AUTOMATIC HEART LOCALIZATION

IN ULTRASOUND FETAL IMAGES

Mozart Lemos de Siqueira and Philippe Olivier Alexandre Navaux

Institute of Informatics, Federal University of Rio Grande do Sul

PO Box: 15064, ZIP: 91501-970, Porto Alegre-RS, Brazil

Keywords:

Medical Imaging, Fetal Cardiology, Ultrasound, Bhattacharyya Coefﬁcient, Texture Feature.

Abstract:

This paper presents the research developed in order to detect the cardiac structure in echocardiography gray

images from the fetal heart. It is based on patten recognition and use a density probability function with the

scales of gray. The function is also used for search of the similar cardiac structure, where it is applied on the

whole image, and then compared with the pattern of structure in which one interested. In order to obtain the

similarity that deﬁnes the choice of the structure of interest we use the Bhattacharyya coefﬁcient.

The method uses texture features to isolate the region of interest inside the ultrasound image to improve the

results and performance. A prototype was developed to evaluate the proposed method. The results of the

experiments are also presented in this paper.

1 INTRODUCTION

Scientiﬁc Research in medical imaging area grows

constantly and its results generate many beneﬁts to

people’s health. Usually, these researches cover many

aspects of image processing and medicine such as dis-

ease predicting and more accurate diagnostics. In this

context, this article presents a computational tech-

nique for automatic localization of cardiac structure

in images. The images used is fetal echocardiography

and, more speciﬁc, fetal’s ultrasound images (Duncan

and Ayeche, 2000; Sheehan, 2000). Such images are

important to the prenatal phase, because an early di-

agnosis of congenital cardiopathy can help the med-

ical treatment. Therefore, this work may be used to

help automatic analysis, mainly when the physician

involved is not a heart specialist.

Although ultrasound images provide a lot of in-

formation about cardiac structures, the resulting im-

ages are contaminated by speckle noise, which cor-

rodes the borders of the cardiac structures (Kang and

Hong, 2002; Zong et al., 1998; Crimmins, 1985; Bur-

ckhardt, 1978). This characteristic turns difﬁcult the

automatic image processing, and specially the pattern

recognition. Besides this kind of noise, other factors

inﬂuence the outcome of fetal ultrasound image. For

instance, the transducer

and the fetus positioning, the

rotation and the scale variations in images of different

patients and the composition of the tissue separating

the fetus heart are issues that must be taken into ac-

count when dealing with heart images (Mattos, 1999).

The method presented in this paper makes use of a

density probability function to obtain the mapping of

the region of interest and generate a searching mold to

be used in other target images. This mold is based on

scales of gray image in the region where the cardiac

structure is positioned. Besides, the mold calculation

considers the distance of the pixels of the center.

The search for the region of interest begins with

the mapping of the candidate regions in the target

image. Each candidate region is compared to the

searching mold using the Bhattacharyya coefﬁcient

(Djouadi et al., 1990). This calculated coefﬁcient pro-

vides a index that deﬁnes the degree of similarity be-

tween the candidate region and the searching mold.

A prototype for the cardiac structure localization

was developed in order to evaluate the automatic

searching pattern. The search of the candidate re-

gions was implemented using the mold and local-

The electronic device used to capture ultrasound im-

ages

107

Lemos de Siqueira M. and Olivier Alexandre Navaux P. (2007).

AUTOMATIC HEART LOCALIZATION IN ULTRASOUND FETAL IMAGES.

In Proceedings of the Second International Conference on Computer Vision Theory and Applications, pages 107-112

DOI: 10.5220/0002059101070112

 SciTePress

ization in different images that was analyzed with a

moving window shifted in 10 pixels along the image

columns and lines. The moving window size was de-

ﬁned based on the mold size. Although this search

covers the whole image and can provide an accurate

result, the time to process on each image is usually

greater than one minute considering a resolution of

640x480. Because this, it was necessary to ﬁnd new

manners to increase the search performance.

Allowing the search performance in considera-

tion, we developed a module to images preprocessing

and to isolate the region of interest (ROI). The entropy

texture feature was used for selection of region of in-

terest. The Figure 1 (a) presents an original image and

a preprocessed image (b) with entropy texture feature

to select the ROI. This approach improves the search

performance because it does not need search regions

out of ultrasound, i.e. the black frame on images. The

use of texture features on this ultrasound images is

justiﬁed by the speciﬁc characteristics of them, there-

fore many works about ultrasound processing using

this features (Valdes-Cristerna et al., 2004; Brusseau

et al., 2004; Hope et al., 2005).

The paper is organized in ﬁve sections. After this

introduction, we present a brief section with related

research. The Section 3 present the description of the

proposed model. The Section 4 show the obtained re-

sults. Finally, the paper presents the conclusion sec-

tion.

