Efficient Watershed Approach Applied to Children Heart Chambers Segmentation

S.Josevin Prem PSN College of Engineering and Technology Email:-mailtojosevin@gmail.com

AbstractSegmentation of echocardiographic images presents a great challenge because these images contain strong speckle noise and artifacts. Besides, most ultrasound segmentation methods are semiautomatic, requiring initial contour to be manually identied in the images. In this paper, we propose an algorithm based on the phase symmetry approa ch and level set evolution, in order to ex-tract simultaneously all heart cavities in a fully automatic way. The level set evolution uses a new logarithmic-based stopping function, which demonstrates to perform well in the boundary extraction. We compared our method with other level set approaches, the watershed technique, and the manual segmentation made by two physicians. The experimental work was based on echocardiography images of children. Similarity metrics, namely Pratt function, pixel mean error, and similarity angle have been used for the performance evaluation of the different methods. The results indicate that our method has a performance of at least 4% superior to the other methods able to segment the four chambers. Even for the two worst boundary extraction cases (right ventricle and left atrium), the performance of the proposed method is still better than the other techniques. Index TermsEchocardiographic images, heart segmentation, level set, phase symmetry, similarity index.

1. INTRODUCTION The Active contour models are today one of the most used approaches for medical image boundary detection, due to its ability to overcome the speckleinduced error. Kass et al. [1] were the rst to propose this methodology, based on an energy minimization scheme. In the recent years, Stoitsis et al. [2] used the Hough transform as an initial estimation for the active contour model, based on the gradient vector ow eld, to automatically extract the carotid artery from sequences of B-mode images. The performance of their active contours was analyzed by comparing the extracted boundaries with the ones traced manually by an expert. Good similarity was observed between both contours. An important drawback related to these classical (parametric) approaches is their incapacity to split and merge, which is fundamental for the detection of more than one region of interest in an image. That formulation can handle complex shapes and topological changes. The idea is to embed the contour into a higher dimensional function (x, y, t) [see (3)], called level set, where the function can evolve from its initial stage to the image boundaries, according to different models. If the velocity term F of the level set equation is based on the image edges, the model is called edge based. Most edge-based models consist of two components: the regularity component, which determines the shape of contours, and the edge

detection component, which attracts the contour toward the edges. Mumford and Shah [8] proposed an alternative segmentation model called region based. This segmentation approach relies on the statistical information of the image intensity to minimize an energy function instead of searching edges. Using the Mumford and Shah Segmentation model, Chan and Vese [9] proposed a method that is a particular case of the minimal partition problem, where the velocity term F is restricted to piecewise constant functions. That model automatically detects inward contours starting with an initial curve that can be anywhere in the image. A model that uses intensity information in local regions was also proposed by Li et al. [10]. First, they dene a region-scalable tting energy functional and two tting functions that locally approximate the image intensities on the two sides of the contour. Paragios et al. [11], [12] proposed the geodesic active region model for supervised segmentation, integrating edge and region based in an energy function. Zhanget al. [13] also applies this association to achieve robust and ac-curate segmentation results. He tried to solve the problem from a more geometric perspective, describing the advantages of the model over Chan and Vese [9] method on medical images with complex backgrounds. The watershed algorithm is also widely applied in the medical domain as a powerful segmentation tool. Cheng et al. [14] combine the advantages of the watershed methodology with those

