Research Journal of Applied Sciences, Engineering and Technology 4(24): 5416-5422, 2012 ISSN: 2040-7467 © Maxwell Scientific Organization, 2012 Submitted: March 18, 2012 Accepted: April 13, 2012 Published: December 15, 2012 Evaluation of Level Set Segmentation for Medical Images with Intensity Inhomogeinity 1 M. Renugadevi, 1V. Vaithiyanathan, 1K.R. Sekar and 2N. Raju 1 School of Computing, 2 School of EEE, SASTRA University, Thanjavur, India Abstract: Image segmentation is an important area in the medical image guide surgery. Major advances in the field of medical imaging provide richer information in clinical applications and support the advancement in the biomedical knowledge. With the growing research on image segmentation, it has become crucial challenge for the images with inhomogeneity in intensity. Level set is the most powerful and broadly used segmentation technique in the medical image processing. This study aims at making a review on the current level set methods such as Region Scalable Fitting (RSF), Statistical and Variational Multiphase Level Set (SVMLS) and Local Clustering based Variational Level Set (LCVLS) developed for intensity inhomogeneous medical image segmentation. Experiments that apply these algorithms to segment the medical images are presented to highlight the distinct characteristics of each method. Results prove that the LCVLS method is most suitable and accurate for intensity inhomogeneous medical image segmentation. Keywords: Biomedical imaging, Intensity Inhomogeneity (IIH), level set method, segmentation INTRODUCTION Medical image analysis plays a crucial role in many image processing applications. It aids in developing computational algorithms and methods that helps to analyze biomedical data in various clinical applications such as diagnosis, study of anatomical structure, computed integrated surgery and treatment planning. The current research on the biomedical imaging focuses on the segmentation which facilitates the visualization of medical data and delineation of anatomical structures. Various techniques such as threshold based methods, watershed transformation based methods and partial differentiation equation based methods have been employed so far for image segmentation problems. Among all these techniques, active contour method has been proved to be an efficient framework in the past two decades. Active contour model is a framework for marking out the object outline from the target image driven by the sum of internal and external energy (Kass et al., 1988; Blake and Isard, 1998; Paragios and Deriche, 2000). Edge based methods (Xu and Prince, 1998; Caselles et al., 1993; Kichenassamy et al., 1996) and Region based methods (Ronfard, 1994; Mumford and Shah, 1989; Chan and Vese, 2001; Vese and Chan, 2002; Paragios and Deriche, 2002; Caselles et al., 1997; Tsai et al., 2001; Li et al., 2007; Zhang et al., 2010a, b) are the two kinds of the existing active contour models. Edge based active contour methods are suitable for segmenting the images with well defined boundaries by utilizing the image gradient to evolve the contour. Region based active contour model builds the region descriptor to guide the motion of the contour and has better performance than the former model by segmenting the images having weak boundaries. Implicit implementation of the active contour model is represented as level set which is one of the most attractive and popular region based method (Sethian, 1999). It has been introduced by Osher and Sethian. (1988) and