ENDOCARDIAL SEGMENTATION IN CONTRAST

ECHOCARDIOGRAPHY VIDEO WITH DENSITY BASED

SPATIO-TEMPORAL CLUSTERING

Prashant Bansod, U. B. Desai

SPANN Lab., Electrical Engg. Department, Indian Institute of Technology, Bombay, India

Nitin Burkule

Cardiologist, Asian Heart Hospital and Research Center, Mumbai, India

Keywords:

Contrast echocardiography, Left ventricle, Segmentation, Spatio-temporal clustering.

Abstract:

We present a spatio-temporal clustering algorithm for detection of endocardial contours in short axis (SAX)

contrast echocardiographic image sequences. A semiautomatic method for segmentation of left ventricle in

SAX videos is proposed which uses this algorithm and at the same time requires minimal expert intervention.

Expert is required to specify a few candidate points belonging to the contour, only in the ﬁrst frame of the

sequence. The initial contour is approximated by ﬁtting an ellipse in the region deﬁned by the points speci-

ﬁed. This region is identiﬁed as the principal cluster corresponding to the left ventriclular cavity. Later the

density based clustering was applied for regularization on the inital contour. We have extended the DBSCAN

algorithm for identiﬁcation of the principal cluster corresponding to the left ventricle from the image. The al-

gorithm also incorporates the temporal information from the adjacent frames during the segmentation process.

The algorithm developed was applied to 10 data sets over full cardiac cycle and the results were validated by

comparing computer generated boundaries to those manually outlined by one expert. The maximum error in

the contours detected was ±2.9mm. The spatio-temporal clustering algorithm proposed in this paper offers an

efﬁcient semiautomatic segmentation of heart chambers in 2D contrast echocardiography sequences.

1 INTRODUCTION

Amongst the various medical imaging modalities, two

dimensional (2D) echocardiography is valuable for

patients with heart diseases. It is noninvasive, real

time, easy to use in clinical environment and of-

fers relatively low cost solution as compared to other

modalities (Bridal et. al, 2003). However, for eval-

uation of cardiac functional parameters, segmenta-

tion is to be carried out. Manual segmentation as

routinely carried by experts is time consuming and

tedious due to large image data in different stan-

dard echo views over a full cardiac cycle. Again

the manual method also suffers from inter-observer

and intra-observer variability in measurements (Maes

et. al, 1993). Many researchers have shown image

processing applications to enhance clinical utility of

echocardiography by automated and semiautomated

endocardial border delineation and for evaluation of

functional cardiac parameters (Noble and Boukerroui,

2006). In fact there is a continuous growing de-

mand for the automated segmentation and quantiﬁ-

cation to support professionals in diagnosis. In re-

cent years automated segmentation of heart cham-

bers and in particular the left ventricle has received

signiﬁcant attention in 2D and 3D echocardiograms.

However automatic edge deﬁnition and subsequent

segmentation in echocardiograhic images is difﬁcult

due to presence of speckle noise, poor contrast, inher-

ent dropouts, inter-cavity structures and variability of

data along with orientation and positioning of trans-

ducer (Setaredhan and Soragham, 1996).

In recent years numerous clinical studies have

shown the clinical utility of myocardial contrast

echocardiography (MCE) in quantiﬁcation of my-

ocardial perfusion, left ventricle (LV) volumes, LV

contours and cardiac functional parameters (Cohen

et.al., 1998). There have been few reports of research

attempts towards the semiautomatic and fully auto-

matic segmentation of left ventricle from 2D con-

trast enhanced echo images (Wolfer et. al, 1999).

A very rigorous work for the segmentation problem

204

Bansod P., B. Desai U. and Burkule N. (2008).

ENDOCARDIAL SEGMENTATION IN CONTRAST ECHOCARDIOGRAPHY VIDEO WITH DENSITY BASED SPATIO-TEMPORAL CLUSTERING.

In Proceedings of the First International Conference on Bio-inspired Systems and Signal Processing, pages 204-209

DOI: 10.5220/0001064102040209

 SciTePress

in low mechanical-index contrast echocardiography

is reported (Zwirn et. al, 2006). It has been shown

that the use of temporal continuity results in better

segmentation as it follows the approach of human ex-

pert in delineation (Mullet-Parada and Noble, 1998).

Typically the dropouts present in the image can be re-

covered by the use of boundary information from the

neighboring frames (Choy et. al, 1998). Researchers

have reported active contour approach (Morales et.

al, 2002), trained deformable models (Garcia et. al,

2003) and active shape model (Pickard et. al, 2004).

