Medical Image Analysis 40 (2017) 60–79 

Contents lists available at ScienceDirect 

Medical Image Analysis 

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

Automatic segmentation of the lumen region in intravascular images 

of the coronary artery 

Danilo Samuel Jodas a , c , Aledir Silveira Pereira 

b , João Manuel R.S. Tavares c , ∗

a CAPES Foundation, Ministry of Education of Brazil, Brasília - DF, 70040-020, Brazil 
b Universidade Estadual Paulista “Júlio de Mesquita Filho”, Rua Cristóvão Colombo, 2265, 15054-0 0 0, S. J. do Rio Preto, Brazil 
c Instituto de Ciência e Inovação em Engenharia Mecânica e Engenharia Industrial, Faculdade de Engenharia, Universidade do Porto, Rua Dr. Roberto Frias, 

s/n, 4200-465, Porto, Portugal 

a r t i c l e i n f o 

Article history: 

Received 29 December 2016 

Revised 3 June 2017 

Accepted 9 June 2017 

Available online 10 June 2017 

Keywords: 

Medical imaging 

Intravascular ultrasound 

Image pre-processing 

Image segmentation 

a b s t r a c t 

Image assessment of the arterial system plays an important role in the diagnosis of cardiovascular dis- 

eases. The segmentation of the lumen and media-adventitia in intravascular (IVUS) images of the coro- 

nary artery is the first step towards the evaluation of the morphology of the vessel under analysis and the 

identification of possible atherosclerotic lesions. In this study, a fully automatic method for the segmenta- 

tion of the lumen in IVUS images of the coronary artery is presented. The proposed method relies on the 

K-means algorithm and the mean roundness to identify the region corresponding to the potential lumen. 

An approach to identify and eliminate side branches on bifurcations is also proposed to delimit the area 

with the potential lumen regions. Additionally, an active contour model is applied to refine the contour 

of the lumen region. In order to evaluate the segmentation accuracy, the results of the proposed method 

were compared against manual delineations made by two experts in 326 IVUS images of the coronary 

artery. The average values of the Jaccard measure, Hausdorff distance, percentage of area difference and 

Dice coefficient were 0.88 ± 0.06, 0.29 ± 0.17 mm, 0.09 ± 0.07 and 0.94 ± 0.04, respectively, in 324 

IVUS images successfully segmented. Additionally, a comparison with the studies found in the literature 

showed that the proposed method is slight better than the majority of the related methods that have 

been proposed. Hence, the new automatic segmentation method is shown to be effective in detecting the 

lumen in IVUS images without using complex solutions and user interaction. 

© 2017 Elsevier B.V. All rights reserved. 

 

 

 

 

 

 

 

 

 

 

 

b  

e  

t  

t  

d  

g  

i  

f  

r

 

e  

t  
1. Introduction 

According to the World Health Organization (WHO), coronary

artery diseases were responsible for the death of 7.3 million peo-

ple around the world in 2008 ( Mendis et al., 2011 ). Atheroscle-

rosis is an underlying disease responsible for the occurrence of

heart attacks and strokes. Atherosclerotic plaques are formed when

fatty material and cholesterol are deposited inside the lumen of

the artery, reducing the blood flow through the vessel and increas-

ing the risk of blood clots that can cause heart attacks and strokes.

Thus, in order to prevent such risks a treatment plan or even stent-

ing procedures should be established based on image-based tech-

nologies. 
∗ Corresponding author. 

E-mail addresses: danilojodas@gmail.com (D.S. Jodas), aledir@sjrp.unesp.br (A.S. 

Pereira), tavares@fe.up.pt (J.M.R.S. Tavares). 

URL: http://www.fe.up.pt/˜tavares (J.M.R.S. Tavares) 

t  

a  

(  

B  

a  

t  

http://dx.doi.org/10.1016/j.media.2017.06.006 

1361-8415/© 2017 Elsevier B.V. All rights reserved. 
Technological advances in computerized systems for imaging-

ased diagnosis are able to detect and analyse cardiovascular dis-

ases. Intravascular Ultrasound (IVUS) is an imaging procedure

hat allows the evaluation of the arterial morphology by means of

he introduction of a catheter equipped with an ultrasound trans-

ucer inside the vessel to be studied. The catheter slides through a

uidewire placed in the blood vessel near the segment of interest

n order to acquired images of the affected region. Images acquired

rom IVUS imaging systems allow experts to envisage atheroscle-

otic lesions and the shape and size of the vessels under analysis. 

The segmentation of IVUS images plays an important role in

valuating the morphology of the vessel under study and ob-

aining important information such as the area and diameter of

he lumen, the presence and volume of atherosclerotic plaques

nd the identification of the atherosclerotic plaque components

 Nair et al., 2002; Diethrich et al., 2006; König and Klauss, 2007;

roersen et al., 2016 ). Since a large number of image frames

re acquired from a single IVUS exam, the manual identifica-

ion of the structures of interest becomes a time-consuming and

http://dx.doi.org/10.1016/j.media.2017.06.006
http://www.ScienceDirect.com
http://www.elsevier.com/locate/media
http://crossmark.crossref.org/dialog/?doi=10.1016/j.media.2017.06.006&domain=pdf
mailto:danilojodas@gmail.com
mailto:aledir@sjrp.unesp.br
mailto:tavares@fe.up.pt
http://www.fe.up.pt/~tavares
http://dx.doi.org/10.1016/j.media.2017.06.006


D.S. Jodas et al. / Medical Image Analysis 40 (2017) 60–79 61 

l  

a  

o  

s

 

i  

e  

r  

t  

s  

t  

t  

t  

s  

s  

o  

t  

t

 

m  

s  

i  

2  

i  

T  

n  

t  

r  

m  

i  

i  

b  

a  

b  

t  

t  

r  

n  

t  

A  

o  

t  

s  

b

 

S  

t  

t  

o  

d  

i  

m  

d

2

 

t  

u  

h  

b  

t  

m

 

p  

i  

t  

t  

t  

p  

c  

b  

c

 

m  

t  

w  

t  

A  

o  

f  

T  

m  

t  

t  

o

 

f  

a  

p  

t  

l  

n

 

l  

b  

t  

l  

b  

a  

t  

d  

h

 

m  

t  

s  

(  

p  

i  

i  

g  

t

 

s  

o  

p  

t  

b  

s  

e  

t  

m  

m

 

a  

a  

c  

o  

s  

p  

o  
aborious task. Therefore, the automatic segmentation of IVUS im-

ges is demanded for expediting the assessment of the morphol-

gy of the vessel and the treatment planning of atherosclerotic le-

ions. 

The segmentation of the lumen and media-adventitia regions

n IVUS images has been an intensive focus of research ( Katouzian

t al., 2012 ). The presence of calcifications, shadows, transducer

eflection, speckle noises and bifurcations, as well as the varia-

ions of the grayscale intensities inside the same structure, repre-

ent a challenge to the development of a fully automatic segmen-

ation method. The presence of the first two artefacts, i.e. calcifica-

ions and shadows, is an obstacle to the efficient segmentation of

he media-adventitia region, whereas the transducer reflection and

peckle noises represent a challenge to identify the region corre-

ponding to the lumen successfully. Additionally, the segmentation

f the lumen in bifurcation regions represents a challenge due to

he extension of the low-intensity values from the lumen region to

he borders of the IVUS image. 

This article proposes a method to automatically detect the lu-

en boundary in IVUS images of coronary arteries. An initial ver-

ion of this method was applied in the segmentation of the lumen

n magnetic resonance (MR) images of carotid arteries ( Jodas et al.,

016 ). The approach was based on the assumption that the lumen

s a low-intensity region with an approximately circular shape.

hus, a circularity index, combined with a centre index and the

umber of regions at the border of the image were used to find

he region corresponding to the lumen under analysis. In the cur-

ent study, besides the assessment concerning the suitability of the

ethod to detect the lumen boundary of coronary arteries in IVUS

mages using the same parameters as were established for the MR

mages, the method has been improved in order to enhance its ro-

ustness, efficiency, automaticity and competency. Hence, a new

pproach was developed to identify and remove side branches of

ifurcation regions so only the regions that potentially correspond

o the lumen are selected for further processing, avoiding bifurca-

ion parts at the border of the image and enhancing therefore, the

obustness of the method. Moreover, the regions at the border are

ow discarded before computing the circularity index of each po-

ential lumen region, which boosts the performance of the method.

dditionally, noisy regions are now automatically eliminated with-

ut the use of any predefined parameter, which increases the au-

omaticity. Finally, a post-processing step has now been added to

mooth the contour found for the lumen region, which leads to

etter segmentation results. 

The remainder of the article is organized as follow:

ection 2 presents the previous studies related to the segmenta-

ion of IVUS images. Section 3 provides a detailed description of

he proposed automatic lumen segmentation method. The results

f the segmentation, as well as the comparison with the manual

elineations and related methods found in the literature, are given

n Section 4 . The advantages and limitations of the proposed

ethod are pointed out in Section 5 . Finally, the conclusions are

rawn up in Section 6 . 

. Previous studies 

Several studies have been proposed to address the segmenta-

ion of the lumen and media-adventitia in IVUS images acquired

sing different ranges of frequencies and with several artefacts that

amper the segmentation successful. In summary, the studies can

e divided into approaches based on machine learning, probabilis-

ic functions, deformable models, region growing, thresholding and

orphological operations. 

Lo Vercio et al. (2016) proposed a learning machine-based ap-

roach to segment the lumen and media-adventitia regions in IVUS

mages of coronary arteries. From the input images, a set of fea-
ures obtained using noise-reduction filters, texture and edge de-

ector operators was acquired and used as input to a Support Vec-

or Machine (SVM) classifier, which classify the likelihood of the

ixels to belong to the lumen and background regions. Hence, the

lassification between lumen and non-lumen, as well as between

ackground and non-background, was performed to determine the

orrect location of the lumen and media-adventitia contours. 

Another learning machine-based approach to segment the lu-

en and media-adventitia contours in IVUS images of coronary ar-

eries was tackled by Su et al. (2016) . Two Artificial Neural Net-

orks (ANN) are used to classify the pixels inside a region of in-

erest (ROI) as belonging to the media-adventitia region. The first

NN is used to perform the initial classification, whereas the sec-

nd one performs the classification of the binary image resultant

rom the first ANN in order to remove noises and refine the results.

hen, the pixels inside the region corresponding to the identified

edia-adventitia are classified as lumen and non-lumen by using

he same ANNs. The contours of the regions are then submitted to

he Snake active contour model proposed by Kass et al. (1988) in

rder to adjust them to the true boundary of each region. 

A study carried out by Destrempes et al. (2014) proposed a

ast-marching method (FMM) to segment the lumen and media-

dventitia boundaries in IVUS images of coronary arteries. The pro-

osed method relies on the minimization of a function that uses

he gradient of the image and the probability of the pixels to be-

ong to a region corresponding to one of the structures of the coro-

ary artery: lumen, media, adventitia, and surrounding tissues. 

Mendizabal-Ruiz et al. (2013) proposed an approach to find the

umen contour in IVUS images of coronary arteries based on proba-

ilistic functions. After the transformation of the IVUS images into

he polar coordinates domain, the probabilities of the pixels be-

onging to the lumen region are used to find the correct contour

y means of the minimization of a cost function. The probabilities

re calculated by using a sigmoid function and an SVM classifier

hat receives texture features from Laws filters, which are used to

eal with pixel intensities that increase due to the catheters with

igher frequencies. 