is defined as a numerical technique for capturing dynamic interfaces and shapes. It is widely used because of its advantageous properties such as topology adaptability and robustness to initialization. Nowadays segmentation is a compelling challenge due to the presence of Intensity Inhomogeneity (IIH) in the images. It often exists in the ordinary or medical images due to various factors such as imperfection of imaging devices, static field inhomogeneity, non-uniform daylight and artificial illumination. In recent times, many level set algorithms (Li et al., 2007, 2008, 2011; Zhang et al., 2010a, b) are used to segment the non-uniform intensity homogeneous images especially medical images. Yun-Jen (Chiu et al., 2010) explains the three active contour models and concludes that RSF model (Li et al., 2008) serves as good one for heterogeneous medical image segmentation among them. So in this study, we review three recently proposed level set methods to segment the intensity inhomogeneous medical images. They are Region Scalable Fitting (RSF) energy method proposed by Li et al. (2008), Statistical and Variational Corresponding Author: M. Renugadevi, School of Computing, SASTRA University, Thanjavur, India 5416 Res. J. Appl. Sci. Eng. Technol., 4(24): 5416-5422, 2012 2 i =1 ⎛ ⎛ ∇φ ⎞ ⎛ ∇φ ⎞ ⎞ ⎟⎟ + µ ⎜ ∇ 2φ − div⎜⎜ ⎟⎟ ⎟ div ⎜⎜ ⎜ φ ∇ ⎝ ⎠ ⎝ ∇ φ ⎠ ⎟⎠ ⎝ Region-scalable fitting model: Li et al. (2008) proposed the Region-Scalable Fitting (RSF) energy method to segment the medical images with intensity inhomogeneities. It performs well for the images with weak object boundaries. Let K: Un6[0, +4) be a nonnegative kernel function with a localization property which is defined as 8(u) $ 8(v), if |u|<|v| and lim |u|÷48(u) = 0 plays a key role in this model. Let C be a closed contour in the image domain S dUn, that separates S into two regions S1 (outside (C)) and S2 (inside (C)). For a given point x 0 S, the local intensity fitting energy is defined as: σ 2 Ωi F (φ , f 1 , f 2 ) = ε ε (φ , f 1 , f 2 ) + µP(φ ) where, : is a positive constant and 2σ (2) 2 ε (C, f 1 ( x ), f 2 ( x )) = ∫ ε xFIT (C , f 1 ( x ), f 2 ( x))dx + v C (3) Which is converted to a level set formulation with the Heaviside function H as in Eq. (4) where M1(K) = H(K), M2(K) = 1! H(K) and the last term in Eq. (4) computes the length of the zero level contour. Here the Heaviside function H is approximated: ε (φ , f 1 , f 2 ) = ∑ λi ∫ ∫ κ σ ( x − y ) I ( y ) − f i ( x ) M i (φ ( y ))dy i =1 dx + v ∫ ∇ H (φ ( x ) ) dx ( ) 2 1 ∇ φ ( x ) − 1 dx . 2 ei ( x ) = ∫ κ σ ( x − y ) I ( x ) − f i ( y ) dy , i = 1, 2 and 2 f i ( x) = κ σ ( x ) * ⎣ M ε i (φ ( x )) I ( x ) ⎦ κ σ ( x ) * M ε i (φ ( x )) , i = 1, 2 The first term is the data fitting term which is responsible for driving the active contour, the second term has smoothing effect on the zero level contour to maintain the regularity of the contour and the third term is called level set regularization term. n ( P(φ ) = ∫ (8) 2 The energy functional is defined for a contour C as: 2 (7) The energy functional in Eq. (4) is now approximated by replacing H with HJ and it can be expressed as in Eq. (5) where Mg1(K) = Hg (K) and Mg2(K) = 1 ! Hg (K). The energy functional along with a level set regularization term P(N ) is defined as: where, 81 and 82 are positive constants and f1(x) and f2(x) are two values that approximate image intensities in S1 and S2 respectively. Equation (1) is called a RegionScalable Fitting (RSF) of the contour C at a point x since the intensities