Many of the proposed methods have shown results

comparable to expert delineation for good quality im-

ages (Mishra et. al, 2003). However none of the

methods has a generalized applicability for fully auto-

matic or semiautomatic segmentation for the images

acquired in routine clinical environment.

Few researchers have extended the application of

well established data clustering approaches in the

ﬁeld of medical image segmentation (Celebi et. al,

2005). In this work we have extended the Density-

based Clustering (DBSCAN) approach by including

temporal data and applied for the segmentation of

contrast echo sequences. Our spatio-temporal cluster-

ing algorithm has shown good results in the segmen-

tation of endocardial borders in frames of a sequence

by accommodating temporal information. The user

intervention is minimal and is of the form of specify-

ing ﬁve or more candidate points for contour on the

ﬁrst frame of the sequence.

The paper is organized as follows: In section II

we discuss the density based clustering and its ex-

tension in spatio-temporal clustering technique. In

section III, we present the application of the algo-

rithm for segmentation of endocardial border after ﬁt-

ting the ellipse in the ﬁrst frame through the points

speciﬁed by the user and then to subsequent frames

in the sequence. The contours thus obtained are post

processed and smoothened to obtain ﬁnal endocardial

borders. In section IV we present the results of the

proposed algorithm and, ﬁnally conclusions drawn

and future work is discussed in Section VI.

2 CLUSTERING

Clustering is an important technique in data min-

ing for ﬁnding data distributions and patterns in the

underlying one or more dimensional data (Jain and

Dubes, 1988). It has been a active ﬁeld of research

since last two decades and many novel approaches

have been reported in the literature (Jain et. al,

1999). Clustering has number of upcoming appli-

cation ﬁelds, such as statistical data analysis, pat-

tern recognition, image processing, segmentation and

many others. It is the task of grouping similar objects

together with respect to a distance , connectivity, con-

tinuity, relative density in the space or other similarity

measure.

In formal mathematical deﬁnition cluster is de-

ﬁned as (Fung, 2001): Let X ∈ R

m×n

be a set of data

items representing a set of m points in x

in R

. The

goal is to partition X into K groups C

such that ev-

ery data that belongs to the same group are more alike

than data in different groups. Each of the K groups is

called a cluster. The result of the algorithm is an injec-

tive mapping X 7−→ C of data items X

to clusters C

The number K might be preassigned by the user or it

can be unknown determined by the algorithm . There

are many approaches to data clustering that vary in

their complexity and effectiveness. For our applica-

tion we have focussed our attention on a single cluster

(K = 1), pertaining to the heart chamber speciﬁcally,

the left ventricle in the contrast echcocardiographic

view. In our work we call it as principal cluster. The

assumption of deﬁning only one principal cluster is

valid because the spatial coordinates of the boundary

objects of this principal cluster reﬂect the endocardial

contour.

2.1 Dbscan for Principal Cluster

Density-based algorithms typically regard clusters as

dense regions of objects in the data space separated by

regions of low density. Thus the main objective lies

in ﬁnding regions of high and low densities (Bradley

and Fayyad, 1998). This approach is also capable of

ﬁnding arbitrarily shaped clusters in the data space.

Another advantage of these algorithms is that they

are independent of the prior knowledge of the number

of clusters. Hence these are very useful in situations

very clustering can be conﬁned to only in the region

of interest (Han and Kamber, 1998). In contrast en-

hanced short axis echo sequence, the chamber cavi-

ties are ﬁlled with micro-bubbles which contribute in

achieving their opacity. This results in a bright re-

gions corresponding to the blood ﬁlled areas in an

echo image (Fedele et. al, 1998). In the SAX images

of the left ventricle (LV), a single bright region in the

center of the acoustic window corresponds to the LV

cavity. We treat this central bright region as a single

cluster of interest. As stated earlier it is termed as the

principal cluster for this application. The two global

parameters of density based clustering algorithms are:

• E ps: Maximum radius of the neighborhood.

• MinPts: Minimum number of points in the Eps

neighborhood of a point.

