A HYBRID BOUNDARY–REGION LEFT VENTRICLE

SEGMENTATION IN COMPUTED TOMOGRAPHY

Antonio Bravo, Jos´e Clemente

Grupo de Bioingenier´ıa, Decanato de Investigaci´on, Universidad Nacional Experimental del T´achira

San Crist´obal 5001, Venezuela

Miguel Vera, Jos´e Avila, Rub´en Medina

Grupo de Ingenier´ıa Biom´edica, Facultad de Ingenier´ıa, Universidad de Los Andes, M´erida 5101, Venezuela

Keywords:

Segmentation, Generalized Hough transform, Mathematical morphology, Unsupervised clustering, Cardiac

images, Left ventricle.

Abstract:

An automatic approach based on the generalized Hough transform (GHT) and unsupervised clustering tech-

nique to obtain the endocardial surface is proposed. The approach is applied to multi slice computerized

tomography (MSCT) images of the heart. The ﬁrst step is the initialization, where a GHT–based segmentation

algorithm is used to detect the edocardial contour in one MSCT slice. The centroid of this contour is used

as a seed point for initializing a clustering algorithm. A two stage segmentation algorithm is used for seg-

menting the three–dimensional MSCT database. First, the complete database is ﬁltered using mathematical

morphology operators in order to improve the left ventricle cavity information in these images. The second

stage is based on a region growing method. A seed point located inside the cardiac cavity is used as input for

the clustering algorithm. This seed point is propagated along the image sequence to obtain the left ventricle

surfaces for all instants of the cardiac cycle. The method is validated by comparing the estimated surfaces

with respect to left ventricle shapes drawn by a cardiologist. The average error obtained was 1.52 mm.

1 INTRODUCTION

In medical image processing, segmentation is an im-

portant tool to analyze anatomic tissue features types,

and spatial distribution of functional regions (active

and pathological) (Bankman, 2000). Additionally,

this technique is useful to extract information for di-

agnosis or quantiﬁcation (Angelini et al., 2001), visu-

alization (Nelson and Elvins, 1993), and ﬁnally com-

pression, storage and transmission (Field, 1996; DI-

COM, 1999).

Image segmentation techniques are based on the

organization and grouping of a set of shapes, being

the proximity, similarity and continuity the main or-

ganization characteristics. The segmentation process

partitions an image into homogeneous regions, also

called classes or subsets, considering one or more

similar characteristics (Fu and Mui, 1981; Duda et al.,

2000). Most of segmentation methods for medical im-

ages, are based on delineation of a curve that deﬁnes

the anatomical structures, which allows to discrimi-

nate the structure of interest from other structures that

appear in the image (Kervrann and Heitz, 1999). An-

other kind of techniques are based on application of

classiﬁcation methods, where the image is processed

until represented as a non–overlapped set of two (2)

regions (subject of interest and background) (Mitchell

et al., 2001).

Images studies in cardiology are used to obtain

both qualitative and quantitative information of the

heart and vessels morphology and function. Several

clinical parameters can be extracted from dynamics

images of cardiovascular structures with the objective

of reproducing the heart space–time behavior (Rabit,

2000). Assessment of cardiovascular function is im-

portant since Cardio–Vascular Disease (CVD) is con-

sidered the most important cause of mortality. Ap-

proximately 17 million people die each year, repre-

senting one third of the deaths in the world (WHO,

2002a). About 85% of overall mortality of middle-

and low-income countries is due to CVD and it is es-

timated that CVD will be the leading cause of death

107

Bravo A., Clemente J., Vera M., Avila J. and Medina R. (2010).

A HYBRID BOUNDARY–REGION LEFT VENTRICLE SEGMENTATION IN COMPUTED TOMOGRAPHY.

In Proceedings of the International Conference on Computer Vision Theory and Applications, pages 107-114

DOI: 10.5220/0002849301070114

 SciTePress

in developed countries in two years (WHO, 2002b).

Multi Slice Computerized Tomography is prob-

ably the term most commonly used to describe the

latest developments in Spiral computed tomography

(CT) which is based on simultaneous acquisition of

more than one tomography plane, and it is closely re-

lated to acquisition systems with multiple detectors

(Fuchs et al., 2000).

