients, which should be set in the range [0, 1] in order to
suppress the background signal. a and b are crucial parameters
that control the degree of enhancement and denoising, which
will be discussed later.
In the ACO algorithm, and control the relative importance of intensity of pheromone versus visibility. Therefore, as
a result of high values for , ants tend to choose edges which
previously have been chosen by other ants, where low values of
make the algorithm similar to stochastic algorithms. Experimental results recommend setting = 1 and from 2 to 5.38
In this study, we adopt = 1 and = 4.

1801

According to previous experimental observation,16 the

parameters in the scale multiplication are 1 = 1.5 and 2 = 15,
and the parameters in the snake model are = 0.1 and = 1.4.
The other factors are = 0.6, Q = 1, m = 20, i = 10, and = 1.
The experiment result shows these values worked well enough
for all images that have been tested, and no adjustment for
specific images is required.
3.D. Comparison of the segmentation methods

The CGACO is validated by comparing its performance

with several other methods. First, to evaluate the contribution
of the proposed CBOS filter, the same ACO based algorithm
with the CBOS filter removed was also tested. Second, linked
dual line detection (LDLD), a previous method for automatic
IMT measurement presented by Zhou et al.,16 was involved
in the comparison. Finally, the manual tracings conducted by
OP2 was tested to evaluate the interobserver variability of
manual measurement.

4. RESULT
4.A. Example images

Figure 6 shows sample performance of CGACO on three

different images. Figure 6(a) shows a sample image with
excessive noise. As shown in Fig. 6(b), the speckle noise that
makes the intima-media complex blurry does not preclude
the identification of the carotid artery. Figure 6(c) documents
the case of an artery that presents curvature and Fig. 6(e)

F. 6. Three examples illustrating the performance of CGACO. The left column presents the original images; the right column presents the automatic tracings
overlaid on the original images. [(a) and (b)] Carotid artery image with excessive noise. [(c) and (d)] Carotid artery with curvature. [(e) and (f)] Carotid artery
with thickened intima-media complex.
Medical Physics, Vol. 43, No. 4, April 2016

1802

Li et al.: Intima-media thickness measurement in ultrasound images

T I. Mean absolute errors and standard deviations for CGACO, LDLD,

and manual tracing (OP2).

LI error
MA error
IMT error

Units

CGACO

LDLD

Manual tracing (OP2)

Pixel
mm
Pixel
mm
Pixel
mm

0.48 0.43
0.030 0.027
0.63 0.57
0.039 0.036
0.66 0.58
0.041 0.036

0.51 0.42
0.032 0.026
0.61 0.51
0.038 0.032
0.68 0.55
0.042 0.035

0.57 0.50
0.036 0.031
0.64 0.53
0.040 0.034
0.85 0.73
0.053 0.045

shows the intima-media complex with obvious local thickening. Figures 6(d) and 6(f) show the corresponding correct
performance.
4.B. Accuracy performance

According to the judgment of the operator who carried out

the examination, CGACO, ACO without the CBOS filter, and
LDLD did not correctly identify 3, 12, and 15 images, respectively. The segmentation errors were calculated on the respective remaining images. The mean absolute distance (MAD)
metrics6 were adopted in this paper. This metrics compute the
point-by-point distances between the GT and the automated
computer tracings. The number of columns of each image depends on the size of ROI which the operator chooses. Overall,
about 68 000 error values were calculated.
Table I summarizes the mean absolute segmentation errors
for CGACO, LDLD, and manual tracing performed by another
operator (OP2). The LI and IMT errors of CGACO were lower
than those of LDLD (0.48 0.43 vs 0.51 0.42 mm for LI
errors and 0.660.58 vs 0.680.55 mm for IMT errors), while
the MA error was higher (0.63 0.5 7 vs 0.61 0.51 mm).
Both of two approaches outperformed the manual tracings in
all LI, MA, and IMT errors. All differences was statistically
significant (paired t-test, p < 103).
To investigate the biases in the IMT estimation, we derived
a single value of IMT for each image by averaging the IMT
estimation series. Table II presents a comparison of the IMT
measures from different approaches and the GT. The mean
IMT measurements were calculated from 206 images. The
algorithm without CBOS filter (ACO) was also included in
the comparison. All the present approaches demonstrated very
accurate IMT estimation, with very small biases of mean
IMT values. Among them, LDLD achieved the lowest bias.
CGACO had a correlation coefficient as high as 0.992, whereas