of active contours (water snakes) for boundary detection and tracking of left ventricle. The method uses watershed to presegment the image, which speeds up the boundary detection and generates the initial position of the snake. Deka and Ghosh [15] proposed a segmentation model having four stages. It starts with the preprocessing stage that consists of applying a histogram equalization method and a ltering, using an adaptive median lter to reduce the effect of the speckle noise. In the second stage, a gradient map suitable for watershed segmentation based on a multiscale morphological gradient is produced. This gradient image is then threshold before applying the watershed transformation to reduce the over segmentation problem. Junior et al. [16] used a semiautomatic method to extract the left ventricle internal border using classical image processing techniques combined with radial search. The preprocessing phase includes a sequence of lters and morphological operations. Then, the watershed algorithm is used to identify and label regions in the preprocessed image followed by contour correction before proceeding with area calculation. In this paper, we propose a segmentation method based on level set that is capable to identify simultaneously the four heart chambers. The phase symmetry approach, which acts in the frequency domain extracting low-level features, is of great importance for the success of the algorithm [18], [19]. The segmentation results are compared with those provided by the region-based [9], [10], hybrid [13], and watershed [20] methods. That evaluation was carried out using as reference the contours drawn by the physicians, using three metrics of similarity. The idea with the segmentation of all four cavities is related to the diagnosis time, for instance when some parameters need to be calculated (as ejection fraction, wall thickness, etc.). II. METHODS In the following section, the proposed level set method with an improved stopping function is described for the simultaneous heart cavities segmentation as well as three alternative level set functions [9], [10], [13] and the classical watershed method [20]for comparison purpose. Then, the performances of the different methods are quantied with three gures of merit, using as reference the contours drawn manually by expert physicians. A. Proposed Method The rst step with the proposed method consists of the detection of low-level features in the

echocardiographic image that stay invariant to changes in image brightness or contrast, i.e., we are searching for points where there is a specic order in the frequency domain. Then, the level set segmentation algorithm is applied to extract the cavity boundaries. 1) Phase Symmetry: Any discrete signal can be represented by the sine and cosine functions with specic amplitudes, giving rise to a set of scaled waves in the time domain that synthesizes the original signal. Phase congruency is a low-level feature detector type, derived from the Fourier analysis; a general dentition about how signal features can be described and detected depending on the phase of the Fourier components is given by Morrone and Owens [21].The points detected by the phase congruency are signicant if they occur over a wide range of frequencies. At the points of the edge transition, all Fourier components are exactly in phase. For all other points, the phase of the sinusoidal components varies resulting in a low phase congruency. To calculate the local frequency, in particular the phase information, wavelets in symmetric or antisymmetric pairs must be used, having band-pass quadrature lters as mother wavelets. This is a bank of lters tuned for different frequency ranges (scales) allowing the identication of features either in the space or in frequency at various scales making it possible to obtain local information along with scale space. Besides being difcult to implement, phase congruency is also very sensitive to noise. Inspired by the properties of the phase congruency, Kovesi [22] showed that symmetries and antisymmetries give rise to special phase pat-terns in the image intensity values. All Fourier series at points of symmetry are either at minima or at maxima of their cycle.

Fig. 1. (a) Original echocardiographic image. (b) Output of the phase symmetry. In this paper, log-Gabor wavelets have been used to extract simultaneously spatial and frequency information. These lters are important, because they avoid the problem of redundancy at low frequencies. Kovesi has adapted the phase congruency measuring

to the detection of symmetry and antisymmetry. The symmetry measure is dened as [ ] [ ][ [ ] [ ] [ ]

[ [ ] [ ] [ ]]

evolutions depending on the applied evolution methodology. For the proposed method, we implement F depending on the edge indicator function (P) F = [div /|| + ] P (4)

[ ] ]

(1)

and the antisymmetry is given by [ ]

[ [ ] [ ] [ ]]

(2)

Where es [n] is the even symmetric part and os [n] is the odd symmetric part of the lter at scale s. The s joins implicitly the information of the various lters at different scales. At a point of symmetry, the absolute value of es will be large and the absolute value of os is small. At points of antisymmetry, the response is reversed and as a consequence the odd part of the lter will have a large response while the absolute value of es is small [22]. The As [n] parameter is the magnitude of the lter response at a given wavelet scale and point n in the image, the term is a small constant, assumed with the same value (0.0001) as the one used by Kovesi, to prevent division by zero in the case where the signal is uniform and no lter response is obtained, and the factor T is a noise compensation term. Extension to 2-D images [23] is made by applying the1-D analysis over several orientations and combining the results to provide a single measuring of the feature signicance. To calculate 2-D phase symmetry of a given image, the image is rst convolved with a bank of log-Gabor lters. 2) Level Set Evolution: The main idea of the level set is to minimize a function (x,y) by solving the corresponding partial differential equation (PDE) using the level set evolution equation as a numerical method. The method evolves a contour (2-D) or surface (3-D) implicitly by manipulating a higher dimensional function. For this particular paper, where 2-D images are used, the evolving contour is extracted from the zero-level curve of : C = {x|(x)=0}. The general curve evolution PDE in the level set framework is /t= ||F (3)