ENDOCARDIAL SEGMENTATION IN CONTRAST ECHOCARDIOGRAPHY VIDEO WITH DENSITY BASED

SPATIO-TEMPORAL CLUSTERING

205

The ellipse ﬁtted through the expert speciﬁed points

in the ﬁrst frames is taken as the starting point for

the density based cluster algorithm. The maximum

radius of the neighborhood E ps is chosen as half of

the major axis of the ellipse. The parameter MinPts

was chosen to be 100 after a study of end systole im-

ages in 44 patients. The core point of the principal

cluster is chosen as the center point of the ellipse ﬁt-

ted through the points speciﬁed by the expert in the

ﬁrst frame. We have chosen four local parameters for

grouping the objects (pixels) in a cluster. These pa-

rameters include the features of the objects like pixel

intensity, gradient threshold, gradient angle and the

angular gradient with respect to the center point of the

region. For cavity boundary, only negative intensity

changes are identiﬁed along radial lines from center

point. Again the threshold for gradient (G

) was ob-

tained automatically from the histogram statistics and

the coordinates and intensities of the pixels speciﬁed

by the expert. The algorithm for our application is

framed as:

• The center of the ellipse is taken as core point.

• The maximum radius is assigned the value of semi

major axis in the ﬁrst frame.

• Gradient threshold is obtained by histogram of the

frame.

• Density reachable points around the core point

are identiﬁed.

• Above steps are repeated for all the frames in the

sequence.

• Border objects of the cluster are determined.

2.2 Spatio-temporal Dbscan

An image is a 2-dimensional(2D) array of pixels de-

ﬁned on a W × H rectangular lattice S = [ (x, y) :

1 ≤ x ≤ W, 1 ≤ y ≤ H ], and is indexed by the co-

ordinate (x, y). Each pixel in a given frame can be

represented by a feature vector. In a video stream, im-

age frames are continuous along the time axis. Thus

a video sequence can be expressed in spatio-temporal

domain. Temporal dimension can be incorporated in

many ways. One of the way is separating the frames

of the sequence with respect to discrete time and to

stack consecutively. We follow this approach in our

application. In a video sequence, the frame to frame

variation in shape and dimension of a given object

depends upon its deformity and forces acting on it.

Hence it is possible to recover a missing segment or

to correct any outlier in the contour from the adjacent

frames if the frame to frame variation is not signiﬁ-

cant. The outliers are detected with the radius of cur-

vature of the extracted contour and the corresponding

points in adjacent frames. We propose temporal con-

tinuity in the neighborhood of three frames:

1. For i

frame the j

border pixel will lie in the

bounds setup by (i − 1)

and (i + 1)

frame.

2. Presence of drop out pixels in a frame was taken

by temporal continuity from adjacent frames over

an interval of three fames.

We have used parameter λ as the correction factor

which governs the closeness of the corrected seg-

ment with the temporal frames. Correction is incor-

porated at 360 equidistant points on the contour i.e

j = [1, 2, .....360]. The clustering is recursively car-

ried out to regroup the cluster on the basis of modiﬁed

distance and density parameters. These parameters in

turn are function of λ during recursive calls of DB-

SCAN.

Figure 1: Flow chart for segmentation procedure.

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing

206

3 SEGMENTATION WITH

T-DBSCAN

In this work, we have applied the spatio-temporal

clustering algorithm to segment the endocardial bor-

der in the contrast enhanced echocardiography videos

of 10 patients. Figure 1. shows the ﬂowchart of the

segmentation procedure. After preprocessing of the

image frames, user speciﬁes candidate boundary pix-

els of the left ventricle in the ﬁrst frame of the ﬁrst

frame of the sequence by mouse. The ellipse is ﬁtted

through these points and its parameters are stored for

subsequent processing.

3.1 Image Processing

The echocardiographic videos used in this study were

contrast enhanced short axis apical images at various

levels of LV. These were obtained from different sub-

jects for two to four cardiac cycles. The videos were

acquired on GE Vingmed Ultrasound, VIVID7 in hos-

pital environment under expert guidance. The frames

in each video were 434 x 636 true color with 8 bit bit-

depth in DICOM format. Gray scale conversion with

256 levels was done. The video sequences for one

complete cardiac cycle were used for estimation of

LV border. Echo images contain speckle noise which

lead to incorrect gradient estimation. Hence speckle

reducing anisotropic diffusion (SRAD) ﬁltering was

used (Yongjian and Scott, 2002). They have sug-

gested edge sensitive diffusion for reducing speckles.

In the numerical implementation we used δt = 0.008

and threshold of 5. This reduced the speckles and at

the same time preserved the edge information for fur-

ther feature extraction.

3.2 Elliptical Boundary Approximation

The Initial boundary approximation is carried out in

the ﬁrst frame of the sequence by ﬁtting a ellipse

through the points speciﬁed by the expert. The best

ﬁt ellipse through the points speciﬁed is done using

Least Squares Criterion (Fitzgibbon et. al, 1999) A

minimum of ﬁve points are to be speciﬁed by the ex-

pert, which strongly belong to the endocardial bor-

der for that particular frame. This is the only user

intervention which is required in our scheme. The

standard impixel function of MATLAB is used which

gives the spatial coordinates of the selected points

along with their intensities. The intensities returned

by the function were used in the subsequent proce-

dure for the search.