1802

the correlation coefficients of LDLD and ACO without CBOS

filter were 0.989 and 0.988. All the three automatic approaches
outperformed manual tracing, which is 0.986.
Figure 7 reports the scatter diagrams showing the IMT
estimates of CGACO, LDLD, and OP2s manual approach
with respect to GT. Figure 8 shows the BlandAltman plots
comparing CGACO, LDLD, and OP2s manual approach with
GT measurements. It can be observed that although the bias
is slightly bigger, CGACO measurements were more accurate
than those of LDLD. Manual measurements had a bigger bias
and worse agreement compared with automatic approaches.
4.C. Antinoise performance

Success rate (SR) was also shown in Table II. CGACO

correctly segmented 221 of the 224 images, and thus had an
SR as high as 98.7%, whereas LDLD correctly segmented 212
images and had an SR of 94.6%. ACO without CBOS only
correctly segmented 209 and had an SR of 93.3%.
Moreover, 270 simulated speckle images were used to
test and compare the antinoise ability of CGACO, ACO, and
LDLD. Figure 9 shows a sample performance of the three
algorithms under different levels of noise. CGACO traced the
LI and MA interface correctly even when the noise variance
= 40. In this condition, the intima-media complex is difficult
to be identified even by naked eyes. However, ACO and LDLD
failed to trace the correct boundaries of LI and MA interfaces
frequently when simulated noise exists. Figure 10 denotes the
SR of CGACO, ACO, and LDLD under different levels of
noise. The SR of ACO and LDLD declines dramatically when
the noise variance increases, whereas SR of CGACO maintains
in a high success rate.
5. DISCUSSION
We have presented an automatic technique for IMT
measurement in longitudinal ultrasound images and evaluated
its performance. The accuracy performance of CGACO is
comparable to or slightly better than the LDLD approach
proposed by Zhou et al.16 Besides, both CGACO and LDLD
outperformed the manual measurements performed by OP2.
On the other hand, CGACO achieved superior robustness
to noise. It correctly segmented 98.7% of the database (221 of
224 images) without any manual correction, which was higher
than LDLD and ACO without CBOS filter. More remarkable
results can be found on simulated database. As shown in Figs. 9
and 10, the algorithm combined with CBOS filter showed

T II. Mean IMT values, bias, correlation coefficient, and success rate of automatic methods, manual method,
and GT.

IMT
Bias
Corr.a
SR (%)
a

Ground truth

CGACO

ACO

LDLD

Manual tracing (OP2)

0.75 0.24

0.74 0.23
0.008
0.992
98.7

0.73 0.24
0.016
0.988
93.3

0.75 0.23
0.005
0.989
94.6

0.76 0.24
0.009
0.986

Corr. denotes the correlation coefficient between GT IMT values and the measured IMT values.

Medical Physics, Vol. 43, No. 4, April 2016

1803

Li et al.: Intima-media thickness measurement in ultrasound images

1803

F. 7. Regression analysis plots comparing CGACO, LDLD, and OP2s measurements with GT measurements. (a) CGACO. (b) LDLD. (c) OP2.