Where the term is an outward growing force, providing a faster convergence and div (/||) is the curvature term. The stopping function P is responsible for attracting the contour toward the image boundaries. The classical stopping term is the convolution of the image intensity with a Gaussian lter. As already mentioned, such a stopping function presents some drawbacks; thus, we have used an alternative stopping function based on a logarithmic variation P = log (I /+1) (5)

Where I is the image, is the average intensity value of the image, and is the dynamic range of the region being analyzed. The logarithmic variation has some advantages when compared to the linear variation. Thus, for low |I | values, the level set is nearby the region to be segmented reducing its evolution speed (low P). For high |I | values, the level set is far away from the region to be segmented; then, the evolution speed remains high and almost constant. An important advantage of using this stopping function is the capability of the level set to adjust itself to the low-intensity regions. For the curve evolution PDE, we have used an initial mask0 as a signed distance function in the Euclidean space, dened as follows: the central pixel of the mask image has the largest value and decrease at each neighbor element, ending with a zero value at the four corners of image. The function is then iteratively modied according to the mean curvature term div (/||) and the stopping function P being updated at each time interval. It suffers some shrinking or expansion iterations until the functional has converged to the boundary. The level set will produce surfaces with zero value on the contour, and positive or negative whether inside or outside the segmented objects. Due to the echocardiography image characteristics and the level set nature, several small noisy regions are also detected and should be eliminated to put in evidence only the four cardiac cavities [see Fig. 2(a)]. Since the result of the segmentation is a binary image, the following procedure was carried out. First, to detect each

Where F is a speed function designed for the boundary detection. The level set function and the evolving curve C at the zero-level change together with time. The speed function F has different

contour, an initial pixel is selected for one contour, and then one proceeds through its adjacent contour pixels until the starting pixel is reached again. In this way, all pixels that form the contour are identied. This procedure is repeated for each contour and the respective area is calculated. The contours corresponding to the four biggest areas are the heart chambers [see Fig. 2(b)]. As a nal step, the irregularities in the detected con tours are smoothed using the dilation and erosion morphological operations [see Fig. 2(c)].

image intensities are statistically homogeneous for each region. The minimization of the energy is achieved by minimizing the integral over the points in the image, and then smoothing the contour by penalizing its length. Note that || is replaced by the smoothed Dirac function, (6) can be written as t= () [ div ( /| |) (1 e1 2 e2) + (2 div ( /| |)] (7) The term (div (/||)) has a smoothing effect and is fundamental to maintain a regularized contour. The data tting term (1 e1 2 e2) is responsible for driving the contour toward the object boundary. The last term, called level set regularization term searches for the regularity of the function. The work of Zhang et al. [13] was also considered in our study as a hybrid model, once it integrates edge and regionbased segmentation in the energy functional t= || [(I )+ (div (P /||))] (8)

Fig. 2. (a) Output of the proposed level set, (b) after discarding small regions, and (c) nal output, after smoothing with morphological operations. B. Alternative Level Set Methods In this section, three alternative level set formulations are described and their performances are evaluated in order to deter-mine the best one for the segmentation of B-mode echo cardio-graphic images. Due to the edge-based nature of our proposed method, we focus our attention on alternative speed terms F based on the image features. Some authors replace the gradient term || by the regularized Dirac functional () to remain close to the minimization problem. In the region-based method-ology, the region information of the target objects is used without image gradient related terms. The authors also have used the regularized Dirac functional (always different from zero) t= () [2 (I out) 2 1 (I in) 2 + (div (/||))] (6) Where the rst two terms of F (1 and 2) measure variations inside/outside the contour; the area inside the contour is given by the third term and the length of the curve is measured by the fourth term. The last two ones are regularization terms. Since medical images may have complex backgrounds, the use of the thresholds in and out can fail. Indeed, their global nature is based on the assumption that the

