Segmentation of Left Ventricle of 2D Echocardiographic image using Active contouring

Shweta Deopujari Electronics engineering department

Yeshwantrao Chavan College of Engineering Nagpur, India

shweta.deopujari@gmail.com

Yogita Dubey Electronics & Telecommunication department Yeshwantrao Chavan College of Engineering

Nagpur, India yogeetakdubey@yahoo.co.in

Abstract- Cardiovascular diseases are a major health concern worldwide. The left ventricle and in particular the endocardium is a structure of particular interest since it performs the task of pumping oxygenated blood to the entire body. Therefore, segmentation of the left ventricle in echocardiographic images is a task with important diagnostic power. Cardiac function is evaluated quantitatively using echocardiography via the analysis of shape attributes, such as the heart wall thickness or the shape change of the heart wall boundaries. This requires that the complete boundaries of the heart wall be detected from a sequence of two-dimensional ultrasonic images of the heart. The image segmentation process is made difficult since these images are plagued by poor intensity contrast and dropouts caused by the intrinsic limitations of the image formation process. This paper describes a method for obtaining the border of the heart’s left ventricle which is useful in the detection of certain heart parameters. The limitation of speckle noise is overcome by smoothening the image using Gaussian filtering.

Keywords— Echocardiography, Left ventricle, speckle noise, Segmentation, Active-contour.

I. INTRODUCTION The term ‘‘echocardiography” refers to a group of tests

where the ultrasound is used to examine the heart and record information in the form of echoes (sound wave reflections) . The maximum limit for a sound to be audible is 20,000 cycles/s or 20 KHz. The frequency used in ultrasounds is from 1 to 10 million cycles/s . The frequencies used for adults are usually 2.0–5.0 MHz, whereas the ones used for children are usually higher, from 3.5 to 10 MHz’s. The resolution of the image registration, which is the ability to distinguish two objects near in space, is directly proportional to the frequency and inversely proportional to the length of the wave. A high frequency ultrasound (short wave) identifies objects less than 1 mm away. Rays with the lowest frequency and longest wave have a more deficient resolution. On the other hand, the intensity of sound penetration, which is the ability to transmit enough ultrasonic energy into the thorax to provide an adequate range, is inversely proportional to the signal’s frequency. A high frequency ultrasonic ray (e.g. 5 or 10 MHz) does not

penetrate a thick thoracic wall, thus lower frequency ultrasound rays are used for adults. Although this enables the rays to penetrate the thoracic wall, it partially sacrifices the resolution; however, even with a transducer that produces 2.5 MHz rays, which is frequently used in adult echocardiographies, it is still possible to identify objects at 1 and 2 mm away from each other.

Two-dimensional echocardiography is an ultrasonic imaging technique that is used as an important noninvasive technique in the characterization of the left ventricular structure and function. Quantitative analysis of the cardiac function often uses shape attributes such as the thickness of the heart wall, the enclosed area, and the measurement of the variation of these shape attributes throughout the cardiac cycle .These analyses require the complete determination of inner (endocardial) and outer (epicardial) boundaries of the heart wall. Current studies in this area often require the tedious and time-consuming process of having expert operators outline the boundaries. This becomes increasingly labor intensive when the analysis of a complete cardiac cycle is needed to provide a description of the systolic and diastolic wall motion pattern, or when a number of frames taken from different views have to be processed to obtain points in three-dimensional space for surface reconstruction. Furthermore, automated definition of the boundaries would improve the reliability of the quantitative analysis by eliminating the subjectivity of manual tracing. Finding boundaries in echocardiograms automatically by computers is often difficult because of the poor quality of the images. This is due to the intrinsic limitations of echo imaging such as low image intensity contrast, dropouts in the image, and boundary discontinuity in any given frame. In echo imaging, a pulse is sent along a ray from a transducer towards the organ that is being imaged. The pulse is attenuated and reflected when it hits a medium with acoustic impedance different from that of the medium in which the pulse is traveling. The time it takes in transit is a measure of the distance of the boundary from the transducer, and the amount of energy that is reflected is a measure of the difference of the acoustic impedance across the boundary. In practice, since the energy of the pulse diminishes as it travels, the post processing of the reflected signal includes