A parametric active contour model for the segmentation of lu-

en and media-adventitia contours in IVUS images of coronary ar-

eries was proposed by Vard et al. (2012) . For the lumen contour

egmentation, a method based on the short-term autocorrelation

STA) is used to remove speckle noise from the lumen region. The

roposed STA method, called normalized cumulative STA (NCSTA),

s then used to produce a new grayscale IVUS image without noise

n the lumen region. The image resulting from the NCSTA method

enerates a pressure force for a parametric active contour model

hat is used to find the lumen contour. 

A 3D-FMM method was proposed by Cardinal et al. (2006) to

egment the lumen and media-adventitia contours in IVUS images

f femoral arteries. The 3D-FMM method is based on the Rayleigh

robability density function (PDF) and a gradient function to find

he correct contour of each region, i.e. lumen and media-adventitia,

y means of the refinement of contours manually defined. Another

tudy tackled by Cardinal et al. (2010) combines the PDF and gradi-

nt intensity obtained from the input images in order to calculate

he propagation speed of the 3D-FMM. In addition, the proposed

ethod automatically detects the initial contours of the lumen and

edia-adventitia borders to be used by the 3D-FMM method. 

Taki et al. (2008) proposed a method based on thresholding

nd deformable models to identify the contours of the lumen

nd media-adventitia regions in IVUS images. Regarding the lumen

ontour identification, the evaluation of the pixels in the polar co-

rdinates domain is made in order to identify the ones with inten-

ities higher than a pre-defined threshold T . If I ( r, θ ) > T , then the

ixel is assumed to belong to the lumen border. The initial contour

f the lumen is then submitted to a parametric active contour and



62 D.S. Jodas et al. / Medical Image Analysis 40 (2017) 60–79 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

3

 

fi  

i  

n  

t

 

t  

b  

a  

m  

c

T  

w

i  

a

c  

a

 

w  

m  

a  

T  

P  

(

α  

w  

a  

o  

p  

t  

w

3

 

s  

s  

i  

e  

(  

n  

b  

v  

c  

t  

a

 

a  

c  

p  

t  

n  

p  

G  

r  

u

 

u  
a geometric model to adjust it to the true borders of the corre-

sponding lumen region. 

The combination of fuzzy clustering with morphological opera-

tion was proposed by dos Santos et al. (2005) to identify the lumen

region in IVUS images. The fuzzy clustering defined with two clus-

ters was applied to the images after the transformation into the

polar coordinates domain. Then, the morphological closing opera-

tion was performed to obtain more regular borders. The extracted

contour was converted into the Cartesian coordinate domain and

overlapped with the IVUS images to present the final segmentation

result. Segmentation methods based on the Otsu’s threshold algo-

rithm and morphological operations for the identification of the

lumen region in IVUS image were also proposed by Sofian et al.

(2015) and Moraes and Furuie (2011) . 

Most of the above mentioned studies still required manual in-

terventions, re-training of the learning-based model and the trans-

formation of the image into the polar coordinates domain. In ad-

dition, the use of threshold values as proposed by Taki et al.

(2008) can fail to identify the initial contour of the lumen in differ-

ent datasets due to the usual variability of the grayscale intensities.

Hence, the development of a more robust, efficient, automated and

less complex solution was the goal of this study. 

3. Materials and methods 

3.1. IVUS images used 

This study was accomplished by using the IVUS images of coro-

nary arteries selected for the IVUS Segmentation Challenge de-

scribed in Balocco et al. (2014) that were kindly provided by the

authors. The images were acquired using a Si5 imaging system

(Volcano Corporation) equipped with an Eagle Eye catheter oper-

ating at a frequency of 20 MHz ( Balocco et al., 2014 ). A total of

326 images with ground truths provided were selected for validat-

ing the proposed method. The manual delineations of the contours

corresponding to the lumen and media-adventitia regions were

performed by two experts, and one of them repeated the manual

delineations about one week after the first delineations ( Balocco

et al., 2014 ). The images have a resolution of 384 × 384 pixels

and the pixel size is 0.026 mm × 0.026 mm. More details about

the IVUS images are available in Balocco et al. (2014) . 

3.2. Proposed method 

The proposed lumen segmentation method is made up of three

main stages: pre-processing, segmentation and lumen identifica-

tion. The diagram of the proposed method is shown in Fig. 1 . 

The pre-processing stage is necessary to minimize noise and ad-

just the contrast of the input image. Then, the enhanced image

is submitted to the segmentation stage in order to separate the

regions with low pixel values, which include the background and

the lumen. Relatively to the initial version ( Jodas et al., 2016 ), the

method has three new steps in this stage: the Gaussian pyramid,

which is adopted to reduce the resolution of the input image; the

elimination of regions at the border of the image; and the iden-

tification and removal of side branches at the bifurcation regions.

Additionally, connected component labelling is employed here to

separate all regions of the binary image corresponding to the low

intensity values instead of using the region growing algorithm pro-

posed in the initial version. The lumen identification stage uses

two classification indexes to identify the region corresponding to

the lumen of the artery under analysis, which is then inputted to

an active contour algorithm for further refinement of the bound-

ary. A post-processing step is now included in the method in order

to smooth the contour resultant from the active contour algorithm,

which leads to better segmentation results. 
.2.1. Pre-processing 

The first step of the pre-processing stage is the use of a median

lter with a mask of 5 × 5 to minimize the noise in the original

mage. The median filter was chosen due to its ability to remove

oise without compromising the boundaries of the regions of in-

erest. 

The contrast enhancement step improves the brightness of

he dark regions of the input image. The gamma correction-

ased method proposed by Huang et al. (2013) was employed to

void the overestimation of regions with low-level intensities. The

ethod relies on the probability density function (PDF) and the

umulative density function (CDF) of the intensity values as: 

 (l) = l max × (l/l max ) 
1 −cdf w (l) , (1)

here cdf w 

( l ) is the weighting CDF of the intensity value l and l max 

s the highest possible intensity value. The weighting PDF ( pdf w 

)

nd weighting CDF ( cdf w 

) are defined as: 

df w 

(l) = 

l max ∑ 

l=0 

pdf w 

(l) ∑ 

pdf w 

, (2)

nd 

pdf w 

(l) = pdf max ×
(

pdf (l) − pdf min 

pd f max − pd f min 

)α

, (3)

here pdf min is the minimum probability of the PDF, pdf max is the

aximum probability of the PDF, 
∑ 

pdf w 

= 

∑ l max 

l=0 
pdf w 

(l) and α is

 parameter that controls the amount of contrast enhancement.

he value of the α parameter is determined by partitioning the

DF of the grayscale intensities as proposed in our previous work

 Jodas et al., 2016 ): 

= 

t ∑ 

i =1 

P DF min i −
N ∑ 

j= t+1 

P DF max j , (4)

here PDF min and PDF max represent the probabilities of the low

nd high intensities of the input image, respectively, t is the value

btained by the Otsu’s threshold algorithm and N is the highest

ossible intensity value. Here, if the value of α is less than 0 (zero),

he contrast correction is not necessary since there are more pixels

ith high intensities. 

.2.2. Segmentation 

The K-means clustering algorithm is a well-known method to

eparate regions with similar characteristics (of intensity, for in-

tance) in images. Finding the correct cluster centroids to be used

n the K-means algorithm is a challenging task because differ-

nt images have different cluster centroids. Subtractive clustering

 Bataineh et al., 2011; Chiu, 1994 ) has been proposed as an alter-

ative approach to find the adequate number of cluster centroids

ased on the potential of each pixel in the neighbourhood. The ad-

antage of the subtractive clustering algorithm is that the cluster

entroids do not change in different runs. This is due to the fact

hat the potential function relies on the pixel values only (or on

nother feature calculated from the pixels of the input image). 

Although subtractive clustering has been widely used in im-

ge segmentation and classification problems, the computational

omplexity of O ( d 2 , N 

2 ), with N representing the number of data

oints and d the dimensional number, has limited its application

o large-scale problems ( Sun et al., 2012 ). Hence, a reduction in the

umber of data points, i.e. pixels, is necessary to reduce the com-

utational time of the subtractive clustering algorithm. Here, the

aussian pyramid is applied to the input IVUS image in order to

educe its resolution and consequently, the number of pixels to be

sed in the subtractive clustering algorithm. 

The Gaussian pyramid is an image processing technique

sed to reduce the resolution of images in repeated steps



D.S. Jodas et al. / Medical Image Analysis 40 (2017) 60–79 63 

Fig. 1. Diagram of the proposed method. 

(  

a  

T  

a  

r  

t

g  

w  

a  

w  

t  

a  

b  

s  

o  

i  

n  

t  

i  

i  

o  

p  

a  

i

 

c  

r  

o  

t  

w  

l  

i  

s  

t  

i  

t  

i  

s  

i  

a  

i  

r  

m  

t  

a

 

r  

(  

t  

m  
 Adelson et al., 1984 ). Pattern recognition, texture analysis and im-

ge compression are examples for the use of the Gaussian pyramid.

he pyramid is a sequence of copies of the original image gener-

ted by recursively smoothing the image with a low-pass filter and

educing its resolution by half. The following equation represents

he basis of the Gaussian pyramid: 

 l (i, j) = 

2 ∑ 

m = −2 

2 ∑ 

n = −2 

w (m, n ) g l−1 (2 i + m, 2 j + n ) , (5)

here l is a level of the pyramid, i and j are the pixel coordinates

t the l th level, g l−1 is the smoothed image at the level l − 1 and

 ( m, n ) is a 5 × 5 low-pass filter applied to the neighbourhood of

he pixel ( i, j ). The original image is represented by g 0 , i.e. the im-

ge at the level l = 0 . The image at the level l = 1 is generated

y smoothing g 0 with w ( m, n ) followed by downsampling the re-

ulting image by a factor of two. The image at the level l = 2 is

btained by applying the same procedure on the image g 1 , i.e. the

mage at the level l = 1 . The process is repeated until the desired

umber of levels is reached. The final result is a pyramid in which

he base is the original image and the top represents the smoothed

mage with the smallest resolution. The smoothing of the image

s necessary to avoid the aliasing effect generated when the res-

lution is reduced. An example of the application of the Gaussian

yramid in order to reduce the resolution of an IVUS image and

ccelerate the use of the subtractive clustering algorithm is shown

n Fig. 2 . 
Although the most common use of the Gaussian pyramid is

ompressing an input image and multiscale processing at lower

esolutions, here the goal is to generate a pyramid of lower res-

lution images for the input image in order to select the one with

he smallest resolution, i.e. the image at the top of the pyramid,

here the subtractive clustering algorithm is then applied with

ower computational time than if it was applied on the original

mage. Here, the number of reductions performed by the Gaus-

ian pyramid is defined as equal to 2, leading to a pyramid with

hree levels in which the base is represented by the input IVUS

mage with 384 × 384 pixels of resolution and the top represents

he IVUS image with the smallest resolution of 96 × 96 pixels that

s inputted to the subtractive clustering algorithm. Thereafter, the

ubsequent steps of the proposed method are performed on the

mage with the smallest resolution in order to obtain a binary im-

ge with the identified lumen, which is then restored to the input

mage resolution without performing any processing on the higher

esolution images of the pyramid. The choice of the Gaussian pyra-

id was due to its simple implementation and ability to reduce

he input image resolution without losing important information

bout the structure of interest. 

After the downsampling of the input image, the K-means algo-

ithm with subtractive clustering suggested in Dhanachandra et al.