Left ventricle (LV) is considered the main cav-

ity of the heart. In this sense, the assessment of the

LV function allows to assess the cardiovascular func-

tion. The LV function analysis requires the accurate

description of ventricular shape.

1.1 Related Work

There are several research studies on cardiac segmen-

tation especially focused on left ventricle segmenta-

tion. Lynch et al. (Lynch et al., 2008) developed a LV

segmentation method from magnetic resonance imag-

ing. The method uses prior knowledge about ventric-

ular motion to guide a parametric model of the car-

diac cavity. The model deformation was initially con-

trolled by a level–set formulation. The state of the

model attained by the level–set evolution was reﬁned

using the expectation maximization (EM) algorithm.

The objective was to ﬁt the model to MRI data. The

method was tested using a set of six clinical databases.

The correlation coefﬁcientobtained by a linear regres-

sion analysis of segmentation results with respect to

manual segmentation was 0.76.

Fleureau et al. (Fleureau et al., 2006), re-

ported a semi–automatic and multi–structure three–

dimensional (3–D) segmentation method. The

method was applied to extract the heart cavities from

MSCT sequences. The method associated basic

agents to the objects of interest. Each agent learnt the

image region characteristics through a Support Vector

Machine formulation. The cardiac cavities were ob-

tained by maximizing the region associated to each

basic agent. The approach allowed to discriminate

structures such as, left ventricle, left atrium, right ven-

tricle and right atrium. However, the clinical valida-

tion was not performed at the time of publication.

Chen et al.(Chen et al., 2004) developed a hybrid

model for LV CT segmentation. The model couples

a segmenter, based on a Gibbs prior models and de-

formable models with a marching cubes procedure. A

external force based on a scalar gradient was consid-

ered in order to achieve convergence. The approach

was tested using 8 CT studies. Results obtained re-

veals the good behavior of the method.

Recently, a semi–automatic segmentation method

based on a 3–D active shape model has been pro-

posed by Van Assen et al. (Assen et al., 2008). The

method has the advantage of being independent with

respect to the imaging modality. The LV shape was

obtained for the whole cardiac cycle in 3D MRI and

CT sequences. A point–to–point distance was one of

the metrics used to evaluate the method performance.

The average value of the distances obtained for the

CT sequences was 1.85 mm.

1.2 Purpose

In this research, an automatic image segmentation ap-

proach useful to extract the left ventricle cavity from

multi–slice computerized tomography 4–D (3–D +

time) images is proposed. Morphological ﬁlters are

used to improve the LV information in the images,

and thus to facilitate the process of segmentation. The

approach uses a region growing algorithm for the seg-

mentation. A pixel called initial seed is located in the

cavity of interest using the generalized Hough trans-

form. This seed is established in the ﬁrst volume of

the MSCT dataset, and then compared with certain

neighborhood pixels characteristics such as intensity

and topological relationship according to a region–

growing algorithm. The segmentation algorithm al-

lows to obtain a binary image with the LV information

and the background. From this initial binary image a

new seed is automatically generated for segmenting

the next MSCT volume. The ﬁrst seed is then propa-

gated along the 3–D image sequence to obtain the LV

surfaces for all instants of the cardiac cycle.

2 METHOD

2.1 Dataset

A human MSCT database is used. The dataset con-

tains 20 volumes to describe the heart anatomical

information for a cardiac cycle. Each volume has

262 CT slices where the spacing between pixels is

0.488281 mm and the slice thickness is 0.625 mm.

The data acquisition was triggered by the R wave of

the ECG signal. Each image is quantized with 12 bits

per pixel and the size is 512× 512 pixels.

2.2 Hough Transform Seed Localization

In this work, the Generalized Hough Transform

(GHT) is applied to obtain the left ventricle border

in one MSCT slice. From this contour, the seed point

required to initialize an unsupervised clustering algo-

rithm is computed.