The speed term F integrates both boundary and region information. The rst term encourages the contour to enclose the regions with gray levels greater than a specic value. The second term aids the contour to attach areas with high image gradients. C. Watershed The watershed segmentation is based on region morphological dilation. Instead of working on the image itself, this technique is applied to the image gradient aiming to search for regions of highintensity gradients (watersheds) that divide neighboring local minima (basins) [26]. In this context, the image is considered to be a topographic surface where high-amplitude pixels correspond to ridge points, and low-amplitude pixels identify valley points. A water source is placed at each regional minimum (catchment basin) lling up the basins gradually. In each subsequent step (dilations), 3-D topography is ooded from the valleys and the pixels covered by the rising water are selected. Valley regions are grown until at least two catchment basins have merged. Whenever different sources meet, vertical dams are erected on those ridgelines, preventing overow between catchment basins. As a common property of all water-shed algorithms, the

ridgelines appear as the equidistant points between two adjacent minima. D. Similarity Metrics In the medical domain, the evaluation of methodologies is mandatory before they could be used in clinical practice. Taking into account that rule, the proposed segmentation algorithm is compared with the aforementioned alternative methods, having as reference the contour drawn by the physicians. There are different approaches to compare two contours [27].It can be made based on the local distances between the two contours or comparing the regions inside the contours. The rst approach starts detecting the mean contour, a line equidistant to both contours under analysis, computed after over-lapping both contours [see Fig. 3(a) and (b)]. In the regions where the contours coincide, the mean contour also coincide; in the regions where the contours do not coincide, an equidistant line is dened. Each point in this mean contour is identied and a perpendicular line is traced that will intersect the two contours under analysis; the two points where this interception occurs are called corresponding points [see Fig. 3(b)]. The distance di used in several similarity metrics is the distance between corresponding points.

PF =1/N

(9)

Where di is the distance between corresponding points (size of the perpendicular lines in Fig. 3) and N is the number of pair of points. The normalization parameter is related to the contour size. The PF varies between 0 and 1, where 1 is the exactly overlapping of the two contours. Pixel mean error (PME): This metric gives a global average view of the distances between contours and is dened as PME=1/N (10)

This metric uses the binary images resulting from the lled contours to be compared. Two vectors H and I are dened. Each vector is the linearization of each image. If one image has n m pixels, the linearization vector will be a 1-D vector with m n pixels. These two vectors H and I can be considered elements of a Hilbert space, where SA denotes the angle between the two binary image vectors. Using the denitions of the Euclidean inner product we dened the similarity angle as: SA=arc cos (H I/||H|||| I||) (11)

The angle varies between 0 and 90 for two similar and dis-similar contours; respectively .A statistical analysis was carried out using these three gures of merit to compare the performance of each method with the experts manually drawn contours.

Fig. 3. (a) Contour produced by the algorithm (dark) overlapped on the manual drawn (light), (b) two contours with the mean contour and the pairing lines, and(c) another example of the pairing lines. We used two similarity metrics based on di; the Pratt function and the pixel mean error. The second approach compares the binary images obtained by lling the inside of the contours. The metric that we used based on this approach was the similarity angle. Pratt function (PF): The Pratt function is a nonlinear metric, characterizing the global behavior of the discrepancies between the comparing contours, which is given by

Fig. 4. Contours obtained using the proposed method overlapped with con-tours manually drawn by the two physicians. III. RESULTS AND DISCUSSION As already mentioned, the phase symmetry approach was used in all studied methods as an important initial segmentation procedure to emphasize the edges. Then, each segmentation algorithm was evaluated to characterize its ability to extract simultaneously the four cardiac cavities. The

described methods need no user intervention and convergence is achieved after a few iterations. The resulting boundaries were then compared to the ones drawn by two physicians. Contours, obtained using the proposed method overlapped with contours manually drawn by physicians, are shown in Fig. 4. The segmentation performance for all the methods is illustrated in Fig. 5.The comparative analysis was made by the previously de-scribed statistical metrics using the 240 chambers. Thus, that analysis was applied to 240 chambers using the proposed method, to 172 chambers using the watershed, to 200 chambers with the hybrid algorithm proposed by Zhang, to 96 chambers using Lis region-based scalable tting method, and to only36 chambers by Chans algorithm. A. Performance of the Five Methods Fig. 6 shows a box plot graphic using the PF for all segmentation methods. Here, only 36 contours were considered.