2015) is employed to separate the regions of the image according

o the grayscale intensity. The centroid of each cluster is found by

eans of the subtractive clustering algorithm. After the centroids



64 D.S. Jodas et al. / Medical Image Analysis 40 (2017) 60–79 

Fig. 2. Example of the Gaussian pyramid applied two times on a pre-processed IVUS image (the second row shows the Gaussian pyramid results expanded to the size of 

the original image). 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

t

E  

 

i  

b  

e  

r

M  

w  

a  

a  

i

 

i

I  

w  

d

 

t  

c  

i  

t  

f  

a  

t

 

a  
have been found, the K-means clustering is applied to separate the

regions of the image. As proposed in our previous study, four clus-

ters are used here to correctly identify the region corresponding

to the lumen. Since the lumen and background regions have low

intensities in IVUS images, the same approach employed is used

to find the cluster having the regions with low-intensity pixels for

identifying the one representing the correct lumen. A binary im-

age with such regions is returned, in which the identified regions

are represented by white pixels and the background by black pix-

els. The image is then submitted to the connected component la-

belling algorithm in order to obtain all regions of the image sep-

arately. Here, the connected component labelling is performed on

the image corresponding to the cluster with low-intensity values

to obtain a set with all regions of interest. Firstly, the algorithm

performs a pixel-by-pixel scan of the binary image from the top

to bottom and left to right in order to identify a white pixel and

assign a label to it that relies on the evaluation of the adjacent pix-

els that share the same intensity. If none of the neighbour pixels

have until then been labelled, a new label is assigned to the pixel

under analysis. Otherwise, the pixel under analysis receives a label

assigned to a neighbour pixel. The process continues until all pix-

els of the binary image have been labelled. Then, pixels having the

same label are merged to form a single region. Thereafter, each re-

gion is used as the input to the lumen identification stage in order

to find the one that represents the lumen. 

3.2.3. Lumen identification 

In our previous work, a set of measures was used to evaluate

each region of interest obtained by the region growing algorithm

in order to identify the one corresponding to the lumen in the

MR images of carotid arteries. Hence, the mean roundness ( MR ),

irregularity ( Ir ) and centre ( d ) indexes are included in the follow-

ing circularity function, which is used to evaluate each region of
he binary image: 

 = MR + 

1 

Ir 
+ 

1 

d 
. (6)

The MR was proposed to determine the circularity of objects in

mages ( Ritter and Cooper, 2009 ). It consists of calculating the ratio

etween the average radius and the distance between the radius of

ach border pixel of the object and the average radius. The mean

oundness is calculated according to: 

R = 

1 

N 

N ∑ 

i =1 

r̄ b 
| r i − r̄ b | + r̄ b 

, (7)

here N is the number of pixels of the contour of the object under

nalysis, r̄ b is the average radius of the object, and r i is the radius

t the contour pixel i . The larger the mean roundness ( MR ) index

s, the more circular the object under analysis is. 

The following irregularity index is used to avoid regions with

rregular contours: 

r = P ∗
(

1 

SD 

− 1 

GD 

)
, (8)

here P is the number of pixels of the contour, SD is the shortest

iameter and GD is the greatest diameter ( Jain et al., 2015 ). 

A centre index is also used to identify the correct location of

he lumen. As described in our previous work, the lumen is a

ircular-shaped region located close to the centre of the input MR

mage. Hence, the distances between the centre of the image and

he centre of each region are calculated and used to penalize those

ar from the centre of the MR image. The same concept can also be

pplied here due to the fact that the lumen is also located close to

he centre of an IVUS image. 

The segmentation of the lumen close to bifurcations represents

 challenge due to the extension of the low-intensity values from



D.S. Jodas et al. / Medical Image Analysis 40 (2017) 60–79 65 

Fig. 3. Example of the identification and removal of the side branch of a bifurcation region: a) Image pre-processed and reduced by the Gaussian pyramid; b) Binary image 

corresponding to the cluster with low-intensity pixels; c) Binary image with the identified bifurcation region (the circle generated by the proposed approach is shown in 

red); d) Binary image without the part of the bifurcation region on the border. (For interpretation of the references to colour in this figure legend, the reader is referred to 

the web version of this article.) 

t  

t  

c  

t  

a  

t  

t  

r  

g  

i  

g  

l  

h  

e  

r  

d  

n  

o  

p  

t  

i  

i  

r

r  

w  

s  

i  

c  

i  

n  

o  

c

 

t  

p  

t  

c  

i  

i

 

n  

m  

d  

l  

s  

I  

a  

s  

c  

s

 

g  

o  

g  

t  

r  

b  

c  

v  

f  

r  

i

 

t  

i  

t

E  

 

l  

a  

t  

H  

r  

i  

s  

e  

s  

i  

b  

m  

t  

t  

i  

c  

e  

s  

t

S

he lumen region to the border of the input image, which causes

he establishment of a region representing this extension when the

luster corresponding to the low-intensity values is built. Hence, in

he initial version of the proposed method, these regions were not

ssessed in the subsequent processing steps, since the regions at

he border of the image are discarded before the identification of

he lumen. Therefore, a new approach was developed to identify

egions representing bifurcations in order to distinguish these re-

ions from the ones at the border of the binary images correspond-

ng to low-intensity values. The proposed approach is based on the

eneration of circles surrounding the regions of possible lumens,

ike the approach proposed in de Macedo et al. (2016) . However,

ere, the distances from the centres of the input image to the pix-

ls of the bifurcation regions at the borders are used to define the

adius of the circles, instead of the distance transform suggested in

e Macedo et al. (2016) . Hence, the regions obtained by the con-

ected component labelling algorithm are evaluated to identify the

ne that is in the centre of the input image. Then, the number of

ixels of this region that are at the border are computed in order

o verify if the region represents a bifurcation. Once the region is

dentified as a bifurcation, the distances between the centre of the

mage and each pixel of the region at the border is calculated. The

adius of the circle is then calculated according to: 

 = min (dist(Cp, Bp)) − min (dist(Cp, Rp)) , (9)

here Cp is the pixel at the centre of the IVUS image, Bp is the

et of pixels of the bifurcation region that are at the border of the

mage and Rp is the set of all pixels of the contour of the bifur-

ation. The circle generated is centred at the centre of the IVUS

mage and the region outside this circle is removed from the bi-

ary image corresponding to the low-intensity values. An example

f the proposed approach applied to an IVUS image with a bifur-

ation region is illustrated in Fig. 3 . 

As depicted in Fig. 3 c, the bifurcation region can be divided into

wo parts, i.e. one inside the circle and the other outside. Since the

art inside the circle contains the regions with the possible lumen,

he regions outside the circle are removed from the binary image

orresponding to the low-intensity values, leading to a new binary

mage that is then submitted to the subsequent steps of the lumen

dentification stage. 

Concavities, irregularities and holes inside the regions of the bi-

ary image corresponding to the cluster with low-intensity values

ay be present due to high intensity values inside the lumen un-

er analysis. Since the high intensity values are identified as be-

onging to another cluster when the K-means algorithm is applied,

uch artefacts are generated in the regions of the binary image.

n order to attenuate the effect of these artefacts, the convex hull
lgorithm is used here. The use of the convex hull algorithm in-

tead of morphological operations is to avoid underestimation and

utting of parts of the region of interest when concavities are pre-

ented. 

The convex hull of a set of points S is the smallest convex poly-

on containing all the points of S ( Berg et al., 2008 ). The definition

f convex hull relies on the concept of a convex set, which is a re-

ion defined in a way that for every pair of points [ a, b ] belonging

o a region, the line joining such points must be totally inside the

egion. Here, for each region present in the binary image obtained

y the connected component labelling algorithm a correspondent

onvex hull is generated i.e., the set of pixels representing the con-

ex polygon that includes all the white pixels of the region, be-

ore calculating the classification indexes used to find the lumen

egion. An example of the application of the convex hull algorithm

s shown in Fig. 4 . 

As depicted in Fig. 4 , the regions become more regular after

he application of the convex hull algorithm. Hence, the irregular-

ty index is now redundant and the function defined by Eq. (6) is

herefore, simplified to: 

 = MR + 

1 

d 
. (10)

In the initial version of the proposed method, the regions with

ess than 1.5% of the total number of pixels of the input im-

ge were discarded from the lumen identification procedure, since

hese regions usually represent small regions associated to noise.

owever, this empirically defined discarding criterion is not always

obust. Hence, a new approach based on the morphological open-

ng operation was developed in order to identify and disregard the

mall regions associated to noise. The morphological opening op-

ration is the erosion followed by the dilation of an image I by a

tructuring element SE . When applied to binary images, the open-

ng operation can smooth the borders of the regions represented

y white pixels, split regions connected by thin bridges and re-

ove small regions that represent noisy artefacts. The choice of

he shape and size of the structuring element is usually based on

he shape of the regions of interest, and it plays an important role

n effectively achieving the desired results. Since the lumen is a

ircular-shaped region in IVUS images, a disk-shaped structuring

lement is used to perform the opening operation. In addition, the

ize of the structuring element is adaptively defined according to

he approach proposed by Gao et al. (2015) and formulated as: 

 o = 

1 

2 

√ 

A r 

π
, (11) 



66 D.S. Jodas et al. / Medical Image Analysis 40 (2017) 60–79 

Fig. 4. Example of the application of the convex hull algorithm: a) Image pre-processed and reduced by the Gaussian pyramid; b) Result of the K-means with the subtrac- 

tive clustering algorithm; c) Binary image corresponding to the cluster with low-intensity pixels (regions at the border corresponding to the background were previously 

removed); d) Result of the convex hull algorithm applied to each region (white pixels) of the binary image. 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

r

S  

w  

t

i  

r  

o  

t  

o  

w  

t

 

m

3

 

m  

c  

a

 

t  

r  

s

J  

 

r  

r  

u  

s

D  

 

t  

g  

m

H  

w  

a

 

e  
where A r is the area of the region under analysis. The opening op-

eration with the adaptive structuring element is applied to each

region obtained by the connected component labelling algorithm.

Hence, a region is not considered in the lumen identification step

if its size is equal or less than the size of the structuring ele-

ment, and the region is removed by the opening operation. Here,

the morphological opening operation is only employed to iden-

tify small regions corresponding to noise. Therefore, the smooth-

ing of the regions is not performed by the opening operation since

the convex hull is previously applied to correct the shape of the

regions under analysis. A region identified as noise and removed

from the binary image by the opening operation is not selected for

calculating the circularity index defined in Eq. (10) . 

Although the terms in Eq. (10) are defined in different contexts,

the objective of using the centre index in the proposed circular-

ity index is to penalize regions with larger values of roundness

that are distant from the centre of the input image. The value of

the circularity index proposed in Eq. (10) is calculated for each re-

gion resultant after the opening operation and the one with the

maximum value is considered to be the lumen of the artery under

study. This region is resized to the original resolution of the in-

put IVUS image and then submitted to the Chan–Vese active con-

tour algorithm ( Chan and Vese, 2001 ) for further refinement of the

contour. The binary image representing the lumen region may not

fit the true boundary of the lumen in the IVUS image. Hence, the

contour of such a region is used as the input of the Chan–Vese ac-

tive contour model, which is applied to the original IVUS image in

order to fit the contour to the true boundary. This refinement step

plays an important role in avoiding under- or over-estimating the

contour, leading the contour closer to the true boundary of the lu-

men under analysis and, consequently, to better results. Since the

gradient of the image is not used in the Chan–Vese model, the

method is recommended for the segmentation of medical images

which commonly have weak boundaries of the structures of inter-

est ( Ma et al., 2010; Wang et al., 2010; Santos et al., 2013; Huang

and Zeng, 2015 ). 

The post-processing step consists in smoothing the lumen con-

tour by using the morphological opening and dilation operations.