VISAPP 2010 - International Conference on Computer Vision Theory and Applications

108

The GHT proposed by Ballard (Ballard, 1981)

has been used to detect objects, with speciﬁc shapes,

from images. The proposed algorithm consists of two

stages: 1) training and 2) detection. During the train-

ing stage, the objective is to describe a pattern of the

shape to detect. This pattern is parameterized using

the gradient direction (θ) at each pattern point, the dis-

tance (r) between the reference point (x

, y

) and each

pattern point, and the angle (α) with respect to x–axis

of the line formed by the reference point and each pat-

tern point. The parameters describing the shape are

stored in an array known as the R–Table.

The second stage is implemented to detect a sim-

ilar shape in an image not used during the training

step. The parameter θ is calculated for each point in a

pre–processed image to segment. The θ value is used

as an entry to the R–table. The corresponding r and

α values found in the table are used to compute the

candidates to reference points (x

, y

) according to 1.

= x+ rcos(α)

= y+ rsin(α)

, (1)

where (x, y) are object boundaries points. An accumu-

lator is used to count the occurrences of each refer-

ence point calculated. The reference point with more

occurrence is selected.

The overviews of the training and detection stages

for the LV segmentation are shown on the ﬂowcharts

in Figures 1 and 2. In the training stage, the heart

structures boundaries are initially estimated using the

Canny edge detection algorithm (Canny, 1986). The

LV contour is manually labeled in the Canny’s con-

tours map. The centroid of the LV contour is used

as the reference point (x

, y

). The R–table is con-

structed when the values of the (θ, r, α) are computed.

During the detection step (Figure 2), the applica-

tion of the Canny edge detection algorithm is also re-

quired. A thresholding technique is applied in order

to discriminate the size of the regions delimitated by

the borders labeled in the Canny image. The thresh-

olds used correspond to areas measured in pixels. The

regions obtained after the thresholding are considered

as candidates for the left ventricle shape. The border

points of the LV candidates are used to calculate the

gradient direction (θ). For the θ values found in the

R–table, the corresponding r and α values are used

to calculate the reference points using 1. The ﬁnal

contour correspond to the candidate whose reference

point has the best match with respect to the reference

point of the pattern.

Figure 3 illustrates the training process applied to

one MSCT slice while Figure 4 shows the results of

the segmentation for other MSCT slice.

Figure 1: GHT training stage.

Figure 2: GHT detection stage.

2.3 Unsupervised Segmentation

2.3.1 Pre-processing

Mathematical morphological operators are used for

implementing the ﬁlters aimed at enhancing the LV

information. These morphological operators are

based on non–linear operations between the original

image (I) and a set of additional points known as

structuring element (B) (Serra, 1982). The applied

ﬁlters are based on the top–hat transform. This trans-

A HYBRID BOUNDARY-REGION LEFT VENTRICLE SEGMENTATION IN COMPUTED TOMOGRAPHY

109

(a) (b)

(c)

Figure 3: Results of GHT training stage. (a) Original image.

(b) Canny image. (c) Pattern obtained.

(a) (b)

Figure 4: Results of GHT segmentation stage. (a) Original

image. (b) LV segmented.

form is a composite operation deﬁned by the set dif-

ference between the image processed by a closing op-

erator and the original image. The closing (•) oper-

ator is also a composite operation that combines the

basic operations of erosion (⊖) and dilation (⊕). The

top–hat transform is expressed according to (2).

I• B− I = (I⊕ B) ⊖ B− I . (2)

A modiﬁcation of the basic top–hat transform def-

inition is introduced. A Gaussian ﬁlter is applied to

the original image. The discrete Gaussian operator

with standard deviation σ is used as a ﬁlter mask (3).

K(i, j) =



2πσ



−

+ j

2σ

;0 ≤ i, j ≤ n , (3)

where n denotes the mask size and σ is set as the stan-

dard deviation of the original image. The processed

(a) (b)

Figure 5: Pre–processing stage. (a) Original image. (b)

Gaussian smoothed image. (c) Eroded image. (d) Dilated

image.

(a) (b)

Figure 6: The top–hat transform. (a) Original image. (b)

Processed image.

image (I

Gauss

) is a blurred version of the input.