2011 International Conference on Computational Intelligence and Communication Systems

978-0-7695-4587-5/11 $26.00 © 2011 IEEE

DOI 10.1109/CICN.2011.146

680

2011 International Conference on Computational Intelligence and Communication Systems

978-0-7695-4587-5/11 $26.00 © 2011 IEEE

DOI 10.1109/CICN.2011.146

674

2011 International Conference on Computational Intelligence and Communication Systems

978-0-7695-4587-5/11 $26.00 © 2011 IEEE

DOI 10.1109/CICN.2011.146

672

time gain control that compensates the attenuation of the signal over time. The poor quality of the echocardiograms is attributed to the reverberations of the pulse and speckle noise, caused by the backscattering of the incident wavefront after it hits the tissue microstructures. Another limitation of this imaging technique is that the reflection is not very pronounced when the angle between the boundary of the organ and the ray that the pulse is traveling along is small. As the lateral parts of the heart wall boundaries are usually not very well defined in the images an approximation to the ventricle’s contours in a long- and short-axis views is carried out.

The outline of the paper is as follows. In Section II, we review principal works on image segmentation on echocardiography images. This section focuses on application areas in which the majority of effort has been focused.

II. HISTORY OF LEFT VENTRICLE SEGMENTATION This section includes some prior studies carried out in

the field of echo-cardiographic images, specifically left ventricle segmentation starting from contour extraction models. The methods suggested by Chu et al. [1] require mostly the gray level information along with user defined initial contours to extract the boundary in the images. The method proposed by Staib et al. [2] implemented a probabilistic deformable model considering the boundary as two-dimensional deformable object using maximum posteriori estimate. Chalana et al. [3] reports an interesting approach to detect epicardial and endocardial boundary of short axis echocardiographic sequences using a multiple active contour model which is an extension to the original model proposed by Kass et al. [4]. The multiple-active-contour model is a special case of active surface model where the surface is represented as a sequence of planar contours. The algorithm requires user defined initial approximation for epicardial boundary that detects the contour by computing gradients using Canny’s edge detection method. The variance of the Gaussian kernel used to convolve the gradient image progressively decreases to intensify the convergence. The optimized contours of the epicardial boarders are used as initial approximation for endocardial boundary with empirically determined values of the snake model parameters. A recent work on segmentation of medical images has been reported using geometric active deformable models where the contour propagates with a velocity profile as a function of curvature. A Hough transform technique by Antonio Fernandez-Caballero et al. [5] is used to find an initial approximation of the object boundary at the first frame of the sequence . Then, an active-contour model is used in a coarse-to-fine framework, for the estimation of a noisy LV (Left Ventricle) boundary. The automatic detection of the boundary of left ventricle in a sequence of cardiac images has been proposed by Mishra et al. [6]. The contour detection algorithm is formulated as a

constrained optimization problem based on active-contour model. The optimization problem has been solved using a genetic algorithm (GA). The result obtained by the proposed GA based approach is compared with conventional nonlinear programming methods. Although digital techniques have improved, the echocardiographic images are inherently noisy due to the air in the lungs; therefore, initial point detection relying on thresholds can be complicated, since these thresholds can change from one image to the next.