The morphological opening operation was employed to smooth the

boundaries of the lumen region generated from the result of the

Chan–Vese active contour. A binary image is generated from the

lumen contour such that the region inside the contour is repre-

sented by white pixels. Then, the region is smoothed by first apply-

ing the opening operation. Finally, the morphological dilation op-

eration is performed on the resulting region in order to restore its

size as close as possible to the size of the original region. The sizes

of the structuring elements are defined according to the size of the

m  

e  
egion under analysis and formulated as: 

 o = 

1 

2 

√ 

A l 

π
, S d = 

1 

N 

N ∑ 

i =1 

√ 

(C o i − C r k ) 
2 
, (12)

here S o and S d are the sizes of the structuring element used in

he morphological opening and dilation operations, respectively, A l 

s the area of the lumen region, C o is the contour of the region

esulting from the opening operation, C r is contour of the region

btained from the lumen contour and N is the number of pixels of

he contour C o . Co i is the i th pixel of Co , and Cr k is the k th pixel

f Cr closest to Co i . The size S o was proposed by Gao et al. (2015) ,

hereas S d is defined as the average distance between the con-

ours C o and C r . 

The steps performed in the automatic segmentation of the lu-

en in one IVUS image are depicted in Fig. 5 . 

.3. Quantitative analysis 

The contours obtained by the proposed method and the related

anual delineations were compared based on four measures: Jac-

ard measure ( JM ), Dice coefficient ( DC ), Hausdorff distance ( HD )

nd percentage of area difference ( PAD ). 

The Jaccard measure is calculated by means of the ratio be-

ween the size of the intersection and the size of the union of the

egions corresponding to the automatic ( S auto ) and manual ( S manual )

egmentations: 

M = 

| S auto ∩ S manual | 
| S auto ∪ S manual | . (13)

The Jaccard measure is important to assess the overlap of the

egion identified by the proposed method with respect to the cor-

esponding manual delineation. Similarly, the Dice coefficient is

sed to calculate the overlap between the automatic and manual

egmentations: 

C = 

2 ∗ | S auto ∩ S manual | 
| S auto | + | S manual | . (14)

The Hausdorff distance is important to assess the closeness of

wo contours, and it is defined here as the maximum between the

reatest distances between the pixels of the automatic ( C a ) and

anual ( C m 

) contours: 

D (C a , C m 

) = max { max 
a ∈ Ca 

min 

b∈ Cm 

d(a, b) , max 
b∈ Cm 

min 

a ∈ Ca 
d(a, b) } , (15)

here a and b are the pixels of contours C a and C m 

, respectively,

nd d ( a, b ) is the Euclidean distance between these pixels. 

The percentage of area difference ( PAD ) represents the differ-

nce between the areas of the contour obtained from the auto-

atic segmentation ( A auto ) and the corresponding manual delin-

ation ( A ) with respect to the area of the manual delineation:
manual 



D.S. Jodas et al. / Medical Image Analysis 40 (2017) 60–79 67 

Fig. 5. Example of the output images resulting from each step of the automatic segmentation of the lumen in one IVUS image (the green contour in the image on the right 

of the lumen identification step represents the final segmentation result). (The identification of the bifurcation region step is not illustrated in this example since there was 

no side branch in the IVUS image used.). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) 

P

 

m  

t  

B  

t  

a

A  
AD = 

| A auto − A manual | 
A manual 

. (16) 

The lumen area and the average lumen diameter of the auto-

atic and manual segmentations were also calculated to compare

he segmentation results by means of the regression analysis and
land–Altman analysis. The lumen area is the area inside the con-

our of the lumen and the average lumen diameter ( ALD ) is defined

s: 

LD = 

1 

N 

N ∑ 

i =1 

2 ∗ r i , (17)



68 D.S. Jodas et al. / Medical Image Analysis 40 (2017) 60–79 

Table 1 

Details about the datasets and the validation measures of the studies used to compare the segmentation 

results of the proposed method. 

Authors Catheter’s frequency Number of frames Validation measure(s) 

Lo Vercio et al. (2016) 20 MHz 149 JM and PAD 

Su et al. (2016) 20 MHz 461 JM and HD 

Sofian et al. (2015) 20 MHz 30 JM, HD, PAD and DC 

Destrempes et al. (2014) 40 MHz 3207 HD 

Mendizabal-Ruiz et al. (2013) 20 MHz and 40 MHz 585 JM, HD and DC 

Cardinal et al. (2010) 20 MHz 1593 HD 

Vard et al. (2012) 30 MHz 40 HD 

Taki et al. (2008) 30 MHz 60 HD 

∗JM = Jaccard measure; HD = Hausdorff distance; PAD = percentage of area difference; DC = Dice coeffi- 

cient; MHz = Megahertz 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

4

 

u

 

p  

r  

i  

m  

t  

i  

p

 

s

 

t  

a  

c  

p  

d

 

e  

 

t  

w  

l  

(  

t  

w  

t

 

p  

v  

a  

l  

c  

d  

f  

 

m  

m  

c  

t  

e  

o  

t  

T  

0  

e  

D  

a  
where N is the number of pixels of the contour of the lumen and

r i is the radius of the i th pixel. 

4. Results 

The proposed method was implemented in MATLAB software

(The Mathworks Inc., Natick, USA) and executed in a desktop com-

puter equipped with an Intel i7-4700 HQ processor (2.4 GHz) and

16GB of RAM memory. A comparison between the contours ob-

tained from the proposed method and those generated by manual

delineations was performed in order to evaluate the accuracy of

the segmentation results. In addition, the results of the proposed

method were also compared with the ones reported in related

studies found in the literature. The details about the datasets and

the validation measures of the studies used here to compare the

segmentation results of the new automatic segmentation method

are shown in Table 1 . 

Since there are not any common validation measures in

the studies indicated in Table 1 , all measures described in

Section 3.3 were calculated and the ones used in each of these

studies were selected to compare the performance of the proposed

automatic lumen segmentation. 

The proposed automatic lumen segmentation was also com-

pared with the results obtained from the eight participant groups

of the IVUS Segmentation Challenge held in the 2011 Computing

and Visualization for (Intra)Vascular Imaging (CVII) workshop of

the Medical Image Computing & Computer Assisted Intervention

(MICCAI) conference. 

The results of the proposed method with the contrast enhance-

ment activated and deactivated were also taken into account in or-

der to evaluate the improvement obtained by adjusting the con-

trast of the input image. 

4.1. Initialization of the parameters 

The following parameters were defined to perform the auto-

matic segmentation: the size of the mask of the median filter was

set to 5x5, which proved to be the most suitable template for re-

moving noise without affecting the edges of the lumen; when the

contrast enhancement was necessary, the value of the α parame-

ter was defined as the difference between the probability of the

low and high intensities of the image, as described by Eq. (4) ; the

number of reductions performed by the Gaussian pyramid was au-

tomatically defined as equal to 2, leading to images with a reso-

lution of 96 × 96 to be processed by the K-means with subtractive

clustering; the radius r a and r b of the subtractive clustering algo-

rithm were set to 1.2 and 1.8, respectively; and the number of iter-

ations of the Chan–Vese active contour algorithm was set to 500,

which was higher when compared to our previous study using MR

images (200 iterations) due to the higher resolution of the IVUS

images under analysis. 
.2. Performance of the proposed method 

Examples of the segmentation results and corresponding man-

al delineations are shown in Fig. 6 . 

In Fig. 6 , the green contours are the results obtained by the

roposed method, whereas the blue, red and yellow contours rep-

esent the related manual delineations. The lumen was correctly

dentified in all images and is very close to the corresponding

anual delineations. Due to the subjective analysis of each expert,

he automatically segmented contour is very close to correspond-

ng manual delineation in some cases, while small differences were

roduced in other cases as shown in Fig. 6 b. 

The effect of the post-processing step on the lumen contour re-

ulting from the Chan–Vese active contour is shown in Fig. 7 . 

As shown in Fig. 7 a, the automatically segmented lumen con-

our can be somewhat irregular and leak from the true boundary

s depicted in the image of the third row. However, the lumen

ontours became more regular and smooth after the final post-

rocessing step, which leads to results very similar to the manual

elineations ( Fig. 7 c). 

The effectiveness of the proposed approach to identify and

liminate side branches of bifurcation regions is apparent in Fig. 8 .

As shown in Fig. 8 b, the segmentation errors resulting from

he automatic lumen segmentation are due to the removal of the

hole bifurcation region from the binary image representing the

ow-intensity values before applying the circularity index in Eq.

10) , leading to the identification of another region as belonging

o the potential lumen. However, as shown in Fig. 8 c, the lumen

as successfully segmented after applying the proposed approach

o identify and split the bifurcation region. 

From the 326 IVUS images used in the experiments, the pro-

osed method successfully segmented 324 images. The average

alues of the Jaccard measure, Hausdorff distance, percentage of

rea difference and Dice coefficient of the automatically segmented

umen in comparison with the three manual delineations of the

orrectly segmented images are indicated in Tables 2 and 3 . Ad-

itionally, the average values of the same four measures obtained

rom the intra- and inter-observer analysis are indicated in Table 4 .

As shown in Tables 2 and 3 , the average values of the Jaccard

easure, percentage of area difference and Dice coefficient for the

anual delineations 2 and 3 are better than those obtained in

omparison to the manual delineation 1. However, the results of

he Hausdorff distance are similar to almost all the manual delin-

ations, except in the case where the adjustment of the contrast

f the images was not applied ( Table 2 ), leading to a decrease of

he Hausdorff distance for the manual delineation 2. As shown in

able 3 , the average value of the Jaccard measure increased from

.87 ± 0.07 to 0.88 ± 0.06 after the application of the contrast

nhancement. Additionally, an increase of the average value of the

ice coefficient from 0.93 ± 0.04 to 0.94 ± 0.04 was also obtained

fter adjusting the contrast of the images. The Hausdorff distance



D.S. Jodas et al. / Medical Image Analysis 40 (2017) 60–79 69 

Fig. 6. Examples of segmentation results obtained by the automatic segmentation method (a–e): The original IVUS images are shown in the first row, and the subsequent 

rows depict the segmentation results and the corresponding manual delineations, respectively. (The contours in green are the ones generated by the proposed method, 

whereas the blue, red and yellow contours represent the manual delineations.). (For interpretation of the references to colour in this figure legend, the reader is referred to 

the web version of this article.) 

Table 2 

Average values of the Jaccard measure ( JM ), Hausdorff distance ( HD ), Percentage of area difference 

( PAD ) and Dice coefficient ( DC ) obtained by the proposed method without the application of the con- 

trast enhancement. 

Manual delineation 1 Manual delineation 2 Manual delineation 3 Average 

JM 0.86 ± 0.07 0.88 ± 0.07 0.88 ± 0.07 0.87 ± 0.07 

HD 0.30 ± 0.17 0.29 ± 0.18 0.30 ± 0.19 0.30 ± 0.18 

PAD 0.12 ± 0.08 0.08 ± 0.07 0.08 ± 0.08 0.09 ± 0.08 

DC 0.92 ± 0.05 0.93 ± 0.04 0.93 ± 0.04 0.93 ± 0.04 

∗The values of the Hausdorff distance ( HD ) are presented in millimeters; Manual delineation 1 rep- 

resents the tracing done by the expert 1, whereas the manual delineations 2 and 3 are the first and 

second tracings of the expert 2, respectively. 

Table 3 

Average values of the Jaccard measure ( JM ), Hausdorff distance ( HD ), Percentage of area difference 

( PAD ) and Dice coefficient ( DC ) obtained by the proposed method with the application of the contrast 

enhancement. 