The Gaussian smoothed image is used to calculate

the morphological closing. The structuring element

selected is an ellipsoid that varies in size depending

on the operator. The major axis of the structuring el-

ement used for the erosion is 3 and for the dilation is

5. Figure 5 shows part of the pre–processing stage.

Finally, the top–hat transform is calculated us-

ing (4), the result is an image with enhanced con-

tours. The original image and the processed image

are shown in Figure 6.

BTH

= (I

Gauss

⊕ B

) ⊖ B

− I

Gauss

. (4)

The intensity values of the top–hat image (I

BTH

)

and the Gaussian image (I

Gauss

) are used to create a

feature vector. This feature vector is used to construct

VISAPP 2010 - International Conference on Computer Vision Theory and Applications

110

(a) (b)

Figure 7: Final enhancement process. (a) Original image.

(b) Similarity image.

a similarity matrix based on a similarity criteria (Har-

alick and Shapiro, 1992). The criteria measures the

difference between the gray–level values of pixels in

BTH

and the smoothed image, I

Gauss

. According to

this criteria, pixels p

(x, y) (in I

BTH

) and p

(x, y) (in

Gauss

) have features vectors denoted as: pv

= [i

, a]

and pv

= [i

, b], where i

and i

denote the intensity

associated with the corresponding pixel of I

BTH

and,

a and b are the intensity of the smoothed image. The

pixel values in the similarity matrix are obtained by

using the following expression:

(x, y) = (i

− i

)

+ (i

− b)

+ (i

− a)

. (5)

Figure 7 shows the image enhanced using the sim-

ilarity criteria, where the information related to the

LV cavity is enhanced with respect to other anatomi-

cal structures that are present in the MSCT slice.

The process described previously, is applied to all

volumes of the human MSCT database. Since this

process requires large computing resources, multipro-

gramming based on threads is used to speed up the

enhancement. The performance and a speed test is

applied using 1Gb memory–Pentium IV machine. An

optimal reduction of 48% in processing time, without

saturating the equipment operation, is attained using

4 threads.

2.3.2 Segmentation Stage

A region growing technique is used to segment the

LV. The unsupervised clustering algorithm requires a

seed point located inside the region of interest to iden-

tify the cardiac cavity. The seed is established in one

slice at only one time instant of the MSCT 4–D image

sequence. The procedure used to establish the seed

point is based on the GHT (see section 2.2).

Seed Determination. A seed is used for starting the

region–growing segmentation process. This seed cor-

responds to the reference point of the shape obtained

using the GHT procedure. The seed point is located

in the bi–dimensional (2–D) image I

, where t rep-

resents the time instant of the MSCT dataset, and k is

the slice in the corresponding volumet. When the im-

age I

has been segmented (process described in sec-

tion 2.3.2) a binary image

is obtained. In this im-

age, pixels in white represent the segmented region.

From

the seed points necessary to segment the

entire volume t are estimated. The center of mass of

the segmented region in the image

is calculated

and denoted as r(x, y). The pixel r(x, y) is the new

seed to segment the images I

k+1

and I

k−1

. The seed

generation process is applied upward by diminishing

the value of k until reaching the LV base. In the same

way, the process is applied downward by increasing k

until reaching the apex.

The process for propagating the seed from one

volume to the next, is also based on the calculation

of the gravity center as previously explained. How-

ever, in this case the calculation is performed for the

3-D binary object.

The procedure based on calculation of the center

of mass, results in a point located very near of the LV

anatomical axis. In consequence, the seed is always

located inside the target region (inside the LV).

Region Growing Algorithm. The algorithm is devel-

oped using dynamic linked–lists. The algorithm in-

puts are the enhanced image and a binary image with

all pixels set to zero (0). The lists are implemented as

a First In First Out (FIFO) queue. The list is used to

store temporarily the pixels that fulﬁll the clustering

criterion. The objective is to develop an iterative algo-

rithm highly efﬁcient with respect to memory require-

ments aiming at avoiding memory overﬂows. Each

node in the list contains the pixel information: loca-

tion and gray level intensity. The ﬁrst node inserted

in the list is the seed pixel.