A. Left ventricular segmentation in the short- and long-axis views The left ventricle and in particular the endocardium is a

structure of particular interest since it performs the task of pumping oxygenated blood to the entire body (Paragios et al.[7]) .Therefore, segmentation of the left ventricle in echocardiographic images is a task with important diagnostic power. More concretely, contour extraction is an important criterion for subjective evaluation of the cardiac function and has become an area of focus. Medical image segmentation is a difficult task due to low signal-to-noise ratios and poor contrast. These problems can be solved with the help of active-contours and other segmentation techniques. The application in medicine of image analysis and of computational vision and calculus has been a deciding factor in the increase of medical diagnosis accuracy and treatment plans and hence segmentation is essential to any image analysis system. Active-contour techniques allow integrating different information elements into the segmentation process. It is possible to combine information from regions and boundaries with shape and/or statistical models. However, several problems must be solved in the classical implementation. One of them is requiring good initialization to stop the local minimum energy in the energy minimization process. Although, at first, the curve has to be initialized the active-contours try to find object boundaries by heading for pixels where abrupt changes break the homogeneity of some of the scenes in the image. One of the most widely used features to identify these points is the gradient. High gradient is usually associated with a noticeable boundary. However, when dealing with medical images, it is very common to find boundaries with a gradient similar to that of the object boundary, due to noise or to other objects.

III. PREPROCESSING The major edge detection algorithms fail due to the

presence of noise and the low contrast in heart echocardiograph imagery and the improvement of echocardiograph images is very important for the accurate detection of both heart boundary and movement of heart valves. Therefore, noise reduction must be applied before edge detection, in the recent years several techniques proposed to reduce the noise without distorting the relevant

681675673

clinical details. Smoothening is primarily used for diminishing noise and/or spurious effects in an image. A Gaussian filter will be used by means of the mathematical operations given as below

(1)

A potentially good image can be filtered by convolving

with a 3 × 3 Gaussian low pass filter fig. (1) followed by thresholding at the mean value of the blood pull intensity. Almost all approaches for contours extraction use common image processing procedures, the most important ones being pre-processing and edge detection. The main differences between these approaches are in the level of the pre-processing procedures to achieve an improved image and get an effective boundary tracking.

a. b. Fig .1(a) Left ventricle in Long axis view. (b) smoothened image by

Gaussian 3×3 filter with � = 3.

IV. SEGMENTATION : ACTIVE CONTOUR MODELS Segmentation subdivides an image into its constituent regions or objects. Automatic LV segmentation is a difficult task due to the relatively poor quality of echocardiography images. Many researchers have proposed algorithms such as Active contour or snakes [4], in the past for image segmentation tasks but all of them consume extensive computation time. Snakes [4], or active contours, by Kass, M. et al. are curves defined within an image domain that can move under the influence of internal forces coming from within the curve itself and external forces computed from the image data. The internal and external forces are defined so that the snake will confirm to an object boundary or other desired features within an image. The basic idea of active contour is to use a deformable model and let it evolves in each iteration to minimize a given energy function. Two main categories exist for active contours: edge-based and region-based. The edge-based method is highly sensitive to the noise and dependent to initial curve location (the initial curve should be placed near the edges). The advantage of edge-base method is correct segmentation in most of the cases since there is no global constraint on the image. Advantages of region-based method in comparison with

edge-based method are insensitivity to noise and robustness to initial curve placement. However the disadvantage is that sometimes it leads to erroneous image segmentation. Edge-based active contour models utilize image gradients in order to identify object boundaries, e.g., [9], [10]. This type of highly localized image information is adequate in some situations, but has been found to be very sensitive to image noise and highly dependent on initial curve placement. One benefit of this type of flow is the fact that no global constraints are placed on the image. Thus, the foreground and background can be heterogeneous and a correct segmentation can still be achieved in certain cases. There are many advantages of region-based approaches when compared to edge-based methods including robustness against initial curve placement and insensitivity to image noise. However, techniques that attempt to model regions using global statistics are usually not ideal for segmenting heterogeneous objects. In cases where the object to be segmented cannot be easily distinguished in terms of global statistics, region-based active contours may lead to erroneous segmentations. Heterogeneous objects frequently occur in natural and medical imagery. To accurately segment these objects, a new class of active contour energies should be considered which utilizes local information, but also incorporates the benefits of region-based techniques. Lankton et al.[8] also propose the use of localized energies in 3-D tensor volumes for the purpose of neural fiber bundle segmentation. All of these works focus on a localized energy that is based on the piecewise constant model of Chan and Vese [11].