Manual delineation 1 Manual delineation 2 Manual delineation 3 Average 

JM 0.87 ± 0.06 0.88 ± 0.06 0.88 ± 0.06 0.88 ± 0.06 

HD 0.29 ± 0.15 0.29 ± 0.17 0.29 ± 0.18 0.29 ± 0.17 

PAD 0.11 ± 0.07 0.08 ± 0.07 0.08 ± 0.08 0.09 ± 0.07 

DC 0.93 ± 0.04 0.94 ± 0.04 0.94 ± 0.04 0.94 ± 0.04 

∗The values of the Hausdorff distance ( HD ) are presented in millimeters; Manual delineation 1 rep- 

resents the tracing done by the expert 1, whereas the manual delineations 2 and 3 are the first and 

second tracings of the expert 2, respectively. 



70 D.S. Jodas et al. / Medical Image Analysis 40 (2017) 60–79 

Fig. 7. Illustration of the post-processing step on the lumen contour: a) The contour resulting from the Chan–Vese active contour; b) The result of the smoothing of the 

contour; c) The smoothed contour (in green) along with a corresponding manual delineation (in blue). (For interpretation of the references to colour in this figure legend, 

the reader is referred to the web version of this article.) 

Table 4 

Average values of the Jaccard measure ( JM ), Hausdorff distance ( HD ), Percentage of area difference ( PAD ) 

and Dice coefficient ( DC ) obtained from the intra- and inter-observer analysis. 

Exp 1 vs Exp 2 (1st) Exp 1 vs Exp 2 (2nd) Exp 2 (1st) vs Exp 2 (2nd) Average 

JM 0.88 ± 0.05 0.87 ± 0.05 0.93 ± 0.05 0.89 ± 0.05 

HD 0.28 ± 0.13 0.30 ± 0.13 0.17 ± 0.13 0.25 ± 0.13 

PAD 0.11 ± 0.08 0.13 ± 0.08 0.04 ± 0.05 0.09 ± 0.07 

DC 0.94 ± 0.03 0.93 ± 0.03 0.96 ± 0.03 0.94 ± 0.03 

∗Exp stands for Expert; 1st and 2nd indicate the first and second delineations of Expert 2, respectively; the 

values of the Hausdorff distance ( HD ) are expressed in millimeters. 

 

 

 

 

 

 

 

 

 

 

 

 

7  

m  

v  

l  

e  

A  

m  

t  

t

 

t  

o

also decreased from 0.30 ± 0.18 mm to 0.29 ± 0.17 mm after ad-

justing the contrast of the images. In terms of the intra- and inter-

observer variability, the results of Table 4 show that the proposed

method is also in accordance with the average values computed

from the comparison between Expert 1 and Expert 2. The aver-

age values of the Jaccard measure, Hausdorff distance, percentage

of area difference and Dice coefficient between both experts were

0.89 ± 0.05, 0.25 ± 0.13 mm, 0.09 ± 0.07 and 0.94 ± 0.03, re-

spectively, which are very similar to the results obtained by our

method (see Tables 2 and 3 ). 

The mean values of the lumen area and average lumen diameter

calculated for each manual delineation and for the corresponding

automatically detected lumen are shown in Table 5 . 
The mean lumen area obtained from the proposed method was

.93 ± 3.63 mm 

2 , which is close to the one obtained from the

anual delineation 3 (7.94 ± 3.55 mm 

2 ). Additionally, the mean

alue of the average lumen diameter of the automatically detected

umen was also close to the one obtained from the manual delin-

ation 3 (3.07 ± 0.71 mm and 3.08 ± 0.68 mm, respectively).

lthough the results of the proposed method were closer to the

anual delineation 3, no significant differences were found be-

ween the areas and average diameters of the automatically de-

ected and manually delineated lumen contours. 

A comparison between the proposed automatic lumen segmen-

ation and the methods reported in the literature was also carried

ut, which led to the results shown in Tables 6 and 7 . 



D.S. Jodas et al. / Medical Image Analysis 40 (2017) 60–79 71 

Fig. 8. Examples of segmentation results after applying the proposed approach to identify bifurcation regions: a) The original IVUS images; b) The result of the automatic 

lumen segmentation (in green) along with the corresponding manual delineations (in blue, red and yellow) before the identification of the bifurcation region; c) The result 

of the automatic lumen segmentation (in green) along with the corresponding manual delineations (in blue, red and yellow) after the identification of the bifurcation region. 

(For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) 

Table 5 

Mean values of the lumen area ( LA ) and average lumen diameter ( ALD ) for the manual delineations 

and the automatically segmented lumen after adjusting the contrast of the images. 

Manual delin. 1 Manual delin. 2 Manual delin. 3 Proposed method 

Mean LA ( mm 

2 ) 8.75 ± 3.73 8.03 ± 3.56 7.94 ± 3.55 7.93 ± 3.63 

Mean ALD (mm) 3.11 ± 0.63 3.10 ± 0.68 3.08 ± 0.68 3.07 ± 0.71 

∗Manual delin. 1 represents the delineation done by Expert 1, whereas manual delin. 2 and manual 

delin. 3 are the first and second delineations of Expert 2, respectively. 

 

c  

C  

e  

P  

c  

o  

a  

t  

s  

P  

a  

0  

c

0

 

s  

e  

w  

c

 

t  

w  

s

 

e  

o  

s  

c  

tion of the lumen. 
Table 6 shows the average measures of the general performance

alculated from each participant group of the IVUS Segmentation

hallenge proposed at the MICCAI 2011 CVII workshop ( Balocco

t al., 2014 ). Table 6 demonstrates that the method proposed by

articipant group 3 outperformed all the others presented at the

hallenge. The performance of our method is comparable to the

ne proposed by this group. Regarding the results obtained after

djusting the contrast of the input images, the average value of

he Jaccard measure obtained from the proposed automatic lumen

egmentation was 0.88 ± 0.06 against 0.88 ± 0.05 presented by

articipant group 3 at the IVUS Segmentation Challenge; the aver-

ge Hausdorff distance of our method was 0.29 ± 0.17 mm against

.34 ± 0.14 mm of this participant group; and the average per-

entage of area difference obtained from our method was 0.09 ±
.07 against 0.06 ± 0.05 of the same group. 
Table 7 shows the comparison between our automatic lumen

egmentation approach and related methods proposed in the lit-

rature. The comparison shows that our method is in accordance

ith the ones proposed in the studies found, achieving results

omparable or even better than the ones of those methods. 

The computational cost of the most important procedures in

he segmentation and identification stages of the proposed method

hen applied to each image of the dataset used in this study is

hown in Table 8 . 

The execution time of the Chan–Vese active contour is the high-

st among the other algorithms involved, representing about 68%

f the total time. The K-means with subtractive clustering repre-

ents about 27% of the total time, whereas the smoothing of the

ontour takes only 6% of the whole segmentation and identifica-



72 D.S. Jodas et al. / Medical Image Analysis 40 (2017) 60–79 

Table 6 

Jaccard measure ( JM ), Hausdorff distance ( HD ) and percentage of area difference ( PAD ) obtained from the general 

performance of the IVUS Segmentation Challenge for the test images acquired at 20 MHz ( Balocco et al., 2014 ). 

P1 P2 P3 P4 P5 P7 P8 

JM 0.81 ± 0.12 0.83 ± 0.08 0.88 ± 0.05 0.77 ± 0.09 0.79 ± 0.08 0.84 ± 0.08 0.81 ± 0.09 

HD 0.47 ± 0.39 0.51 ± 0.25 0.34 ± 0.14 0.47 ± 0.22 0.46 ± 0.30 0.38 ± 0.26 0.42 ± 0.22 

PAD 0.14 ± 0.13 0.14 ± 0.12 0.06 ± 0.05 0.15 ± 0.12 0.16 ± 0.09 0.11 ± 0.12 0.11 ± 0.11 

∗P stands for the participant group; Participant group 6 (P6) did not perform the segmentation of the lumen contour 

and it explains the absence of the average values of this participant group; the values of the Hausdorff ( HD ) distance 

are presented in millimeters. 

Table 7 

Average measures obtained from the proposed method and the related ones found in the literature. 

Authors JM HD PAD DC 

Lo Vercio et al. (2016) 0.830 0 ± 0.050 0 – 0.180 0 ± 0.060 0 –

Su et al. (2016) 0.9182 0.2243 – –

Sofian et al. (2015) 0.8624 ± 0.0193 0.54 4 4 ± 0.1290 0.0645 ± 0.0509 0.9260 ± 0.0111 

Destrempes et al. (2014) – 0.330 0 ± 0.070 0 – –

Mendizabal-Ruiz et al. (2013) a 0.8671 ± 0.0341 0.1398 ± 0.0384 – 0.9283 ± 0.0201 

Vard et al. (2012) – 0.3044 ± 0.1853 – –

Cardinal et al. (2010) b – 0.430 0 ± 0.30 0 0 – –

Taki et al. (2008) – 0.7081 ± 0.2491 – –

Proposed method 0.880 0 ± 0.060 0 0.290 0 ± 0.170 0 0.090 0 ± 0.070 0 0.940 0 ± 0.040 0 

∗JM = Jaccard measure; HD = Hausdorff distance; PAD = Percentage of area difference; DC = Dice coefficient. 
a Average values of the comparison between the method and the manual delineations of two observers. 
b Result from the pre-interventional group. The HD for the post-interventional and follow-up groups were 0.46 

± 0.26 and 0.40 ± 0.21, respectively. (See the study of Cardinal et al. (2010) for more details.) 

Table 8 

Computational cost of the most important procedures of the proposed method when applied to 

each image (in seconds). 

K-means Chan–Vese active contour Smoothing of the contour Total time 

Average 1.5279 3.8694 0.3192 5.7165 

Std 0.2322 1.1645 0.1452 1.5419 

∗Std = Standard deviation 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

e  

−  

i  

t  

2  

t  

e  

t  

u  

−  

r

 

e  

t  

−  

e  

d  

−  

t  

w  

t

 

t

5

 

t  

w  

p  
Linear regression analysis for the lumen area and average lu-

men diameter showed a high correlation between the automatic

segmentations and manual delineations, as shown in Figs. 9 and

10 . 

For the lumen area, the correlation coefficients of the proposed

method compared to the manual delineations 1, 2 and 3 were

0.985096, 0.975378 and 0.974 4 40, respectively, before adjusting

the contrast of the input images; while for the case when the

contrast adjustment was applied, the correlation coefficients were

0.985729, 0.974483 and 0.974314, respectively. A slight improve-

ment in the correlation coefficients between the automatic seg-

mentations and the manual delineations of Expert 1 was obtained

after the application of the contrast enhancement. Regarding the

average lumen diameter, the correlation coefficients between the

automatic segmentations and manual delineations 1, 2 and 3 were

0.977080, 0.972486 and 0.971514, respectively, before adjusting the

contrast of the input images. When the contrast enhancement was

applied, the correlation coefficients were 0.977346, 0.972354 and

0.971756, respectively. Similar to the lumen area, a slight improve-

ment in the correlation coefficients of the average lumen diameter

between the automatic segmentations and manual delineations 1

and 3 was also achieved after adjusting the contrast of the input

images. 

The difference between the area calculated by the automatically

segmented lumen and the ones obtained from the three manual

delineations is depicted in Fig. 11 by Bland–Altman plots. For the

average lumen diameter, the difference between the automatic and

manual segmentations is represented in Fig. 12 . 