The generalized CONIC equation of the Ellipse is

given by:

+ by

+ cx +dy + exy + f = 0 (1)

with a, b and c not all zero and b

< 4ac, where all

of the coefﬁcients are real. Again, more than one so-

lution, deﬁning a pair of points (x, y) on the ellipse,

exists. It can be expressed in matrix notation as;

AX = 0 (2)

where X and A are given by

X = [1 x y]

(3)

A =





f 0.5c 0.5d

0.5c a 0.5e

0.5d 0.5e b





(4)

The coordinates of the N chosen points (N >= 5) as

marked by the expert and the equations (2-4) are used

for the determination of the parameter matrix of the

conic representation. The orientation and tilt of the el-

lipse is sought by coefﬁcients in the equation (1) and

incorporated in the evaluation of ﬁnal ellipse parame-

ters using square completion method. Figure 2 shows

the ﬁrst frame of SAX apical sequence with the de-

tected ellipse and its center point.

Figure 2: Ellipse Fitted in the ﬁrst frame of the sequence.

3.3 Spatio-temporal Clustering

The DBSCAN clustering algorithms is recursively

called for incorporating corrections in the outliers

with the parameter λ. In our implementation we have

used λ = 0.5 distance units, which gave optimum re-

sults.

3.4 Post Processing and Smoothing

The contour thus obtained was smoothed out by lo-

cally weighted scatter plot smoothing using least

squares linear polynomial ﬁtting. A span of

10 percent was used to implement this standard

ENDOCARDIAL SEGMENTATION IN CONTRAST ECHOCARDIOGRAPHY VIDEO WITH DENSITY BASED

SPATIO-TEMPORAL CLUSTERING

207

MATLAB function. Further smoothing was carried

out by ﬁtting spline through the data points with min-

imizing the maximum square distance between the

data points.

Figure 3: Detected contour in frame No.1.

Figure 4: Detected contour in frame No.8.

Figure 5: Detected contour in frame No.15.

4 EXPERIMENTAL RESULTS

The proposed methods for ellipse ﬁtting, DBSCAN

and Spatio-temporal DBSCAN were implemented in

MATLAB 2006a on P-IV 2.1 GHz PC. Figures 3 to

7 show the result of application of the proposed algo-

rithm. The endocardial border estimation was done

on more than 10 video sequences of various standard

contrast echo views. The contour estimated by com-

puter in each frame of every sequence was compared

with that drawn by expert.

Figure 6: Detected contour in frame No.25.

Figure 7: Detected contour in frame No.32.

5 CONCLUSIONS

The proposed method for semi automatic estimation

of endocardial border of heart chambers in short axis

contrast echocardiographic sequences is based on el-

lipse ﬁtting and subsequent spatio-temporal recursive

density-based clustering. The results show the effec-

tiveness of the method and its utility in the recovery

of the dropouts during image acquisition. The method

requires user intervention only in the ﬁrst frame of

the sequence. The contour for each frame so obtained

may be utilized for the determination of the cardiac

parameters like, wall motion, area and for 3D visual-

ization. Further work is required before the method

can be employed in clinical environment for evalua-

tion of cardiac functional parameters. The issues in-

volved are the testing robustness, computational com-

plexity of the method along with its sensitivity to the

expert points. The algorithm requires ﬁne tuning of

parameter λ for determination of optimum number of

iterations. In future work, we also intend to test the

proposed method on large number of data sets for its

further validation for images acquired in routine clin-

ical environment.

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing

208

ACKNOWLEDGEMENTS

The authors would like to thank Echocardiology de-

partment of Asian Heart Hospital, Mumbai for ex-

tending its support for contrast echo data and sugges-

tions during analysis and interpretation.

REFERENCES

S.L.Bridal, S.L. Correas, Saied, A. and Laugier, P.(2003)

Milestones on the road to higher resolution, quanti-

tative, and functional ultrasonic imaging. In Proc.

IEEE,vol.91, no.10, pp.1543-1561.

Maes, L. Bijnens, B. Suetens, P. and Verf, F.V. Automatic

Contour Detection of the Left Ventricle in short Axis

views in 2D Echocardiograms. In Machine vision and

Applications, Springer-Verlag., 6:1-9.

Noble, J.A. and Boukerroui, D.(2006). Ultrasound Image

Segmentation: A Survey. In IEEE Trans. Med. Imag.,

vol. 25, pp. 987-1010.