distinct advantages in antinoise ability evaluation, compared

with LDLD and ACO without CBOS filter.
5.A. About CBOS filter

The superior performance in robustness evaluation is

benefited from the introduction of the CBOS filter and scale
multiplication based on Gaussian filters. The effect of these
procedures can be shown in Fig. 11 in which two sample
images are polluted by extreme simulated speckle noise. Clear
edge maps can still be obtained from the simulated speckle
images, although the intima-media complex has been masked
by speckle noise and can hardly be identified even by naked
eyes.
The proposed CBOS filter is based on curvelet transform, which is not only a multiscale but also multidirectional
transform. Compared with traditional 2D wavelet transform,
curvelet transform could achieve more precise representation
of the image. Nowadays, many image denoising and enhancement algorithms based on curvelet transform have been proposed, and most of them employ thresholding or shrinkage
techniques.31,3942 In those techniques, the high-pass bands
coefficients are subjected to a linear or nonlinear threshold that
suppresses low-amplitude values while retaining or enhancing
high-amplitude values.

However, these techniques process the curvelet coefficients

only according to their amplitudes, thus do not take full advantages of the multidirectional property of curvelets. According to the curvelet coefficients distributions by orientations
shown in Fig. 2, we utilized an orientation-selective filter,
which sets different gain factors for sub-bands corresponding
to different orientations. Equations (3)(6) describes the gain
function for high-frequency sub-bands. Essentially, it is a periodic Gaussian function, which sets maximum gain factors for
sub-bands with orientations vertical to the carotid artery and
minimum gain factors for sub-bands with parallel orientations.
The enhancing effect will be stronger with the increase of a and
the decrease of b, which controls the amplitudes and widths of
the peaks of the Gaussian curves, respectively. Figure 12 shows
the enhancing effect of the CBOS filter under different settings
of a and b.
5.B. About ACO based boundary tracking

With the edge map obtained by CBOS filter and scale

multiplication procedure, segmentation of intima-media layers
become much less difficult. A specially designed two-leg ant
colony optimization technique is then proposed as a novel
segmentation method to extract the LI and the MA boundaries. This approach is based on the consideration that the

F. 8. BlandAltman plots comparing CGACO, LDLD, and OP2s measurements with GT measurements. (a) CGACO. (b) LDLD. (c) OP2.
Medical Physics, Vol. 43, No. 4, April 2016

1804

Li et al.: Intima-media thickness measurement in ultrasound images

1804

F. 9. Sample performances of CGACO, ACO without CBOS filter, and LDLD under different levels of speckle noise.

intima-media complex is a double-layered structure with two

parallel interfaces, which can be considered as the paths of ants
in the ACO algorithm.
The state-transfer probability is determined by the pheromone and the intensity of the surrounding pixels. The amount
of pheromone that ants deposit depends on the darkness (1normalized intensity) of the pixel and the error distance between the end point of the ant and the manually selected
point Pe . The darkness item makes the algorithm considerably
biased toward pixels with high intensity, and the introduction
of error distance could assign the tours that deviate from the
supposed end point a lower weight. In this way, the pheromone
is inclined to accumulate in the LI and MA interfaces and the

paths of ants will gradually approach the real boundaries of LI

and MA interfaces.
Furthermore, the introduction of Otsu thresholding and
Sobel detector can provide initial edge segments, which can
be regarded as the mandatory paths of ants. This strategy
divides the whole exploration into several shorter segments,
and each segment can provide the ants local destinations by
stages. In this way, ants are unlikely to move to wrong ways
and the robustness of the tracking procedure is thus improved.
Moreover, this strategy reduces the length of routes that need
to be explored by ants and improves the time efficiency of
the ACO based boundary tracking. The average time cost of
the segmentation procedure (initial edge detection, boundary
tracking and curve evolution) based on platform was
1.37 s/image. Added with the average time cost of enhancement procedure (CBOS filter and scale multiplication), the
total processing time was 2.63 s/image.
5.C. Comparison of the segmentation methods

F. 10. The success rate of CGACO, ACO without CBOS filter, and LDLD
under different levels of speckle noise.
Medical Physics, Vol. 43, No. 4, April 2016