After introducing the seed pixel in the FIFO list,

the algorithm performsthe following steps: 1) the ﬁrst

node of the list is dequeue, 2) the gray level informa-

tion associated to the analyzed node is compared with

pixel intensities in an 8 pixels neighborhood to deter-

mine if these neighbor pixels belong or not to the tar-

get region. Pixels of the neighborhood that fulﬁll the

clustering criterion are inserted at the end of the list

and their values in the binary image are modiﬁed to

one (1). The pixels that do not fulﬁll with the condi-

tion are rejected, and 3) the algorithm continues with

this process while there are nodes in the list. The al-

gorithm output is the binary image where pixels set

to one represent the segmented region. The unifor-

mity criterion for grouping pixels is as follows: pixels

are grouped if the difference between the pixel value

in the neighborhood and the intensity of the pixel ex-

tracted of the list is below

standard deviation of sim-

A HYBRID BOUNDARY-REGION LEFT VENTRICLE SEGMENTATION IN COMPUTED TOMOGRAPHY

111

ilarity image.

In this algorithm multiprogramming is also used,

considering two threads. The ﬁrst thread segments the

slices from I

, I

k+1

,......, to I

apex

, and the second

thread segments the slices from the I

k−1

, I

k−2

,......,

to I

base

. The LV base and apex are detected auto-

matically by our algorithm. A comparison between

current segmented area and previous segmented area

is performed. If the areas are different in more than

80%, the current segmented area does not belong to

the interest region, and then the segmentation process

is stopped. Figure 8 shows the results of the segmen-

tation for two consecutive tomographic slices.

(a) (b)

(e) (f)

Figure 8: (a) Seed in slice I

. (b) Seed in slice I

k+1

. (c) and

(d) LV contours. (e) and (f) LV areas.

After the segmentation process the reconstruction

of the LV surface is performed using the Visualiza-

tion Toolkit (VTK) (Schroeder et al., 2001). The en-

docardial LV wall is reconstructed using the marching

cubes algorithm (Salomon, 1999) (Figure 9).

Figure 9: Result of the segmentation process.

2.3.3 Validation

The proposed method is validated by calculating the

difference between the estimated LV shape with re-

spect to a ground truth shape, traced by an expert.

Two different methodologies for evaluating the per-

formance of the LV segmentation method are consid-

ered. First, the approach proposed by Suzuki et al.

(Suzuki et al., 2004) is incorporated. Suzuki’s quanti-

tative evaluation methodology is based on calculating

two metrics that represent the contour error (E

) and

the area error (E

). See (Suzuki et al., 2004, p. 335)

to show the contour and area errors expressions.

The validation methodology proposed by Chalana

and Kim (Chalana and Kim, 1997) is also used. A

metric based on a mean absolute distance (MAD) of

the distance to the closest point (DCP) is used for as-

sessing the position error (E

) between contours au-

tomatically extracted with respect to contour traced

by an experts. The metric expression can be found in

(Chalana and Kim, 1997, p. 643).

3 RESULTS

The proposed method is implanted using a multi–

platform object–oriented methodology along with

C++ multiprogramming and using dynamic memory

handling. Standard libraries such as the Visualization

Toolkit (VTK) are used. VTK consists of a complete

set of algorithms for 3–D images visualization. The

Fast Light Toolkit (FLTK) open source libraries are

also used to develop the graphic interface. The code is

executed in Microsoft Windows and Linux platforms.

In the application, threads are used to speed up the

process and then to optimize the response times.

The segmentation algorithms are tested with more

VISAPP 2010 - International Conference on Computer Vision Theory and Applications

112

Figure 10: Cardiac structures at 10%, 30%, 50%, 70% and 90% of the cardiac cycle.

Figure 11: Left ventricle surfaces in the new frame of reference.

Table 1: Errors obtained for a total of 262 images processed.

Error Types Result

3.38 % ± 3.09 %

6.23 % ± 3.77 %

1.52 mm ± 0.18 mm

than ten thousand 2–D images, obtaining good results

for all images with very satisfactory processing times.

For instance, for a database including 20 volumes and

262 images per volume, the ﬁltering and segmentation

take approximately 15 minutes. Figure 9 shows the

segmented LV overlaid with three orthogonal slices

extracted from the original database (axial, coronal

and saggital), where it can be shown that for these

planes the structure matches with the LV contours.