2 RELATED RESEARCH

The pattern recognition in images is an area that has

offered many results. The goal of pattern recogni-

tion is to automatically spot speciﬁc objects inside

images, without the intervention from the user. There

are many application possibilities for pattern recogni-

tion in medicine, since several medical routines gen-

erate images. In some routines, physicians look for

image patterns in nodules, intern structures, the be-

havior of the heart (i.e. systole and diastole), for ex-

ample. There are works that already address this rou-

tines (Lee et al., 2001; Brown et al., 2001; Bruijne

et al., 2003; Salvadorls et al., 2003).

Jacob et al. (Jacob et al., 2002) and Sugioka et

al (Sugioka et al., 2003) developed research using

patterns to detect cardiac structures using active con-

tours (snakes) in echocardiographic images. Comani-

ciu (Comaniciu et al., 2004) proposed a methodology

to tracking cardiac edges in echocardiographic im-

ages using several information extract of the images.

One of the principal ways of searching objects in

images is through patterns. In most of the cases, the

(a)

(b)

Figure 1: Fetal echocardiography obtained with ultrasound

device. (a) Original image; (b) The regions of interest iso-

lated with entropy.

objects of interest must be known, and their charac-

teristics are searching in the image. When the object

of interest is not previously known, the complexity

of the search increases. The search algorithm perfor-

mance is important to the processing time do not to

be a bottleneck to the system.

3 PROPOSED MODEL

To developing the automatic localization of the car-

diac structure was studied the approach proposed

by Comaniciu (Comaniciu et al., 2003) for tracking

down dynamic objects. We proposed a model inspired

on Comaniciu and implement a prototype to evaluate

our proposal. The prototype was applied in 640x480

fetal cardiac ultrasound images with four chamber cut

plans (Nelson, 1998).

The searching pattern is calculated based on the

gray scale histogram and the space location of pixels

of a given object of interest. This gray scale was used

VISAPP 2007 - International Conference on Computer Vision Theory and Applications

108

as parameter in the calculation of the searching mold.

The method process is composing by three stages:

the ﬁrst, the second and the third. The ﬁrst stage of

the process consists in the selection of the region of

interest to be found. This structure region is isolated,

and then used in the calculation of the searching mold.

This task is done by the user, who interacts with the

prototype in order to select the limits of the structure

inside of an image.

The second stage of the localization process is the

search for regions that looks like the searching mold.

In this search it is used the same calculation that gen-

erated the mold, applied to several regions of the im-

age, which are called candidates. From the distri-

butions generated in the candidate regions, it is es-

timated the position of the pattern that is more similar

among the candidate regions and the searching pattern

used.

The last stage of the process is the search for the

region of interest inside the image. At this stage, the

similarity is calculated through the Bhattacharyya co-

efﬁcient. The whole process is shown in Figure 2.

2. Candidate molds

1. Searching mold

Result

3. Similarity

Figure 2: Prototype model (q

is the searching mold and

, the candidate regions).

3.1 Calculation of the Searching Mold

To determine the mold of the cardiac structure, the

method follows this procedure: a circular region, ra-

dius h, with center x

, placed in the center of position

of the desired structure of the region. For each point

(or pixel), x = (x

), in the region, a vector of char-

acteristics is extracted and categorized according to

a discrete number of characteristics. This point re-

ceives an index of that characteristic, u = b(x). The

distribution of characteristics, q = {q

}

u=1...m

, which

computes the occurrence of a given characteristic u in

the region of the desired structure, is calculated by:

∑

i=1

k(|x

−x

|/h)δ(b(x

),u)

∑

i=1

k(|x

−x

|/h)

(1)

Where x

is the center of the region, and δ is the

delta Kronecker function

. Notice that the distribu-

tion satisﬁes

∑

i=1

= 1.

The function k(x) is an isotropic kernel that re-

duces the importance of characteristics removed from

the center, in the distribution calculation q. Specif-

ically, the important characteristic is the gray scale

of the pixel. The distribution (q) represents an his-

togram of gray q = {q

}

u=1...,m

which incorporates

spatial and color information of the image pixels.

Figure 3 shows a scheme with the stages of the

generation process of the mold. In this ﬁgure, it

is possible to see the window of the system, devel-

oped at Matlab (Gonzalez, 2004), with an echocar-

diographic image where the region of the image used

for the calculation of the mold is selected. To this

mold, it is applied the equation 1, and the vector q

is calculated and stored for later use in searching for

images of different patients. Figure 4 shows the distri-

bution characteristic of the molds of cardiac structure

of different patients.

Figure 3: Calculation of searching mold.

3.2 Search in the Image of Interest

At this stage of the process, it is necessary to scan

given image searching for a similar mold to the

searching mold generated in the previous stage. The

solution developed for searching was preprocessing

the image with entropy texture feature to select only

the ROI.