In addition to the comparison with LDLD in the described

experiment, CGACO also shows superiority in terms of accuracy compared with the recent remarkable studies. Table III
shows the reported LI, MA tracing errors, and IMT measurements reported in several recent studies. It can be observed that
CGACO achieved errors equal to 3027 m for LI, 3936 m
for MA, and 4130 m for IMT, which were smaller than most
of other reported performance.
Moreover, CGACO showed superior robustness, which was
not considered sufficiently in previous studies. Many studies,

1805

Li et al.: Intima-media thickness measurement in ultrasound images

1805

F. 11. The denoising and edge enhancing effect of CBOS filter and scale multiplication. [(a) and (c)] Images polluted by simulated speckle noise. [(b) and (d)]
The corresponding edge maps obtained from the polluted images.

such as the LDLD technique published by Zhou et al.16 in

2013, did not mention whether incorrect segmentation existed
in the database or not. In 2009, Destrempes et al.26 calculated
and compared the success rate of the two proposed models
in the study, which was 83.2% and 91.5%, respectively. For
obvious detection errors, Molinari et al.29 chose to remove
them from the database, whereas Liang et al.10 introduced
human intervention mechanism to correct the errors. A fully
automatic technique named CAMES proposed by Molinari
et al.14 in 2012 achieved the success rate of 100% for the first
time. But due to the difference between databases, the success
rate can hardly be compared between different studies.
In previous researches, there has always been a lack of ways
to more fully evaluate the robustness. For the first time, we not
only used the real data, but also created simulated speckled
images to evaluate their robustness to different levels of noise.

We calculated and compared the success rates of different

approaches under different levels of noise. The result showed
that CGACO had a greater antinoise ability and maintained
high success rate even in severe noise condition. Therefore,
there are reasons to believe that CGACO could provide more
stable and reliable performance in clinical applications.
5.D. About completely automatic and manual
operations

Many completely automatic approaches have been proposed in recent years. Representatives of them include a series
of approaches proposed by Molinari.14,28,29 These approaches
do not require any user interactions and therefore have higher
efficiency. However, for low-quality images with severe noise
or misleading artifact, necessary manual operations could

F. 12. The performance of the CBOS filter under different settings of a and b. (a) The original image. (b) The images processed by the CBOS filter.
Medical Physics, Vol. 43, No. 4, April 2016

1806

Li et al.: Intima-media thickness measurement in ultrasound images

T III. Reported tracing errors in several recent studies.a

Molinari et al. (Ref. 14)

Molinari et al. (Ref. 29)
Delsanto et al. (Ref. 19)
Destrempes et al. (Ref. 26)
Ilea et al. (Ref. 43)
Xu et al. (Ref. 21)
CGACO

LI errors

MA errors

IMT errors

81 99
91 93
59.4 65
210 130
60 30
29 15
30 27

82 197
25 55
48.1 50
160 70
80 40
47.1 23
39 36

78 112
54 35
63 49.1

38.1 16.4
41 30

All data come from the published papers. The database adopted by different
methods is not the same.

provide effective guidance for the algorithm and improve

its robustness and reliability. Furthermore, user-dependent
methods generally offer better accuracy performance than
completely automatic methods.14
Basically, our strategy requires the operator to select two
points, Ps and Pe , on the carotid vessel wall manually. This is
the only manual operation during the entire process. The two
points not only determine the ROI for IMT measurement but
also play important roles in the follow procedures: First, the
orientation of carotid artery, which the CBOS filter depends on,
is determined by calculating the slope k from the coordinates
of the two points. Second, in the ACO based boundary tracking
procedure, the starting point of ants is determined by Ps . Third,
the distance between Pe and the real ending point of the ant is
set as a penalty item in the calculation of total cost. It also need
to be mentioned that although the manually selected points is
essential for the algorithm, the tracking result is not quite
sensitive to the accurate positions of the points, and thus could
avoid the interobserver bias mentioned by Molinari et al.7,28,29
5.E. Limitations

This study still has limitations. First, our proposed method

is intended for segmenting nearly parallel boundaries, which
may limit its use for plaque. However, because plaque tends
to grow into irregular shapes that are difficult to model analytically, few papers have focused on this topic.16 Moreover,
CGACO has many parameters to tune, although they are not
image-dependent and do not need adjustment for specific
images.
Another possible limitation is that the ground truth in this
study is derived from the manual tracings of only one operator.
However, according to our evaluation, there is not significant
difference between the averages of three manual tracings performed by different operators.