The basic idea of this paper is to start with initial mask (e.g. a square/ rectangle) represented in a form of a closed curves, i.e. contours, and then iteratively modify them which result in shrinking /expansion operations according to the constraints of the image. This operation is known as contour evolution which is performed by minimization of an energy function. The contour is defined in the (x, y) plane of an image as a parametric curve

v(s)=(x(s), y(s)) (2)

Contour is said to possess an energy (Esnake) which is defined as the sum of the three energy terms. (3) • Internal energy depends on the intrinsic properties of

the curve. It is defined as sum of elastic energy and bending energy.

• Elastic Energy (Eelastic): The curve is treated as an elastic rubber band possessing elastic potential energy. It discourages stretching by introducing tension.

(4)

2 2

2 22 22 2

1 1( , ) ( )2 2

x y r

g x y e e g rσ σσ σπσ πσ

+− −

= = =

int intsnake ernal external constraE E E E= + +

21 ( ) | |2elastic sE s v ds= α�

s

( )s

dv svds

=

682676674

Weight α(s) allows us to control elastic energy along different parts of the contour which is considered to be constant α for many applications. It is responsible for shrinking of the contour. • Bending Energy (Ebending):

The snake is also considered to behave like a thin metal strip giving rise to bending energy. It is defined as sum of squared curvature of the contour.

(5)

β(s) plays a similar role to α(s). Total internal energy of the snake can be defined as

(6)

• External Energy:

It is derived from the image. Define a function Eimage(x,y) so that it takes on its smaller values at the features of interest, such as boundaries.

(7)

Key rests on defining Eimage(x,y). Some examples (8)

The problem at hand is to find a contour v(s) that minimize the energy functional

(9)

Each term corresponds to a force produced by the respective energy terms. The contour deforms under the action of these forces.

(a) (b)

Fig.2 (a) Detection of the left ventricle in the long-axis view. (b) Segmented output

Fig.2 (c) Detection of the left ventricle in the short-axis view. (d) Segmented output

V. CALCULATION OF DIMENSION OF LEFT VENTRICLE

After applying the active contour algorithm, we get an image

with just the desired object (i.e. Left Ventricle) as shown in Figure 2(b). From this image a number of parameters are

calculated as shown in Table 2.Parameters such as area, length and volume are determined from the segmented output. Area can be calculated by counting the number of white pixels in the segmented region or by the equation (10). Minor axis (width) of LV can be found by calculating

distance between left and right corner of the image.

(10)

where, (xi, yi) are from {1, 2, …, n} and n is the number of contour co-ordinates . From the area and minor axis (Width), one can calculate major axis (Length) of LV using equations (11 and 12) ,

(11)

To calculate the LV chamber volume (Volume), we use Area-Length method as given in equation (12).

(12)

TABLE 1. STANDARD PARAMETERS OF HUMAN LEFT

VENTRICLE

LVED (cm) Normal

Mildly dilated

Moderately dilated

Severely dilated

Men 4.2-5.9 6.0-6.3 6.4-6.8 >6.9

Women 3.9-5.3 5.4-5.7 5.8-6.1 >6.2

21 ( ) | |2bending ss

s

E s v ds= β�

2 2int

1 | | | | )2elastic bending s ss

s

E E E v v ds= + = (α +β�

( ( ))ext images

E E v s ds= �2( , ) | , )|imageE x y x y= − ∇Ι(

2( , ) | ( ( , )* ( , )) |imageE x y G x y I x yσ= − ∇

2 21 ( ) | | ( ) | | ) ( ( ))2snake s ss image

s

E s v s v E v s ds= (α +β +�

683677675

TABLE 2. LEFT VENTRICLE DIMENSIONS OF SOME PATIENTS

AFTER SEGMENTATION

Image Image view

Area (cm2)