Fig. 11 shows a significant underestimation of the lumen area

of the automatic segmentation compared to the manual delin-
m  
ation 1, leading to average differences of −0.89136 mm 

2 and

0.82174 mm 

2 before and after adjusting the contrast of the input

mages, respectively. In contrast, the average difference between

he automatically segmented lumen and the manual delineations

 and 3 is small. The average differences of the lumen area be-

ween the automatically segmented lumen and the manual delin-

ation 2 were −0.16389 mm 

2 and −0.09497 mm 

2 before and af-

er adjusting the contrast of the images, respectively. For the man-

al delineation 3, the average differences were −0.07421 mm 

2 and

0.00808 mm 

2 before and after the adjustment of the contrast,

espectively. 

Regarding the average lumen diameter, a slight average differ-

nce was found between the automatic and manual segmenta-

ions. For the manual delineation 1, the average differences were

0.05160 mm and −0.03283 mm before and after the contrast

nhancement of the input images, respectively; for the manual

elineation 2, the average differences were −0.04473 mm and

0.02680 mm before and after the contrast enhancement, respec-

ively; and for the manual delineation 3, the average differences

ere −0.02701 mm and −0.00963 mm before and after the con-

rast enhancement, respectively. 

The linear regression and Bland–Altman analysis obtained from

he delineations of the two experts is depicted in Fig. 13 . 

. Discussion 

The development of automatic segmentation methods applied

o medical images plays an important role in providing experts

ith auxiliary diagnosis tools for identifying various types of

athological conditions. For example, the segmentation of the lu-

en and media-adventitia regions in IVUS images represents an



D.S. Jodas et al. / Medical Image Analysis 40 (2017) 60–79 73 

Fig. 9. Area of the automatically segmented lumen versus the three corresponding manual delineations before and after adjusting the contrast of the input images. 

i  

r

 

a  

u  

t  

p  

p  

B  

l  

c  

i  

t  

v  

o  

m  

B  

fi  

t  

p

5

 

c  

i  

a  

m  

f  

p  

i

p  

f  

i  

j  

g  

o  

m  

i  

t  

p  
mportant step to quickly identify and quantify possible atheroscle-

osis in arteries. 

The automatic segmentation of the lumen region in IVUS im-

ges of coronary arteries was successfully tackled in this study. The

se of unsupervised classification and circularity index to identify

he lumen region is the kernel of the proposed method. In our

revious work, the initial version of the proposed method was ap-

lied to identify the lumen region in MR images of carotid arteries.

ased on the fact that the lumen is a circular-shaped region with

ow-intensity values in axial black-blood MR images, the use of a

ircularity index was proposed to identify the region correspond-

ng to the correct lumen among those obtained from the subtrac-

ive clustering algorithm. One goal of this study was to assess the

iability of the method for segmenting the lumen in IVUS images

f coronary arteries. Additionally, improvements were developed to

ake the method more robust, flexible, efficient and competent.

riefly, the original empirical criterion concerning noise was rede-

ned, side branches in bifurcations regions are now successfully

ackle, and the segmentation accuracy has been considerably im-

roved. 
I

a  
.1. Initialization of the parameters 

The values of the parameters adopted in our previous study

oncerning MR images were also employed here to show the abil-

ty of the proposed method in segmenting the lumen in IVUS im-

ges. The mask of the median-filter used has also been used in

any studies related to medical image segmentation; which there-

ore, indicates that it is a suitable choice to attenuate the noise

resent in the input images without leading to excessive smooth-

ng of the borders of the structures of interest. The value of the α
arameter in Eq. (2) is automatically calculated by using the dif-

erence between the probabilities of the low and high grayscale

ntensities of the input image as given by Eq. (4) . Hence, the ad-

ustment of the contrast is not necessary if more pixels with high

rayscale intensities are present in the input image. The number

f clusters that are defined has an important role in the proposed

ethod. In most cases, the grayscale intensity of the lumen in IVUS

mages is well-defined and distinguishable from other structures in

he images. Hence, the number of clusters initially proposed in our

revious study is also suitable in the segmentation of the lumen in

VUS images. On using the subtractive clustering algorithm, the r a 
nd r parameters may affect the number of clusters to be gener-
b 



74 D.S. Jodas et al. / Medical Image Analysis 40 (2017) 60–79 

Fig. 10. Average lumen diameter of the automatically segmented lumen versus the three manual delineations before and after adjusting the contrast of the input images. 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

c  

r  

l  

s  

o  

i  

t  

t  

p  

t  

a

 

a  

r  

l  

i  

g  

i  

w  

p  

i  

t  
ated ( Chiu, 1994 ). However, since the subtractive clustering is only

used to generate the centroids of the expected number of clusters,

the values of these parameters can also be used to separate the

regions in IVUS images according to the same expected number.

Regarding the downsampling of the resolution of the input IVUS

image, the Gaussian pyramid was implemented in such way that it

is automatically performed until the total number of pixels of the

input image is equal or less than 22,500, which was found as ap-

propriate to reduce the computational cost of the subtractive clus-

tering algorithm. Since all the IVUS images used have the same

resolution ( 384 × 384 ), the downsampling process was performed

twice, leading to a final image resolution of 96 × 96 (9216 pixels). 

5.2. Performance of the proposed method 

When compared to the related studies found in the litera-

ture, the proposed automatic segmentation of the lumen does not

require any kind of user interaction and is easily implemented

without using complex algorithms. The K-means with subtractive

clustering is used to separate the regions of the input IVUS im-

age according to the expected number of clusters. The subtractive

clustering is only used as a prior step to find the appropriate initial
entroids to be used in the K-means clustering algorithm. Different

esults could be produced when the centroids are randomly se-

ected and used in the K-means clustering algorithm. Since the re-

ult of the K-means clustering algorithm depends on the selection

f the initial centroids, we decided to use the subtractive cluster-

ng due to its stability to find the same initial centroids even when

he algorithm is executed several times. Additionally, the subtrac-

ive clustering algorithm is easily implemented, although the com-

utational time increases in images with higher resolution. Once

he initial centroids are found, the traditional K-means clustering

lgorithm is applied to the input image. 

Because the low-intensity values are associated to the lumen

nd background regions, the connected component labelling algo-

ithm is applied to obtain all regions of the cluster belonging to the

ow-intensity values as a binary image. The regions correspond-

ng to the background of the IVUS image are at the border of the

enerated binary image. In our previous work, a term represent-

ng the number of pixels of the region at the border of the image

as added to reduce its circularity index. Here, the region is sim-

ly discarded from the binary image to avoid additional process-

ng. Additionally, the morphological opening operation with adap-

ive size of the structuring element is proposed to remove noisy



D.S. Jodas et al. / Medical Image Analysis 40 (2017) 60–79 75 

Fig. 11. Bland–Altman plots of the lumen area of the automatically segmented lumen and the three corresponding manual delineations. 

a  

v  

s  

s  

c

 

t  

I  

f  

b  

a  

t  

c  

g  

d  

t  

m  

p  

t  

s

 

S  

s  

o  

c  

b  

a  

i  

v  

w  

t  

o  

a  

H

0  

a  
rtefacts without the need of a heuristic criterion as in the initial

ersion. The morphological operations in the new post-processing

tep of the proposed method revealed that they were effective in

moothing the lumen contour resulting from the Chan–Vese active

ontour and removing possible irregularities. 

An important contribution of this study is an effective approach

o identify regions corresponding to side branches in the input

VUS images. The low-intensity values of the side branch extend

rom the lumen region to the border of the IVUS images when a

ifurcation is presented. Hence, the bifurcation is represented by

 single region with pixels at the border of the input image when

he cluster with low-intensity values is generated by the K-means

lustering algorithm. In order to avoid the elimination of this re-

ion in the lumen identification step, an approach based on the

elimitation of the potential lumen region by a circle centred on

he centre of the input IVUS image is proposed to separate and re-

ove the branch of the bifurcation region. The proposed approach

roved to be effective in eliminating the side branches of bifurca-
t  
ion regions and therefore, in decreasing the number of erroneous

egmentations. 

The average values of the measures described in

ection 3.3 were obtained from the 324 IVUS images successfully

egmented by the proposed method. The contrast enhancement

f the IVUS images can improve the brightness of the region

orresponding to the intima layer, allowing a better distinction

etween this region and the lumen. Hence, the segmentation

ccuracy obtained before and after adjusting the contrast of the

nput image were compared. A slight improvement in the average

alues of the Jaccard measure and Dice coefficient was found

hen the contrast enhancement was applied. The average value of

he Jaccard measure was 0.87 ± 0.07 before adjusting the contrast

f the images, whereas a slight increasing to 0.88 ± 0.06 was

chieved after the application of the contrast. Additionally, the

ausdorff distance decreased from 0.30 ± 0.18 mm to 0.29 ±
.17 mm after adjusting the contrast. Regarding the Bland–Altman

nalysis, the average differences between the area calculated from

he proposed method and the ones obtained from the manual



76 D.S. Jodas et al. / Medical Image Analysis 40 (2017) 60–79 

Fig. 12. Bland–Altman plots of the average lumen diameter of the automatically segmented lumen and the three corresponding manual delineations. 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

t  

t  

a  

t  

t  

p  

g  

t  

s  

i  

w  

n  

l

5

 

t  
delineations reduced when the contrast enhancement was applied.

Similar reductions of the bias were also obtained for the average

lumen diameter. When compared to the manual delineation 3,

a reduction of the bias from −0.02701 mm to −0.00963 mm

was obtained before and after adjusting the contrast, respectively,

leading to an average distance less than one pixel between the

automatic and manual segmentations. 

The high complexity ( O ( d 2 , N 

2 )) of the subtractive clustering al-

gorithm makes it unfeasible to be applied in images with higher

resolutions due to the high number of pixels to be processed to

find the centroids of the clusters. Hence, the reduction of the reso-

lution of the input image in order to decrease the computational

cost of the subtractive clustering algorithm was tackled in this

study. The Gaussian pyramid proved to be the most suitable choice

due to its simple implementation and effective reduction of the

resolution of the input image without losing important informa-

tion on the structure of interest. The effectiveness of the Gaussian

pyramid can be perceived from the data in Table 8 , which indicates

s  
hat the execution time of the K-means with the subtractive clus-

ering algorithm is low when compared to the average total time

nd the time required by the Chan–Vese algorithm. This is due to

he reduction of the number of pixels to be processed by the sub-

ractive clustering algorithm when the Gaussian pyramid was ap-

lied. The computational cost of the Chan–Vese active contour was

reater than that of the other algorithms. Since re-initialization of

he signed distance function of the contour is necessary at every

tep of the contour evolution, the time consumed by the algorithm

ncreases when applied to images of higher resolution. However, it

as decided to apply the Chan–Vese active contour to the origi-

al IVUS image in order to take into account all the pixels of the

umen region available in the full resolution of the input image. 