6. CONCLUSION
We have described and evaluated a semiautomatic technique (CGACO) for layers extraction of the carotid artery wall
in ultrasound images. Ant colony optimization is proved to
be feasible for the segmentation of intima-media complex.
Based on curvelet transform, a specially designed orientation
selective filter (CBOS filter) was used to improve its robustness
Medical Physics, Vol. 43, No. 4, April 2016

1806

performance. The experiment results showed that CGACO

achieved comparable or better accuracy in comparison with
other techniques. More than that, CGACO showed superior
robustness when testing on simulated speckled images. In
particular, it maintained high success rate in dealing with
images with excessive noise.
ACKNOWLEDGMENTS
The authors would like to thank Y. Zhou of Chinese
Academy of Sciences for providing their codes for comparison
and giving useful suggestions. The authors acknowledge the
support from the National Natural Science Foundation of
China under Grant No. 81571666.
a)Author

to whom correspondence should be addressed. Electronic mail:

zhangjue@pku.edu.cn
1G. Zahnd, S. Salles, H. Liebgott, D. Vray, A. Srusclat, and P. Moulin, Realtime ultrasound-tagging to track the 2D motion of the common carotid artery
wall in vivo, Med. Phys. 42, 820830 (2015).
2M. W. Lorenz, H. S. Markus, M. L. Bots, M. Rosvall, and M. Sitzer, Prediction of clinical cardiovascular events with carotid intima-media thickness: A systematic review and meta-analysis, Circulation 115, 459467
(2007).
3J. J. Badimon, B. Ibanez, and G. Cimmino, Genesis and dynamics of
atherosclerotic lesions: Implications for early detection, Cerebrovasc. Dis.
27, 3847 (2009).
4P. J. Touboul, M. G. Hennerici, S. Meairs, H. Adams, P. Amarenco, M.
Desvarieux, S. Ebrahim, M. Fatar, R. Hernandez Hernandez, S. Kownator, P.
Prati, T. Rundek, A. Taylor, N. Bornstein, L. Csiba, E. Vicaut, K. S. Woo, and
F. Zannad, Mannheim intima-media thickness consensus, Cerebrovasc.
Dis. 18, 346349 (2004).
5I. M. Van Der Meer, M. L. Bots, A. Hofman, A. I. Del Sol, D. A. M. Van
Der Kuip, and J. C. M. Witteman, Predictive value of noninvasive measures
of atherosclerosis for incident myocardial infarction: The Rotterdam study,
Circulation 109, 10891094 (2004).
6F. Molinari, G. Zeng, and J. S. Suri, A state of the art review on intima-media
thickness (IMT) measurement and wall segmentation techniques for carotid
ultrasound, Comput. Methods Programs Biomed. 100, 201221 (2010).
7F. Molinari, K. M. Meiburger, G. Zeng, U. R. Acharya, W. Liboni, A. Nicolaides, and J. S. Suri, Carotid artery recognition system: A comparison of
three automated paradigms for ultrasound images, Med. Phys. 39, 378391
(2012).
8N. D. Nanayakkara, B. Chiu, A. Samani, J. D. Spence, J. Samarabandu,
G. Parraga, and A. Fenster, Nonrigid registration of three-dimensional
ultrasound and magnetic resonance images of the carotid arteries, Med.
Phys. 36, 373385 (2009).
9F. Mao, J. Gill, D. Downey, and A. Fenster, Segmentation of carotid artery
in ultrasound images: Method development and evaluation technique, Med.
Phys. 27, 19611970 (2000).
10Q. Liang, I. Wendelhag, J. Wikstrand, and T. Gustavsson, A multiscale
dynamic programming procedure for boundary detection in ultrasonic artery
images, IEEE Trans. Med. Imaging 19, 127142 (2000).
11C. P. Loizou, A review of ultrasound common carotid artery image and
video segmentation techniques, Med. Biol. Eng. Comput. 52, 10731093
(2014).
12F. Faita, V. Gemignani, E. Bianchini, C. Giannarelli, L. Ghiadoni, and
M. Demi, Real-time measurement system for evaluation of the carotid
intima-media thickness with a robust edge operator, J. Ultrasound Med.
27, 13531361 (2008).
13C. Liguori, A. Paolillo, and A. Pietrosanto, An automatic measurement
system for the evaluation of carotid intima-media thickness, IEEE Trans.
Instrum. Meas. 50, 16841691 (2001).
14F. Molinari, C. S. Pattichis, G. Zeng, L. Saba, A. U. Acharya, R. Sanfilippo, A. Nicolaides, and J. S. Suri, Completely automated multiresolution
edge snapperA new technique for an accurate carotid ultrasound IMT
measurement: Clinical validation and benchmarking on a multi-institutional
database, IEEE Trans. Image Process. 21, 12111222 (2012).