Table 1 shows the comparison of extracted sur-

face with respect to the surface traced by the car-

diologist. The errors estimated are expressed as

mean± standard deviation. The position error varies

between 0.87 mm and 1.72 mm. The average po-

sition error obtained using our segmentation method

was 1.52 mm which is smaller than the average error

(1.85 mm) reported by Van Assen et al. (Assen et al.,

2008). Using the proposed segmentation method the

average contour error and the average area error are

6.23% and 3.38%, respectively.

Additionally, the application allows to establish

a particular frame of reference not dependent on the

original position of the heart. This frame is based on

an axis (z–axis) joining the apex to the joint between

the mitral and aortic valves as determined in the en-

docardial LV wall, and a plane perpendicular to this

axis that is the x–y plane. The frame of reference cor-

responds with the image acquisition geometry for the

MRI images. Figure 10 shows the segmentation re-

sults while Figure 11 shows the left ventricle in the

new frame of reference. The LV structures for in-

stants located at 10%, 30%, 50%, 70% and 90% of

the cardiac cycle are shown.

4 CONCLUSIONS

An automatic method for LV image segmentation

from 4–D MSCT datasets was proposed. The soft-

ware system is a platform independent tool devel-

oped using C++ and open–source libraries. The

pre–processing stage was based on gray level math-

ematical morphology ﬁlters aimed at performing the

smoothing and enhancement of image contours. A

region growing algorithm was controlled by a seed

point located in one volume using generalized Hough

transform, which was propagated to the rest of vol-

umes in order to segment the entire MSCT database.

A valuable contribution was the utilization of

threads since they improve the processing time for the

whole process. The segmentation method evaluation

was performed by comparing the estimated contours

with respect to contourstraced by a cardiologists. The

comparison was performed based on the methodolo-

gies proposed in (Chalana and Kim, 1997; Suzuki

et al., 2004) which are also used in (Oost et al., 2006)

and (Bravo and Medina, 2008). The validation stage

shows that errors are small. The method allowed to

detect LV important features as the papillary muscles.

As a future research we propose to use the method

for automatic segmentation of right ventricle, left and

right atrium in multi–slice computerized tomography

(MSCT) images. Additionally, we plan to develop the

approach in a 3–D domain in order to take into ac-

count three–dimensional topological features of the

left ventricle and for speeding up the segmentation

procedure. We also have considered to use the pro-

posed method for heart structures segmentation in

other cardiac imaging modalities.

A HYBRID BOUNDARY-REGION LEFT VENTRICLE SEGMENTATION IN COMPUTED TOMOGRAPHY

113

ACKNOWLEDGEMENTS

The authors would like to thank the Investigation

Dean’s Ofﬁce of Universidad Nacional Experimental

del T´achira, LOCTI grant PR0100401, and CDCHT

from Universidad de Los Andes (projects I-1075-07-

02B and NUTA C–24–07–02–C) for their support to

this research. Authors would also like to thank H. Le

Breton and D. Boulmier from the Centre CardioPneu-

mologique and M. Garreau from Laboratoire Traite-

ment signal et de l’image (LTSI) in Rennes, France

for providing the human MSCT database.

REFERENCES

Angelini, E., Laine, A., Takuma, S., Holmes, J., and

Homma, S. (2001). LV volume quantiﬁcation via spa-

tiotemporal analysis of real–time 3–D echocardiogra-

phy. IEEE Trans. Med. Imag., 20(2):457–469.

Assen, H. V., Danilouchkine, M., Dirksen, M., Reiber, J.,

and Lelieveldt, B. (2008). A 3D active shape model

driven by fuzzy inference: Application to cardiac CT

and MR. IEEE Trans. Inform. Technol. Biomed.,

12(5):595–605.

Ballard, D. (1981). Generalizing the hough transform to

detect arbitrary shapes. Pattern Recog., 13(2):111–

122.

Bankman, I. (2000). Handbook of Medical Imaging: Pro-

cessing and Analisys. Academic Press, San Diego.

Bravo, A. and Medina, R. (2008). An unsupervised clus-

tering framework for automatic segmentation of left

ventricle cavity in human heart angiograms. Comput

Med Imaging Graph., 32(5):396–408.