Volume (cm3)

Length (cm)

Width (cm)

1 Long axis 14.76 38.31 4.83 4.29

Short axis 5.61

2 Long axis 22.41 59.786 7.14 4.39

Short axis 16.42

3 Long axis 9.51 22.32 3.45 3.67

Short axis 7.4

4 Long axis 13.14 30.34 4.84 3.93

Short axis 12.52

5 Long axis 11.9 27.04 4.45 3.62

Short axis 8.56

VI. RESULT AND DISCUSSION

All the echocardiographic images of patients are obtained from Wockhardt Hospital. These images are of size 640×480 and measurements are conducted by extracting the image area of size 571×551 through cropping .The proposed method was coded in MATLAB R2008 followed by some filtering and selection of initial contour and finally the ventricles traced automatically. From the segmented output, we calculate the length, width and area of left ventricle.

VII. CONCLUSION This paper presented a novel approach for detecting ventricle boundaries. The desired speed for analyzing the

patient’s echocardiographic images is achieved through resizing the image. A variety of images in apical long axis view (2 chamber) and short axis view were used for our experiments. We observed that the initial selection of the contour is important to obtain good result. Not all images provide clear cavities and in such cases special filters may help to remove such unwanted regions. Also images with discontinuity along the boundary can be improved by using Gaussian smoothening.

REFERENCES [1] Chu, C. H., Delp, E. J., & Buda, A. J. (1988). “Detecting left ventricular endocardial and epicardial boundaries by digital two dimensional echocardiography”. IEEE Transactions on Medical Imaging, 7, 81–90. [2] Staib, L. H., & Duncan, J. S. (1992). “Boundary finding with parametrically deformable models.” IEEE Transactions on Pattern Analysis and Machine Intelligence, 14, 1061–1074. [3] Chalana, V., Linker, D. T., Haynor, D. R., & Kim, Y. (1996). “A multiple active contour model for cardiac boundary detection on echocardiographic sequences”. IEEE Transactions on Medical Imaging, 15, 290–298. [4] Kass M., Witkin A., & Terzopoulos D. (1988). “Snakes: Active contour models”. International Journal of Computer Vision, 1, 321–331. [5] Antonio Fernandez-Caballero, Jose´ M. Vega-Riesco. “Determining heart parameters through left ventricular automatic segmentation for heart disease diagnosis”. Expert Systems with Applications 36 (2009) 2234–2249©2007ElsevierLtd.All rights reserved doi:10.1016/j.eswa.2007.12.045 [6] A. Mishra, P. K. Dutta, and M. K. Ghosh, “A GA based approach for boundary detection of left ventricle with echocardiographic image sequences”. Image Vis. Compute., vol. 21, pp. 967–976, 2003. [7] Paragios, N., Jolly, M. P., Taron, M., & Ramaraj, R. (2005). “Active shape models & segmentation of the left ventricle in echocardiography”. Lecture Notes in Computer Science, 3459, 131–142 [8] S. Lankton, A. Tannenbaum, “Localizing Region-Based Active Contours” IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 17, NO. 11, NOVEMBER 2008 [9] S. Kichenassamy, A. Kumar, P. Olver, A. Tannenbaum, and A. Y. Jr, “Conformal curvature flows: From phase transitions to active vision,” Arch. Ration. Mech. Anal., vol. 134, no. 3, pp. 275–301, Sep. 1996. [10] V. Caselles, R. Kimmel, and G. Sapiro, “Geodesic active contours,” Int. J. Comput. Vis., vol. 22, no. 1, pp. 61–79, Feb. 1997. [11] T. Chan and L. Vese, “Active contours without edges,” IEEE Trans. Image Process. vol. 10, no. 2, pp. 266–277, Feb. 2001.

684678676