To search for the candidate region, it must be as-

sumed that in this region of the image the distribution

The Kronecker delta function returns 1 if its arguments

are equal and 0, otherwise.

AUTOMATIC HEART LOCALIZATION IN ULTRASOUND FETAL IMAGES

109

Figure 4: Distribution characteristic of the molds in differ-

ent patients.

is similar to the searching mold of the structure. By

doing this, the searching mold equation was used for

the calculation of the candidate regions as in Equa-

tion 2.

(y, h) =

∑

i=1

k(|x

−y|/h)δ(b(x

),u)

∑

i=1

k(|x

−y|/h)

(2)

In the search, the image is examined with a mov-

ing window and a set of candidates generated for later

comparison with the mold. Figure 5 shows the pro-

cess of searching in an image. The number of candi-

date patterns depends on the diameter of the searching

pattern and the image processed.

pu n

pu 2

pu 1

Figure 5: Process of searching for structures.

3.3 Finding of the Structure in the

Image of Interest

Spotting the cardiac structure takes place by compar-

ing the candidate mold found at the previous stage

with the structure mold generated during the ﬁrst

stage of the process. The comparison is done using

the Bhattacharyya coefﬁcient (Djouadi et al., 1990),

which provides an index of similarity between the

mold distribution and candidate mold.

The Bhattacharyya coefﬁcient is a measure of the

statistical separability of classes, and gives an esti-

mate of the probability of correct classiﬁcation. It is

a divergence-type measure that has a straightforward

geometric interpretation. It is the cosine of the angle

between n-dimensional vectors. The closer to 1 the

provided value is, the more similar the vectors are.

The calculation is presented in the Equation 3.

ρ(y) = ρ[p(y),q] =

∑

u=1

(y)

√

(3)

The method spots the searching structure through

the patterns bearing more similarity. Only the can-

didate region bearing more similarity is selected; all

the others are ignored. Even so, the calculation of the

candidate region is done considering the whole image.

The Figure 5 shows this process.

qu − mold

Bhattacharyya coefficient

pu 2 pu npu 1 (candidate mold)

candidate mold = 2

More Similar

Figure 6: Finding of the cardiac structure.

4 RESULTS

The tests was applied in an image sample where the

searching mold has been calculated from an image

and the search occurred over 33 different images. In

the sample showed in Table 1 we classiﬁed the results

in three types: success (second column of table 1),

ﬁnd two chamber (third column of table 1) and fail

(fourth column of table 1). The ﬁrst class is the objec-

tive of work, i.e. success on the search for the heart

structure on the image. When the method ﬁnds the

heart, but not totally we classiﬁed with two chamber

and when the method fail, the third class is select. The

last line in the table shows the means of all tests re-

sults.

The Figure 7 shows images with the results, the

ﬁrst image showed in the Figure 7(a) is a instance of

VISAPP 2007 - International Conference on Computer Vision Theory and Applications

110

(a)

(b)

(c)

(d)

Figure 7: Results obtained on tests.

Table 1: Results obtained with the prototype tests.

search mold and it localization on the image, the oth-

ers two, Figure 7 (b) and (c) are instances of success,

and the Figure 7 (d) is a search result instance of class

”ﬁnd two chamber”.

5 CONCLUSION

This paper presents a model developed for pattern

recognition of cardiac structure in echocardiographic

images, as well as the necessary modiﬁcations for

its improvement. One of reason that motives the re-

search about echocardiographic images has been the

dynamics feature. In reality, these images are extract

of videos of the cardiac dynamics that allow work-

ing with these sequences on future. This approach

can take advantage of the dynamics information on

the heart test that nowadays does not be used.

The images used were kindly provided by the fe-

tal cardiology team at Institute of Cardiology of Porto

Alegre. Those images were captured with an echocar-

diographic machine produced by Siemens (Aspen).

That machine allows the recording of images in DI-

COM format. The resolution of the images was

640x480 pixels.

It has been observed that the size of the search-

ing mold inﬂuenced the results; larger molds showed

better performance. A small region may not represent

adequately the structure. This happens because of the

noise, the size of the heart is variable, and the search

is based exclusively on the intensity of gray.

The use of texture feature, to select the region

of interest had been important to increase the perfor-

mance of the method, because texture can separate the

ultrasound image of the background.

The prototype developed can automatically ex-

tract the pattern of cardiac structure of echocardio-

AUTOMATIC HEART LOCALIZATION IN ULTRASOUND FETAL IMAGES

111

Figure 8: Trial obtained, only mean.

graphic images. Considering the graphic depicted in

Figure 8 and the class ”ﬁnd two chambers” as ob-

jective, we can conﬁrm the success of the proposed

method.