1807

Li et al.: Intima-media thickness measurement in ultrasound images

15D. Cheng and X. Jiang, Detections of arterial wall in sonographic artery im-

ages using dual dynamic programming, IEEE Trans. Inf. Technol. Biomed.
12, 792799 (2008).
16Y. Zhou, X. Cheng, X. Xu, and E. Song, Dynamic programming in parallel
boundary detection with application to ultrasound intima-media segmentation, Med. Image Anal. 17, 892906 (2013).
17D. Cheng, A. Schmidt-Trucksss, K. Cheng, and H. Burkhardt, Using
snakes to detect the intimal and adventitial layers of the common carotid
artery wall in sonographic images, Comput. Methods Programs Biomed.
67, 2737 (2002).
18M. Gutierrez, P. Pilon, S. Lage, L. Kopel, R. Carvalho, and S. Furuie, Automatic measurement of carotid diameter and wall thickness in ultrasound
images, Comput. Cardiol. 29, 359362 (2002).
19S. Delsanto, F. Molinari, P. Giustetto, W. Liboni, S. Badalamenti, and J.
S. Suri, Characterization of a completely user-independent algorithm for
carotid artery segmentation in 2-D ultrasound images, IEEE Trans. Instrum.
Meas. 56, 12651274 (2007).
20C. P. Loizou, C. S. Pattichis, M. Pantziaris, T. Tyllis, and A. Nicolaides,
Snakes based segmentation of the common carotid artery intima media,
Med. Biol. Eng. Comput. 45, 3549 (2007).
21X. Xu, Y. Zhou, X. Cheng, E. Song, and G. Li, Ultrasound intima-media
segmentation using Hough transform and dual snake model, Comput. Med.
Imaging Graphics 36, 248258 (2012).
22S. Petroudi, C. Loizou, M. Pantziaris, and C. Pattichis, Segmentation of the
common carotid intima-media complex in ultrasound images using active
contours, IEEE Trans. Biomed. Eng. 59, 30603069 (2012).
23C. P. Loizou, C. S. Pattichis, A. N. Nicolaides, and M. Pantziaris, Manual
and automated media and intima thickness measurements of the common carotid artery, IEEE Trans. Ultrason., Ferroelectr., Freq. Control 56, 983994
(2009).
24F. Molinari, K. M. Meiburger, L. Saba, G. Zeng, U. R. Acharya, M. Ledda, A.
Nicolaides, and J. S. Suri, Fully automated dual-snake formulation for carotid intima-media thickness measurement. A new approach, J. Ultrasound
Med. 31, 11231136 (2012).
25S. Golemati, J. Stoitsis, E. G. Sifakis, T. Balkizas, and K. S. Nikita, Using the Hough transform to segment ultrasound images of longitudinal
and transverse sections of the carotid artery, Ultrasound Med. Biol. 33,
19181932 (2007).
26F. Destrempes, J. Meunier, M. F. Giroux, G. Soulez, and G. Cloutier,
Segmentation in ultrasonic B-mode images of healthy carotid arteries using
mixtures of Nakagami distributions and stochastic optimization, IEEE
Trans. Med. Imaging 28, 215229 (2009).
27R.-M. Menchn-Lara, M.-C. Bastida-Jumilla, J. Morales-Snchez, and
J.-L. Sancho-Gmez, Automatic detection of the intima-media thickness
in ultrasound images of the common carotid artery using neural networks,
Med. Biol. Eng. Comput. 52, 169181 (2014).