Setaredhan, S. K. and Soragham, J. J.,(1996) Automatic

Cardiac LV Boundary Detection and Tracking Using

Hybrid Fuzzy Temporal and Fuzzy Multiscale Edge

Detection. In IEEE Trans. Med. Imag., vol. 46, pp.

1364-1378.

JL Cohen et.al.(1998) Improved left ventricular endocardial

border delineation and opaciﬁcation with optison, a

new echocardiographic contrast agent. In J Am Coll

Cardiol, 32, pp: 746:752.

Wolfer, J. Lee, SH. Sandelski, J. Summerscales, R. Soble, J.

and Roberge, J.(1999) Endocardial Border Detection

in Contrast Enhanced Echocardiographic Cineloops

using a Pulse Coupled Neural Networks. In IEEE

Conference on Computers in Cardiology, pp. 185-188.

Zwirn, G. Beeri, R. Gilon, D. and Akselrod, S.(2006).

Automatic Endocardial-Boundary Detection in Low

Mechanical-Index Contrast Echocardiography. In

IEEE Trans.on Biomedical Engg., Vol.53, No. 11,

pp:2310- 2322.

Mulet-Parada, M and Noble, J.A.(1998). 2D+T Acoustic

Boundary Detection in Echocardiography. In MIC-

CAI’98. LNCS 1496, pp. 806-813.

Choy, M.M. and Jin, J.S.(1998) Extracting endocardial Bor-

ders from sequential Echocardiographic Images. In

IEEE Engineering in Medicine and Biology, pp. 116-

121.

MA Morales, MA. Positano, V. Rodriguez, O. Passera, M.

Lombardi, M. and Rovai, D., (2002) Tracking of the

Left Ventricle in Contrast Enhanced Echocardiogra-

phy by Anisotropic Filtering and Active Contours Al-

gorithm. In IEEE Conference on Computers in Cardi-

ology, pp. 65-68.

Garcia,J. Rotger, D. Carreras, F. Leta, R. and Radeva, P.

(2003) Contrast Echography Segmentation and Track-

ing by Trained Deformable Models In IEEE Confer-

ence on Computers in Cardiology, pp. 173-176.

J.E. Pickard, J.E. R.L. Janiiczek, R.L. S.T. Acton, S.T. J.

Sklenar, J. J.A. Hossack, J.A. and S. Kaul, S.,(2004)

Segmentation of the Myocardium from Myocardial

Contrast Echocardiography. In IEEE Conference

on Signals, Systems and Computers, Volume 2, pp:

1616-1619.

Mishra, A. Dutta, P. Ghosh, M.K.(2003) A GA Based Ap-

proach for boundary Detection of Left Ventricle with

Echocardiographic Image Sequences, In Image and

vision Computing vol.21, pp. 967-976.

Celebi, M.E. Aslandogan, Y.A. and Bergstresser,

P.R.(2005) Mining biomedical Images with Density-

based Clustering. In Proceedings of IEEE Conference

on Information Technology: Coding and Comput-

ing(ITCC05).

Jain, A.K. Dubes, R.C. (1998). Algorithms for Clustering

Data. Prentice Hall, Upper Saddle River, New Jersy.

Fung, G. (2001) A Comprehensive Overview of Basic Clus-

tering Algorithms In Technical Report

Jain, A.K. Murty, M.N. Flynn, P.J.(1999). Data Clustering:

a Review. In ACM Surveys, 31(3):264-323.

Bradley, PS. Fayyad, U. and Reinna, C.(1998) Scaling

Clustering algorithms to large data bases.

Han,J. and Kamber, M. (1998) Data Mining: Concepts and

Techniques. Morgan Kaufmann, San Francisco, 2000.

Fedele F, Trambaiolo P, Magni G, De Castro S and Cacciotti

L, New modalities of regional and global left ventric-

ular function analysis;State of the art. In American

Journal of Cardiology, 81:49G-57G.

Yu, Y. and Acton, S.T., (2002) Speckle Reducing

Anisotropic Diffusion In IEEE Trans.Image Process-

ing.,vol 11,No.11, pp. 126-1270.

Fitzgibbon, A. Pilu, M. and Fisher, R.B. (1999) Direct Least

Square ﬁtting of Ellipses, In Pattern Analysis and Ma-

chine Intelligence, vol.21, no.5.

ENDOCARDIAL SEGMENTATION IN CONTRAST ECHOCARDIOGRAPHY VIDEO WITH DENSITY BASED

SPATIO-TEMPORAL CLUSTERING

209