Canny, J. (1986). A computational approach to edge de-

tection. IEEE Trans. Pattern Anal. Machine Intell.,

PAMI–8:679–698.

Chalana, V. and Kim, Y. (1997). A methodology for eval-

uation of boundary detection algorithms on medical

images. IEEE Trans. Med. Imag., 16(5):642–652.

Chen, T., Metaxas, D., and Axel, L. (2004). 3–D cardiac

anatomy reconstruction using high resolution CT data.

In Proc. MICCAI (1), pages 411–418.

DICOM (1999). Digital imaging and communication in

medicine DICOM. NEMA Standards Publication.

Duda, R., Hart, P., and Stork, D. (2000). Pattern Classiﬁca-

tion. Wiley–Interscience, New York.

Field, M. J. (1996). Telemedicine: A Guide to Assess-

ing Telecommunications in Health Care. Institute of

Medicine, National Academy Press, Washington.

Fleureau, J., Garreau, M., Hern´andez, A., Simon, A., and

Boulmier, D. (2006). Multi-object and N-D segmenta-

tion of cardiac MSCT data using SVM classiﬁers and

a connectivity algorithm. In Computers in Cardiology,

pages 817–820.

Fu, K. S. and Mui, J. K. (1981). A survey on image seg-

mentation. Pattern Recog., 13(1):3–16.

Fuchs, T., Kachelriess, M., and Kalender, W. (2000). Sys-

tems performance multislice spiral computed tomog-

raphy. IEEE Eng. Med. Biol. Mag., 19(5):63–70.

Haralick, R. A. and Shapiro, L. (1992). Computer and

Robot Vision, volume I. Addison–Wesley, USA.

Kervrann, C. and Heitz, F. (1999). Statistical deformable

model–based segmentation of image motion. IEEE

Trans. Image Processing, 8(4):583–588.

Lynch, M., Ghita, O., and Whelan, P. (2008). Segmentation

of the left ventricle of the heart in 3-D+t MRI data

using an optimized nonrigid temporal model. IEEE

Trans. Med. Imag., 27(2):195–203.

Mitchell, S., Lelieveldt, B., van der Geest, R., Bosch, H.,

Reiber, J., and Sonka, M. (2001). Multistage hybrid

active appearance model matching: Segmentation of

left and right ventricles in cardiac MR images. IEEE

Trans. Med. Imag., 20(5):415–423.

Nelson, T. R. and Elvins, T. T. (1993). Visualization of

3D ultrasound data. IEEE Comput. Graph. Appl.,

13(6):50–57.

Oost, E., Koning, G., Sonka, M., Oemrawsingh, P. V.,

Reiber, J. H. C., and Lelieveldt, B. P. F. (2006). Auto-

mated contour detection in X–ray left ventricular an-

giograms using multiview active appearance models

and dynamic programming. IEEE Trans. Med. Imag.,

25(9):1158–1171.

Rabit, O. (2000). Quantitative analysis of cardiac function.

In Bankman, I. N., editor, Handbook of Medical Imag-

ing: Processing and Analysis, pages 359–374. Aca-

demic Press, San Diego.

Salomon, D. (1999). Computer Graphics and Geometric

Modeling. Springer Verlag, New York.

Schroeder, W., Martin, K., and Lorensen, B. (2001). The

Visualization Toolkit, An Object-Oriented Approach

to 3D Graphics. Prentice Hall, New York.

Serra, J. (1982). Image Analysis and Mathematical Mor-

phology. A Press, London.

Suzuki, K., Horiba, I., Sugie, N., and Nanki, M. (2004). Ex-

traction of left ventricular contours from left ventricu-

lograms by means of a neural edge detector. IEEE

Trans. Med. Imag., 23(3):330–339.

WHO (2002a). Integrated management of cardiovascular

risk. The World Health Report 2002 Geneva, World

Health Organization.

WHO (2002b). Reducing risk and promoting healthy life.

The World Health Report 2002 Geneva, World Health

Organization.

VISAPP 2010 - International Conference on Computer Vision Theory and Applications

114