From Fig. 6, the median values for each method are, respectively, 0.520, 0.465, 0.480, 0.450, and 0.445. It is observed that the proposed method presents the best median value, while the Chan, Zhang, and Watershed methods have the lower median values, suggesting the worst segmentation performance. The Zhang and the Watershed methods have the upper quartile in a lower position (from 0.51 to 0.58) when compared to the position of the same quartile of the proposed method (from 0.56to 0.66) that suggests a lower performance for the Zhang and Watershed methods. B. Comparison of the Three Methods Since the region-based methods proposed by Chan and Liare only able to segment 36 and 96 chambers, respectively, we have discarded these two methods and restricted the analysis to the three methodologies able to segment the larger number of chambers (e.g., the proposed, Zhang and Watershed ones). This procedure permitted using 172 chambers in the evaluation of those three methods. Fig. 7 illustrates the performance of the three methods applied to each cavity measured by the similarity angle metric.

Fig. 5. Contours from all methods overlapped with a reference drawn by a physician: (a) Proposed, (b) Chan, (c) Li, (d) Zhang, and (e) Watershed. Fig. 7. Similarity angle of the three segmentation methods with better performance, for each cavity using a total of 172 chambers. It is observed that the median value for the four cavities is always lower for the proposed method, having similarity angles of 27, 25, 25 and 38 for the right ventricle, left ventricle, right atrium, and left atrium, respectively. The most problematic regions are the right ventricle and the left atrium that present the larger dispersion range and the higher median

Fig. 6. Pratt functions of the ve segmentation methods using 36 contours.

values (high similarity angles mean high disagreement between the contours).That is veried for the Zhang and Watershed methods. Nevertheless, it is observed that the proposed method performs better for the referred cavities, providing always lower median values and a smaller dispersion range.

Fig. 9. Worst cases of the segmentation, corresponding to (a) the right ventricleand (b) to the left atrium.

Fig. 8. Pixel mean error of the proposed method for each cavity on the 240chambers. C. Evaluation of the Proposed Method The proposed model was applied to 240 chambers. Its performance was evaluated considering each heart cavity separately using the PME metric. The results are shown in Fig. 8.The median value of the PME for all cavities is around 5pixels, except for left atrium that presents an error of 9 pixels. Observing the second and third quartiles as well as the entire range, better agreement is veried between the reference con-tours, and the left ventricle and the right atrium. On the contrary, for the problematic cavities (right ventricle and left atrium), it was not always possible to get good segmentation. This is justied not only by the outliers (right ventricle), but also by the large dispersion range of values (left atrium) as shown in Fig. 8.Two examples of these regions, corresponding also to the two worst boundary extraction cases for the proposed method, are illustrated in Fig. 9. The low performance of the proposed algorithm related to the segmentation of the referred two cavities is connected to the variability of physician interpretations about the shape of the contours. The right ventricles moderator band that is visible in some frames justies the incorrect boundary detection by the level set algorithm [see Fig. 9(a)]. Fig. 9(b) shows the imperfect segmentation of the left atrium. The difculty experimented by the algorithm is directly related to the problems that physicians have to delineate similar boundaries for this cavity

Fig. 10. Heart cavities delineated by two different physicians. Fig. 10(a) and (b) shows two echocardiographic images, illustrating contours made by the two expert physicians. On the left image, the physician considered part of the pulmonary vein as belonging to the left atrium, and the result was a rough con-tour contrasting with the delineation illustrated on the right image, where the other physician produced a smooth and regular contour. This demonstrates the importance to have an automatic segmentation algorithm that provides the same result for a given image. IV. CONCLUSION In general, the physicians carry out the heart cavity measurements in a manual way according to their own knowledge, skill, and perspective. Full automatic segmentation methods are therefore important for the standardization of heart measurements and to assist the physicians detecting congenital defects in the clinical practices, as well as for the 3-D (4-D) heart reconstruction. We proposed a method that segments simultaneously the four heart cavities using echocardiographic B-mode images. The method consisted of two steps. The rst one