.3. Comparison with the intra- and inter-observer variability 

In terms of the inter-observer variability, i.e. the comparison be-

ween the manual delineations of Expert 1 and Expert 2, the re-

ults are close to the ones computed from our method. The aver-



D.S. Jodas et al. / Medical Image Analysis 40 (2017) 60–79 77 

Fig. 13. Linear regression and Bland–Altman analysis concerning the lumen area and average lumen diameter obtained from the delineations of the two experts. 

a  

o  

t  

0  

t  

p  

a  

d  

±  

b  

l  

H  

c  

±  

s  

l  

T  

±  

t

 

o  

i  

t  

a  

r  

i  

m  

t  

m  

0  

m  

e  

t  

o  

t  

s  

c  

l  

0  

e  

2  

b  

v  

fi  

t  

T  

i  

t  
ge values of the Jaccard measure, Hausdorff distance, percentage

f area difference and Dice coefficient of Expert 1 in comparison to

he first delineation of Expert 2 were 0.88 ± 0.05, 0.28 ± 0.13 mm,

.11 ± 0.08 and 0.94 ± 0.03, respectively. The average values for

he same measures obtained from the comparison between the

roposed method after adjusting the contrast of the input images

nd the first delineation of Expert 2 (represented by the manual

elineation 2 in Table 3 ) were 0.88 ± 0.06, 0.29 ± 0.17 mm, 0.08

0.07 and 0.94 ± 0.04, respectively. Regarding the comparison

etween the manual delineation of Expert 1 and the second de-

ineation of Expert 2, the average values of the Jaccard measure,

ausdorff distance, percentage of area difference and Dice coeffi-

ient were 0.87 ± 0.05, 0.30 ± 0.13 mm, 0.13 ± 0.08 and 0.93

0.03, respectively. The results of the proposed method for the

ame measures obtained from the comparison with the second de-

ineation of Expert 2 (represented by the manual delineation 3 in

able 3 ) were 0.88 ± 0.06, 0.29 ± 0.18 mm, 0.08 ± 0.08 and 0.94

0.04, respectively, which were slightly better than the ones ob-

ained between Expert 1 and Expert 2 (second delineation). 

The regression analysis showed that the segmentation results

f the proposed method are also close to ones obtained from the

ntra- and inter-observer variability. In terms of the lumen area,

he Pearson correlation coefficients between Expert 1 and the first

nd second delineations of Expert 2 were 0.987866 and 0.986466,
espectively. The regression analysis of the intra-observer variabil-

ty showed a correlation coefficient of 0.992330 between the lu-

en areas of the first and second delineations of Expert 2. After

he application of the contrast enhancement, the automatic seg-

entation method obtained a correlation coefficient of 0.985729,

.974483 and 0.974314 with the lumen area calculated from the

anual delineations of Expert 1 and the first and second delin-

ations of Expert 2, respectively. For the average lumen diameter,

he correlation coefficients between Expert 1 and the first and sec-

nd delineations of Expert 2 were 0.981065 and 0.981453, respec-

ively, whereas the correlation coefficient between the first and

econd delineations of Expert 2 was 0.990351. After adjusting the

ontrast of the input images, the automatic segmentation of the

umen obtained a correlation coefficient of 0.977346, 0.972354 and

.971756 with the average lumen diameter of the manual delin-

ations of Expert 1 and the first and second delineations of Expert

, respectively. The Bland–Altman analysis showed similar results

etween the proposed method and the intra- and inter-observer

ariability. For the lumen area, the bias between Expert 1 and the

rst delineation of Expert 2 were 0.708995 mm 

2 , whereas for

he second delineation of Expert 2 the bias was 0.803415 mm 

2 .

he Bland–Altman analysis of the lumen area with respect to the

ntra-observer variability showed a bias of 0.094420 mm 

2 be-

ween the first and second delineations of Expert 2. The bias be-



78 D.S. Jodas et al. / Medical Image Analysis 40 (2017) 60–79 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

a  

o  

J  

±
 

t  

i  

m

1

6  

w  

o

5

 

r  

K  

p  

s  

t  

p  

g  

e  

m  

a  

p

 

a  

i  

a  

t  

i  

q  

t  

t

6

 

I  

t  

r  

o  

p  

b  

d  

b  

e  

p  

i  

i  

t

 

t  

a  

d  

o  

s  

y  

t  

d  

m  

o

 

t  
tween the proposed method after adjusting the contrast of the

images and the first delineation of Expert 2 was -0.09497 mm 

2 ,

which is closer to zero when compared to the bias between Ex-

pert 1 and the same delineation of Expert 2 (0.708995 mm 

2 ).

For the average lumen diameter, the bias between Expert 1 and

the first delineation of Expert 2 was 0.006241 mm, whereas for

the second delineation of Expert 2 the bias was 0.024817 mm.

For the intra-observer variability, the bias of the average lumen

diameter between the first and second delineations of Expert 2

was 0.018575 mm. The bias of the average lumen diameter be-

tween the proposed method after adjusting the contrast of the

images and the three manual delineations were -0.03283 mm, -

0.02680 mm and -0.00963 mm. Compared to the manual delin-

eations, the proposed method underestimated the average lumen

diameter. However, the results were similar to the intra- and inter-

observer variability. 

5.4. Comparison with other methods 

A comparison with related studies found in the literature was

carried out in order to validate the accuracy of the proposed lumen

segmentation method. The first comparison was performed against

the results obtained in the IVUS Segmentation Challenge ( Balocco

et al., 2014 ). The average values described in Table 6 showed that

the method proposed by Participant group 3 outperformed the

ones obtained from the other participant groups. A slight improve-

ment of the Hausdorff distance was obtained from our method

when compared to the one of Participant group 3 of the challenge

(0.29 ± 0.17 mm vs 0.34 ± 0.14 mm, respectively). Although no

significant improvements were achieved when compared to the re-

sults of this participant group, the segmentation of the lumen per-

formed by our method is fully automatic and there is no need for

an initial user interaction. In contrast, the method proposed by this

group requires the initialization of a number of points to generate

the initial contours. 

The results shown in Table 7 indicate that our method outper-

formed many of the studies found in the literature. The value of

the Dice coefficient of our method (0.9400 ± 0.0400) is better than

the values obtained by Sofian et al. (2015) and Mendizabal-Ruiz

et al. (2013) (0.9260 ± 0.0111 and 0.9283 ± 0.0201, respectively).

The value of the Jaccard measure (0.8800 ± 0.0600) was also bet-

ter than the ones obtained from the studies under comparison, ex-

cept for the value obtained by Su et al. (2016) (0.9182). The value

of the Hausdorff distance (0.2900 ± 0.1700 mm) is only greater

than the ones obtained from Su et al. (2016) and Mendizabal-Ruiz

et al. (2013) (0.2243 mm and 0.1398 ± 0.0384 mm, respectively).

Finally, the value of the percentage of area difference (0.0900 ±
0.0700) is only greater than the one obtained from Sofian et al.

(2015) (0.0645 ± 0.0509). 

Although the results presented by Su et al. (2016) are better

than the ones obtained from our method, a ROI surrounding the

region corresponding to the media-adventitia region must be de-

lineated in the input image before the classification by the two

ANN. In addition, the elimination of noise of the binary image re-

sulting from the first classification is performed by a second ANN.

In contrast, morphological operations and the convex hull algo-

rithm are used here to identify noise and refine the regions to be

evaluated instead of using a complex ANN. In addition, the method

proposed performs the lumen segmentation in the whole image.

Although the value of the percentage of area difference of the

method proposed by Sofian et al. (2015) is less than the one ob-

tained from the proposed method, the authors only used 30 IVUS

images randomly selected from the dataset provided by Balocco

et al. (2014) to perform the experiments. In the study tackled by

Mendizabal-Ruiz et al. (2013) , IVUS images acquired from catheters

operating at frequencies of 20 MHz and 40 MHz were taken into
ccount. However, if only the images acquired from the frequency

f 20 MHz were considered, then the average values of the Dice,

accard and Hausdorff distance would be 0.9215 ± 0.0217, 0.8555

0.0363 and 0.1421 ± 0.0371 mm, respectively. 

In terms of the computational cost, the average total time of

he proposed method is also in accordance with the related stud-

es found in the literature. The average total time of the proposed

ethod to process each one of the 326 IVUS images was 5.7165 ±
.5419 s. Gao et al. (2015) reported an average total time of 8.13 ±
.05 s for the lumen segmentation step of their automated frame-

ork. The average total time of the lumen segmentation carried

ut by Mendizabal-Ruiz et al. (2013) was 4.51 s. 

.5. Limitations 

The proposed method has two limitations. The first limitation

egards the number of parameters of the clustering algorithm. The

-means algorithm with subtractive clustering used requires three

arameters: the number of clusters and the radius r a and r b repre-

enting the neighbourhood of the pixels under analysis. Although

he proposed method with the same parameters adopted in our

revious study for MR images was able to identify the lumen re-

ion in IVUS images, a fully automatic segmentation without any

mpirical parameters increases the reliability and robustness of the

ethod. Hence, future studies will be conducted to effectively sep-

rate the regions present in the input images without any kind of

arameter. 

The second limitation regards the segmentation of IVUS im-

ges acquired from a frequency of 40 MHz. Since speckle noise

n these images is higher when compared to IVUS images acquired

t a frequency of 20 MHz, the application of most efficient filters

o minimize the effects of such noisy artefacts will be considered

n future research projects to enable segmentation on images ac-

uired from catheters operating at different frequencies. In addi-

ion, texture analysis will also be considered instead of using only

he grayscale intensity of the pixels in the K-means algorithm. 

. Conclusions 

The segmentation of the lumen and media-adventitia regions in

VUS images is an intensive focus of research and plays an impor-

ant role in assessing the presence and progression of atheroscle-

osis. A fully automatic segmentation of the lumen in IVUS images

f coronary arteries was proposed in this article. Compared to our

revious study, new solutions were developed to enhance the ro-

ustness, efficiency and automaticity of the proposed method. Ad-

itionally, a new approach to successfully identify and remove side

ranches of bifurcation regions was also proposed to avoid the

limination of the potential lumen regions from the subsequent

rocessing steps. The improved method proved to be effective in

dentifying the regions corresponding to the lumen without user

nteraction and any change in the values of the method parame-

ers. 

Modifications were also accomplished to improve the shape of

he lumen contour and the segmentation accuracy. The qualitative

nalysis showed that the visual shape of the lumen contour pro-

uced by the new contour correction step was better than the one

btained from the Chan–Vese active contour algorithm, leading to

moother and more regular final contours. The quantitative anal-

sis demonstrated that the segmentation results of the new au-

omatic segmentation method are in accordance with the manual

elineations performed by two experts. Additionally, the proposed

ethod showed results close to the ones obtained from the inter-

bserver variability. 

An effective approach to reduce the number of parameters of

he subtractive clustering algorithm, as well as the application of



D.S. Jodas et al. / Medical Image Analysis 40 (2017) 60–79 79 

t  

o  

f

A

 

m  

B

 

F  

a  

R

s

R

A  

B  

 

 

 

 

 

 

B  

B  

 

B  

 

 

 

 

C  

 

 

C  

 

C  

C  

D  

 

 

D  

 

D  

 

G  

 

 

H  

H  

 

J  

 

J  

 

K  

K  

 

 

K  

L  

 

 

M  

 

 

d  

 

 

M  

M  

 

 

M  

 

N  

 

R  

S  

 

 

d  

 

 

S  

 

 

S  

 

S  

 

 

T  

 

 

V  

 

 

W  

 

he method proposed here to IVUS images acquired from catheters

perating at different frequencies, are expected to be addressed in

uture research work. 

cknowledgements 

This work was partially funded by Coordenação de Aperfeiçoa-

ento de Pessoal de Nível Superior ( CAPES ), funding agency in

razil, under the PhD Grant with reference number 0543/13-6 . 