Medical Physics, Vol. 43, No. 4, April 2016

28F.

1807

Molinari, G. Zeng, and J. Suri, An integrated approach to computerbased automated tracing and its validation for 200 common carotid arterial
wall ultrasound images a new technique, J. Ultrasound Med. 29, 399418
(2010).
29F. Molinari, G. Zeng, and J. S. Suri, Intima-media thickness: Setting
a standard for a completely automated method of ultrasound measurement, IEEE Trans. Ultrason., Ferroelectr., Freq. Control 57, 11121124
(2010).
30E. Cands, L. Demanet, D. Donoho, and L. Ying, Fast discrete curvelet
transforms, Multiscale Model. Simul. 5, 861899 (2006).
31J.-L. Starck, E. J. Cands, and D. L. Donoho, The curvelet transform for
image denoising, IEEE Trans. Image Process. 11, 670684 (2002).
32M. Dorigo, V. Maniezzo, and A. Colorni, Positive feedback as a search
strategy, Technical Report No. 91-016 (1991).
33M. Dorigo, G. Di Caro, and L. M. Gambardella, Ant algorithms for discrete
optimization, Artif. Life 5, 137172 (1999).
34M. Dorigo, V. Maniezzo, and A. Colorni, Ant system: Optimization by a
colony of cooperating agents, IEEE Trans. Syst., Man, Cybern., Part B 26,
2941 (1996).
35T. Loupas, W. N. McDicken, and P. L. Allan, Adaptive weighted median
filter for speckle suppression in medical ultrasonic images, IEEE Trans.
Circuits Syst. 36, 129135 (1989).
36F. Sattar, L. Floreby, G. Salomonsson, and B. Lovstrom, Image enhancement based on a nonlinear multiscale method, IEEE Trans. Image Process.
6, 888895 (1997).
37A. Achim, A. Bezerianos, and P. Tsakalides, Novel Bayesian multiscale
method for speckle removal in medical ultrasound images, IEEE Trans.
Med. Imaging 20, 772783 (2001).
38M. Dorigo and T. Sttzle, Ant Colony Optimization (The MIT Press, Cambridge, MA, 2004).
39J.-L. Starck, F. Murtagh, E. J. Cands, and D. L. Donoho, Gray and color
image contrast enhancement by the curvelet transform, IEEE Trans. Image
Process. 12, 706717 (2003).
40J. Ma and G. Plonka, Combined curvelet shrinkage and nonlinear anisotropic diffusion, IEEE Trans. Image Process. 16, 21982206 (2007).
41K. V. Devarapu, S. Murala, and V. Kumar, Denoising of ultrasound
images using curvelet transform, in 2nd International Conference on
Computer and Automation Engineering (ICCAE) (IEEE, Singapore, 2010),
Vol. 3, pp. 447451.
42G. G. Bhutada, R. S. Anand, and S. C. Saxena, Edge preserved image
enhancement using adaptive fusion of images denoised by wavelet and
curvelet transform, Digital Signal Process. 21, 118130 (2011).
43D. E. Ilea, C. Duffy, L. Kavanagh, A. Stanton, and P. F. Whelan, Fully
automated segmentation and tracking of the intima media thickness in
ultrasound video sequences of the common carotid artery, IEEE Trans.
Ultrason., Ferroelectr., Freq. Control 60, 158177 (2013).