enhances the cardiac boundaries, applying the phase symmetry approach along with six orientations of the image. It is the rst time that phase symmetry is used to segment the whole heart cavities, providing substantial noise reduction. Then, the edge-based approach has been chosen for the level set evolution, with an improve d logarithmic-dependent stopping function, which demonstrates better performance than traditional Gaussian or exponential functions. It was demonstrated that the proposed method is capable to segment automatically and simultaneously the four heart chambers that is an important advantage when compared with the available methods that only succeeded segmenting the left ventricle, namely reduced computation time, and analysis of the four cavities in almost real time. A statistical analysis based on three similarity metrics was made comparing the proposed method with the three alternative level set algorithms and the watershed methodology having as reference the contours obtained by two experts. As demonstrated, the segmentation results provided by the proposed method were always superior to the ones used for comparison. The positive results obtained with this paper motivated the authors to proceed with improvements for the proposed method-ology as well as with the development of new approaches for the simultaneous segmentation of the four cardiac chambers, thus assisting image-guided interventions and helping the experts in the clinical diagnosis, namely in the detection of congenital heart diseases REFERENCES [1] M. Kass, A. Witkin, and D. Terzopoulos, Snakes: Active contour models, Int. J. Comput. Vision, vol. 1, no. 4, pp. 321331, 1988. [2] J. Stoitsis, S. Golemati, S. Kendros, and K. S. Nikita, Using the Hough transform to segment ultrasound images of longitudinal and transverse sections of the carotid artery, Ultrasound Med. Biol., vol. 33, no. 12,pp. 191832, 2007. [3] V. Caselles, R. Kimmel, and G. Sapiro, Geodesic active contours, Int.J. Comput. Vision, vol. 22, no. 1, pp. 6179, 1997. [4] R. Malladi, J. A. Sethian, and B. C. Vemuri, Shape modeling with front propagation: A level set

approach, IEEE Trans. Pattern Anal. Mach.Intell., vol. 17, no. 2, pp. 158175, 1995. [5] S. Osher and J. A. Sethian, Fronts propagating with curvature dependent speed: Algorithms based on HamiltonJacobi formulations, J. Comput. Phys., vol. 79, no. 1, pp. 1249, 1988. [6] M. Mora, C. Tauber, and H. Batatia, Robust Level Set for heart cavities detection in ultrasound images, in Proc. Comput. Cardiology. Lyon,France, 2005, pp. 235238. [7] W. Fang, K. L. Chan, S. Fu, and S. M. Krishnan, Incorporating temporal information into active contour method for detecting heart wall boundary from echocardiographic image sequence, Comput. Med. Imag. Graph. vol. 32, no. 7, pp. 590600, 2008. [8] D. Mumford and J. Shah, Optimal approximations by piecewise smooth functions and associated vibrational problems, Commun. Pure Appl .Math., vol. 42, no. 5, pp. 577685, 1989. [9] T. F. Chan and L. A. Vese, Active contours without edges, IEEE Trans. Image Process., vol. 10, no. 2, pp. 266277, Feb. 2001. [10] C. Li, C.-Y. Kao, J. C. Gore, and Z. Ding, Minimization of region-scalable tting energy for image segmentation, IEEE Trans. Image Process.,vol. 17, no. 10, pp. 19401949, Oct. 2008. [11] N. Paragios and R. Deriche, Geodesic active regions: A new framework to deal with frame partition problems in co mputer vision, J. Vis. Commun. Image Represent., vol. 13, pp. 249268, 2002. [12] N. Paragios, Vibrational methods and partial differential equations in cardiac image analysis, in Proc. IEEE Int. Symp. Biomed. Imaging: From Nano to Macro. Arlington, TX, 2004, pp. 1720. [13] Y. Zhang, B. J. Matuszewski, L.-K. Shark, and C. J. Moore, Medical image segmentation using new hybrid level-set method, in Proc. 5th Int. Conf. Biomed. Vis.: Inf. Vis. Med. Biomed. Inf.. London, U.K, 2008,pp. 7176.