The authors thank the funding of Project NORTE-01-0145-

EDER-0 0 0 022 - SciTech - Science and Technology for Competitive

nd Sustainable Industries, co-financed by “Programa Operacional

egional do Norte” (NORTE2020), through “Fundo Europeu de De- 

envolvimento Regional” (FEDER). 

eferences 

delson, E.H. , Anderson, C.H. , Bergen, J.R. , Burt, P.J. , Ogden, J.M. , 1984. Pyramid

methods in image processing. RCA Engineer 29 (6), 33–41 . 
alocco, S., Gatta, C., Ciompi, F., Wahle, A., Radeva, P., Carlier, S., Unal, G., Sanidas, E.,

Mauri, J., Carillo, X., Kovarnik, T., Wang, C.-W., Chen, H.-C., Exarchos, T.P.,

Fotiadis, D.I., Destrempes, F., Cloutier, G., Pujol, O., Alberti, M., Mendizabal-
Ruiz, E.G., Rivera, M., Aksoy, T., Downe, R.W., Kakadiaris, I.A., 2014. Standard-

ized evaluation methodology and reference database for evaluating IVUS im-
age segmentation. Comput. Med. Imaging Graph. 38 (2), 70–90 . Special Issue

on Computing and Visualisation for Intravascular Imaging. http://dx.doi.org/10.
1016/j.compmedimag.2013.07.001 

ataineh, K. , Naji, M. , Saqer, M. , 2011. A comparison study between various fuzzy

clustering algorithms. Editorial Board 5 (4), 335–343 . 
erg, M.d. , Cheong, O. , Kreveld, M.v. , Overmars, M. , 2008. Computational geometry:

Algorithms and applications, third ed. Springer-Verlag TELOS, Santa Clara, CA,
USA . 

roersen, A., de Graaf, M.A., Eggermont, J., Wolterbeek, R., Kitslaar, P.H., Dijkstra, J.,
Bax, J.J., Reiber, J.H.C., Scholte, A.J., 2016. Enhanced characterization of calci-

fied areas in intravascular ultrasound virtual histology images by quantifica-

tion of the acoustic shadow: validation against computed tomography coro-
nary angiography. Int. J. Cardiovasc. Imaging 32 (4), 543–552. doi: 10.1007/

s10554-015-0820-x . 
ardinal, M.H.R., Meunier, J., Soulez, G., Maurice, R.L., Therasse, E., Cloutier, G.,

2006. Intravascular ultrasound image segmentation: a three-dimensional fast-
marching method based on gray level distributions. IEEE Trans. Med. Imaging

25 (5), 590–601. doi: 10.1109/TMI.2006.872142 . 
ardinal, M.-H.R. , Soulez, G. , Tardif, J.-C. , Meunier, J. , Cloutier, G. , 2010. Fast-march-

ing segmentation of three-dimensional intravascular ultrasound images: a pre-

and post-intervention study. Med. Phys. 37 (7), 3633–3647 . 
han, T.F., Vese, L.A., 2001. Active contours without edges. IEEE Trans. Image Process.

10 (2), 266–277. doi: 10.1109/83.902291 . 
hiu, S.L. , 1994. Fuzzy model identification based on cluster estimation. J. Intell.

Fuzzy Syst. 2 (3), 267–278 . 
estrempes, F., Roy Cardinal, M.-H., Allard, L., Tardif, J.-C., Cloutier, G., 2014. Seg-

mentation method of intravascular ultrasound images of human coronary arter-

ies. Comput. Med. Imaging Graph. 38 (2), 91–103. doi: 10.1016/j.compmedimag.
2013.09.004 . 

hanachandra, N., Manglem, K., Chanu, Y.J., 2015. Image segmentation using K-
means clustering algorithm and subtractive clustering algorithm. Procedia Com-

put. Sci. 54, 764–771. doi: 10.1016/j.procs.2015.06.090 . 
iethrich, E.B. , Irshad, K. , Reid, D.B. , 2006. Virtual histology and color flow intravas-

cular ultrasound in peripheral interventions. In: Seminars in Vascular Surgery,

19. Elsevier, pp. 155–162 . 
ao, Z., Hau, W.K., Lu, M., Huang, W., Zhang, H., Wu, W., Liu, X., Zhang, Y.-T.,

2015. Automated framework for detecting lumen and media-adventitia borders
in intravascular ultrasound images. Ultrasound Med. Biol. 41 (7), 2001–2021.

doi: 10.1016/j.ultrasmedbio.2015.03.022 . 
uang, C., Zeng, L., 2015. An active contour model for the segmentation of images

with intensity inhomogeneities and bias field estimation. PLOS ONE 10 (4), 1–

24. doi: 10.1371/journal.pone.0120399 . 
uang, S.-C., Cheng, F.-C., Chiu, Y.-S., 2013. Efficient contrast enhancement using
adaptive gamma correction with weighting distribution. IEEE Trans. Image Pro-

cess. 22 (3), 1032–1041. doi: 10.1109/TIP.2012.2226047 . 
ain, S., Jagtap, V., Pise, N., 2015. Computer aided melanoma skin cancer detection

using image processing. Procedia Comput. Sci. 48, 735–740. doi: 10.1016/j.procs.
2015.04.209 . 

odas, D.S., Pereira, A.S., Tavares, J.M.R., 2016. Lumen segmentation in magnetic res-
onance images of the carotid artery. Comput. Biol. Med. 79, 233–242. http:

//dx.doi.org/10.1016/j.compbiomed.2016.10.021 . 

ass, M., Witkin, A., Terzopoulos, D., 1988. Snakes: active contour models. Int. J.
Comput. Vis. 1 (4), 321–331. doi: 10.10 07/BF0 0133570 . 

atouzian, A., Angelini, E.D., Carlier, S.G., Suri, J.S., Navab, N., Laine, A.F., 2012. A
state-of-the-art review on segmentation algorithms in intravascular ultrasound

(IVUS) images. IEEE Trans. Inf. Technol. Biomed. 16 (5), 823–834. doi: 10.1109/
TITB.2012.2189408 . 

önig, A., Klauss, V., 2007. Virtual histology. Heart 93 (8), 977–982. doi: 10.1136/hrt.

2007.116384 . 
o Vercio, L., Orlando, J.I., del Fresno, M., Larrabide, I., 2016. Assessment of image

features for vessel wall segmentation in intravascular ultrasound images. Int. J.
Comput. Assist. Radiol. Surg. 11 (8), 1397–1407. doi: 10.1007/s11548-015-1345-4 .

a, Z., Tavares, J.M.R.S., Jorge, R.N., Mascarenhas, T., 2010. A review of algorithms
for medical image segmentation and their applications to the female pelvic

cavity. Comput Methods Biomech Biomed Engin 13 (2), 235–246. doi: 10.1080/

10255840903131878 . 
e Macedo, M.M.G., Takimura, C.K., Lemos, P.A., Gutierrez, M.A., 2016. A robust fully

automatic lumen segmentation method for in vivo intracoronary optical coher-
ence tomography. Res. Biomed. Eng. 32 (1), 35–43. doi: 10.1590/2446-4740.0759 .

endis, S. , Puska, P. , Norrving, B. , 2011. Global Atlas on Cardiovascular Disease Pre-
vention and Control. World Health Organization . 

endizabal-Ruiz, E.G., Rivera, M., Kakadiaris, I.A., 2013. Segmentation of the lu-

minal border in intravascular ultrasound b-mode images using a probabilistic
approach. Med. Image Anal. 17 (6), 649–670. http://dx.doi.org/10.1016/j.media.

2013.02.003 . 
oraes, M.C., Furuie, S.S., 2011. Automatic coronary wall segmentation in intravas-

cular ultrasound images using binary morphological reconstruction. Ultrasound
Med. Biol. 37 (9), 1486–1499. doi: 10.1016/j.ultrasmedbio.2011.05.018 . 

air, A. , Kuban, B.D. , Tuzcu, E.M. , Schoenhagen, P. , Nissen, S.E. , Vince, D.G. , 2002.

Coronary plaque classification with intravascular ultrasound radiofrequency
data analysis. Circulation 106 (17), 2200–2206 . 

itter, N., Cooper, J., 2009. New resolution independent measures of circularity. J.
Math. Imaging Vis. 35 (2), 117–127. doi: 10.1007/s10851- 009- 0158- x . 

antos, A.M.F., dos Santos, R.M., Castro, P.M.A., Azevedo, E., Sousa, L.,
Tavares, J.M.R.S., 2013. A novel automatic algorithm for the segmentation

of the lumen of the carotid artery in ultrasound B-mode images. Expert Syst.

Appl. 40 (16), 6570–6579. doi: 10.1016/j.eswa.2013.06.003 . 
os Santos, E., Yoshizawa, M., Tanaka, A., Saijo, Y., Iwamoto, T., 2005. Detection of

luminal contour using fuzzy clustering and mathematical morphology in in-
travascular ultrasound images. In: IEEE 27th Annual Conference Engineering in

Medicine and Biology, pp. 3471–3474. doi: 10.1109/IEMBS.2005.1617226 . 
ofian, H., Ming, J.T.C., Noor, N.M., 2015. Detection of the lumen boundary in the

coronary artery disease. In: 2015 IEEE International WIE Conference on Electri-
cal and Computer Engineering (WIECON-ECE). IEEE, pp. 143–146. doi: 10.1109/

WIECON-ECE.2015.7443882 . 

u, S., Hu, Z., Lin, Q., Hau, W.K., Gao, Z., Zhang, H., 2017. An artificial neural net-
work method for lumen and media-adventitia border detection in IVUS. Com-

put. Med. Imaging Graph. 57, 29–39. doi: 10.1016/j.compmedimag.2016.11.003 . 
un, Z.-H., Wei, X.-H., Zhou, W.-H., 2012. A nyström-based subtractive clustering

method. In: 2012 International Conference on Wavelet Active Media Technol-
ogy and Information Processing (ICWAMTIP), pp. 74–77. doi: 10.1109/ICWAMTIP.

2012.6413443 . 

aki, A., Najafi, Z., Roodaki, A., Setarehdan, S.K., Zoroofi, R.A., Konig, A., Navab, N.,
2008. Automatic segmentation of calcified plaques and vessel borders in

IVUS images. Int. J. Comput. Assist. Radiol. Surg. 3 (3), 347–354. doi: 10.1007/
s11548- 008- 0235- 4 . 

ard, A., Jamshidi, K., Movahhedinia, N., 2012. An automated approach for seg-
mentation of intravascular ultrasound images based on parametric active con-

tour models. Australas. Phys. Eng. Sci. Med. 35 (2), 135–150. doi: 10.1007/

s13246- 012- 0131- 7 . 
ang, X.-F., Huang, D.-S., Xu, H., 2010. An efficient local Chan-Vese model for im-

age segmentation. Pattern Recognit. 43 (3), 603–618. http://dx.doi.org/10.1016/j.
patcog.20 09.08.0 02 . 

http://dx.doi.org/10.13039/501100002322
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0001
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0001
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0001
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0001
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0001
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0001
http://dx.doi.org/10.1016/j.compmedimag.2013.07.001
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0003
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0003
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0003
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0003
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0004
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0004
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0004
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0004
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0004
http://dx.doi.org/10.1007/s10554-015-0820-x
http://dx.doi.org/10.1109/TMI.2006.872142
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0007
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0007
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0007
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0007
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0007
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0007
http://dx.doi.org/10.1109/83.902291
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0009
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0009
http://dx.doi.org/10.1016/j.compmedimag.2013.09.004
http://dx.doi.org/10.1016/j.procs.2015.06.090
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0012
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0012
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0012
http://refhub.elsevier.com/S1361-8415(17)30092-0/sbref0012
http://dx.doi.org/10.1016/j.ultrasmedbio.2015.03.022
http://dx.doi.org/10.1371/journal.pone.0120399
http://dx.doi.org/10.1109/TIP.2012.2226047
http://dx.doi.org/10.1016/j.procs.2015.04.209
http://dx.doi.org/10.1016/j.compbiomed.2016.10.021
http://dx.doi.org/10.1007/BF00133570
http://dx.doi.org/10.1109/TITB.2012.2189408
http://dx.doi.org/10.1136/hrt.2007.116384
http://dx.doi.org/10.1007/s11548-015-1345-4
http://dx.doi.org/10.1080/1025584