REFERENCES

Brown, M. S. et al. (2001). Patient-speciﬁc models for lung

nodule detection and survellience in ct images. IEEE

Transactions on Medical Imaging, 20(12):1242–1250.

Bruijne, M., Niessen, W. J., Maintz, J. B. A., and Viergever,

M. A. (2003). Localization and segmentation of aortic

endografts using marker detection. IEEE Transaction

on Medical Imaging, 22(4):473–482.

Brusseau, E., de Korte, C. L., Mastik, F., Schaar, J., and

van der Steen, A. F. W. (2004). Fully automatic lumi-

nal contour segmentation in intracoronary ultrasound

imaging a statistical approach. IEEE Transactions on

Medical Imaging, 23(5):554–566.

Burckhardt, C. B. (1978). Speckle in ultrasound b-mode

scans. IEEE Transactions on Sonics and Ultrasonics,

SU-25(1):1–6.

Comaniciu, D., Ramesh, V., and Meer, P. (2003). Kernel-

based object tracking. IEEE Transactions on Pattern

Analysis and Machine Intelligence, 25(5):564–577.

Comaniciu, D., Zhou, X. S., and Krishnan, S. (2004). Ro-

bust real-time myocardial border tracking for echocar-

diography: An information fusion approach. IEEE

Transaction on Medical Imaging, 23(7):849–860.

Crimmins, T. R. (1985). Geometric ﬁlter for speckle reduc-

tion. Applied Optics, 24(10):1438–1443.

Djouadi, A., Snorrason, O., and Garber, F. D. (1990).

The quality of training-sample estimates of the bhat-

tacharyya coefﬁcient. IEEE Transactions on Pattern

Analysis and Machine Intelligence, 12(1):92–97.

Duncan, J. S. and Ayeche, N. (2000). Medical image anal-

ysis: Progress over two decades and the challenges

ahead. IEEE Transactions in Pattern Analysis and

Machine Intelligence, 22(1):85–105.

Gonzalez, R. C. (2004). Digital image processing : using

matlab. Pearson Prentice Hall, Upper Saddle River, 1

edition.

Hope, T., Linney, N., and Gregson, P. (2005). Using the

local mode for edge detection in ultrasound images.

In Proc of. Canadian Conference on Electrical and

Computer Engineering, pages 374–377, Canada. Los

Alamitos: IEEE.

Jacob, G., Noble, J. A., Behrenbruch, C., Kelion, A. D.,

and Banning, A. P. (2002). A shape-space-based ap-

proach to tracking myocardial borders and quantifying

regional left-ventricular function applied in echocar-

diography. IEEE Transaction on Medical Imaging,

21(3):226–238.

Kang, S. C. and Hong, S. H. (2002). A speckle reduction ﬁl-

ter using wavelet-based methods for medical imaging

application. In Proc of. 14th International Conference

on Digital Signal Processing, DSP2002, volume 2,

pages 1169–1172, Santorini, Greece. Los Alamitos:

IEEE.

Lee, Y., Ishigaki, T., et al. (2001). Automated detection of

pulmonary nodules in helical ct images based on an

improved template-matching technique. IEEE Trans-

action on Medical Imaging, 20(7):595–604.

Mattos, S. S. (1999). O Corac¸

ao Fetal. Revinter, Rio de

Janeiro/RJ.

Nelson, T. R. (1998). Ultrasound visualization. In Advances

in Computers, volume 47, pages 185–253. Academic

Press, New York.

Salvadorls, A., .Maingourd, Y., Ful, S., and LeralluT, J.-

F. (2003). Optimizaton of an edge detection algo-

rithm for echocardiographic images. In Proc. of the 25

Annual International Conference of the IEEE EMBS,

Cancun, Mexico. Los Alamitos: IEEE.

Sheehan, F. (2000). Echocardiography. In Handbook of

Medical Imaging, volume 2, pages 609–674. Spie,

Bellingham.

Sugioka, K. et al. (2003). Automated quantiﬁcation of

left ventricular function by the automated contour

tracking method. ECHOCARDIOGRAPHY: A Jor-

nal of Cardiovascular Ultrasound and Allied Tech.,

20(4):313–318.

Valdes-Cristerna, R., Jimenez, J., Yanez-Suarez, O., Leral-

lut, J., and Medina, V. (2004). Texture-based echocar-

diographic segmentation using a non-parametric esti-

mator and an active contour model. In Proc. of the

26th Annual International Conference of the IEEE

EMBS, San Francisco, US. Los Alamitos: IEEE.

Zong, X., Laine, A., and Geiser, E. (1998). Speckle reduc-

tion and contrast enhancement of echocardiograms via

multiscale nonlinear processing. IEEE Transactions

on Medical Imaging, 17(4):532–540.

VISAPP 2007 - International Conference on Computer Vision Theory and Applications

112