I(y) in above model are in the local region centered at the point x and its size is controlled by the kernel function: e− u 1⎡ 2 1 ε ⎛ x⎞ ⎤ 1 + atc tan⎜ ⎟ ⎥ δε = Hε' ( x ) = ⎝ ε⎠ ⎦ 2 ⎢⎣ π π ε 2 + x2 The standard gradient descent method is used to minimize this energy functional with respect to N by solving the following gradient flow in Eq. (8) where: ε xFIT (C, f 1 ( x ), f 2 ( x )) = ∑ λi ∫ κ ( x − y) I ( y ) − f i ( x ) dy (1) 2 (6) by a smooth function defined as HJ and its derivative is defined as *J: Hδ = 1 (2π ) n 2 ) (5) ∂φ = − δε (φ )( λ1e1 − λ2e2 ) + vδε (φ ) δt This section presents and illustrates the effectiveness of each method for segmenting the intensity inhomogeneous medical images. These methods are mainly based on the level set technique but used in its distinct way. The level set algorithm of each method is explained as follows: κ= 2 dx + v ∫ ∇ H (φ ( x ) ) dx METHODOLOGY i =1 ( ετ (φ , f 1 , f 2 ) = ∑ λi ∫ ∫ κ σ ( x − y ) I ( y ) − f i ( x ) M iτ (φ ( y ))dy Multiphase Level Set (SVMLS) method proposed by (Zhang et al., 2010b) and Local Clustering based Variational Level Set (LCVLS) proposed by (Li et al., 2011). 2 ) (4) SVMLS model: Zhang et al. (2010b) proposed a novel Statistical and Variational Multiphase Level Set (SVMLS) method to simultaneous intensity inhomogeneous image segmentation and bias correction. With a function I : S ÷ Udefined on a continuous domain S, the given intensity inhomogeneous image is modeled as: I(x) = b(x) J(x) + n (x), x , S (9) where, I(x) is the measured image, b(x) is spatially variant bias field, J(x) is the true signal that is assumed as 5417 Res. J. Appl. Sci. Eng. Technol., 4(24): 5416-5422, 2012 piecewise constant and n(x) is assumed to be Gaussian distributed with zero mean and variance F2 . To accurately model the statistical characteristics of the image, the intensity corresponding to the object domain Si is described as: ( ⎛ ( I ( y ) − b( x ) c ) 2 ⎞ i ⎟,y ∈Ω exp⎜ − 2 ⎜ ⎟ σ 2 2πσ i i ⎝ ⎠ ) p I ( y) αi = i =1 Ω Ωi 1 ∑ I ( y σi ) = mi ( x ) y∈Ω i ∩ Ox i =1 Ω ⎛ where, d i ( y) = Ω∫ κ p ( x, y )⎜⎝ log( 2πσ i ) + ( E (α )∆ − ∫ log p( D α ) dx (13) which is integrated with the likelihood function as: i =1 {( ) D = I x α i , i = 1,..., N N ∏ i = 1 y ∈Ω i ∩ Ox }, ( ( )) p I y αi (14) ( )( ) (18) [ ] [ ] I = bJ +n (20) Equation (13) is called Statistical and Variational Multiphase Level Set (SVMLS). Let us consider the characteristic function of region Ox as: ⎧⎪ 1, y − x ≤ ρ else ⎩⎪ 0, ) LCVLS model: Li et al. (2011) proposed a local clustering based variational level set framework for simultaneous heterogeneous image segmentation and bias correction. Let an image I is considered as a function I : S ÷U defined on a continuous domain S. And to deal with intensity inhomogeneities, it is modeled as: i = 1 Ω Ω i ∩ Ox ( ( )) ∝ ∏ )( where, *(N) is the Dirac function. The smoothness of the bias field is ensured by the normalized convolution. Ω N ) ⎧ ∂ φ1 ⎪ ∂ t = − (d 1 − d 2 − d 3 + d 4 ) H (φ2 ) + d 2 − d 4 δ (φ1 ) ⎪ ⎨ ⎪ ∂ φ2 = − (d − d − d + d ) H (φ ) + d − d δ (φ ) (19) 1 2 3 4 1 3 4 2 ⎪⎩ ∂ t The energy functional is defined as: p( D α ) = ∏ p I x α i ( The energy functional E{b, ci, Fi} is minimized with respect to each variable by fixing other variables. The derived gradient decent by minimizing the energy functional ESVLMS K N ," is as follows: y ∈Ωi ∩ Ox N ( I ( y ) − b( x )cu ) 2 ⎞ ⎟ dx 2σ i2 ⎠ ⎧ M = H (φ ) H (φ ) M 2 = H (φ1 ) 1 − H (φ2 ) 1 2 ⎪ 1 ⎨ ⎩⎪ M 3 = 1 − (φ1 ) H (φ2 ) M 4 = 1 − H (φ1 ) 1 − H (φ2 ) σ2 ⎞ ∑ ∫ ∫ log( p( I ( y αi )))dy dx (17) In the energy minimization for four-phase case, Mi is defined with the Heaviside function H(N) as: (11) ⎛ κ ρ ( x, y) = ⎨ (16) N ∏ p( I ( y αi )) ≈ p( I ( x αi ))mi ( x ) ∝ N ⎜⎜⎝ bci , mi (i x) ⎟⎟⎠ (12) where, ( I ( y ) − b( x ) cu ) 2 ⎞ ⎟⎟ dydx 2σ i2 ⎠ E SVLMS Φ N , α = ∑ ∫ d i ( y ) M i (Φ N ( y ))dy where, mi(x) = ||Si 1 Ox|| and I ( x σ i ) is normally distributed. The product of the Gaussian probability density function is still Gaussian, so that: = cons tan t − ) 2πσ i + Let (Mi (KN (.)) be the characteristic function of region Si , where KN (.)is a function of set{Ki , i = 1 ,... n}. Then Eq. (13) can be rewritten as: (10) mapping T: I ( x α i ) → I ( x σ i ) from the original image intensity domain D(T) to another domain U(T) is defined as: ( ) ⎛ ∫ ∫ κ p ( x − y)⎜⎜⎝ log( 1 where, Fi is a standard deviation of intensity, b(x)ci is spatially varying local mean and ai = {b ,ci , Fi }. To formulate the energy functional, a circular neighborhood center is considered for each position x , S, D. A i . e . , Ox = { y y − x ≤ ρ} w i t h radius I x σi N E (α )∑ (15) By eliminating the trivial constant term and with Eq. (10) and (15) the energy functional E (") can be rewritten as: where, J is the true image which is assumed to be piecewise constant since it measures an intrinsic physical property of the objects, b is a bias field component that denotes the intensity inhomogeneity and assumed to be varying slowly and n is additive noise that is assumed to be zero-mean Gaussian noise. To derive a local intensity clustering property, a circular neighborhood with a radius D centered at each point y , S is defined by Oy ∆ { x: x − y ≤ ρ} . The partition of the neighborhood Oy is induced by the 5418 Res. J. Appl. Sci. Eng. Technol., 4(24): 5416-5422, 2012 partition{Si}Ni = 1 of the entire domain S.Then specifically for classifying the intensities I(x) in the neighborhood Oy , a clustering criterion is defined as: ε y = ∑ ∫ K ( y − x ) I ( x ) − b( y )ci dx N ⎛ N ⎞ 2 N ( ) i =1 2 (27) where, L(N) and Up(N) are the regularization terms defined below. The term L(N) that computes the arc length of the zero level contour of N is defined by: (28) The term Up(N ) is defined by: ℜ p (φ ) = ∫ p( ∇ φ )dx (29) (23) (24) that forms a partition of the domain S. The level set formulation of the energy with N = 2 and N>2, called two phase and multiphase formulation, respectively. Here only the two phase formulation is considered. In the two case level set formulation, the regions S1 and S2 can be represented with their membership function defined by M1(N) = H(N) and M2(N) = 1!H(N) respectively, where H is the Heaviside function [19].Thus, the energy functional can be expressed as: ε = ∫ ∑ ∫ K ( y − x ) I ( x ) − b( y)ci dy Mi (φ ( x ) )dx F (φ , c, b) = ε (φ , c, b) + vL(φ ) + µℜ p (φ ) L( φ ) = ∫ ∇ H (φ ) dx where, a is a normalization constant such that ∫ K (u) = 1 , F is the scale parameter of the Gaussian function and D is the radius of the neighborhood Oy. The above defined energy functional is in terms of the regions S1, ..., SN. To derive the solution for the energy minimization problem, the energy is converted to a level set formulation. Let N:S 6 U be a level set function with two disjoint region: Ω 1 { x:φ ( x ) > 0} andΩ 2 { x:φ ( x ) < 0} (26) The variational level set formulation uses the above defined energy as the data term which is defined by: (22) Here the kernel function K is truncated Gaussian function defined by: ⎧1 2 ⎪ e − u / 2σ 2 , for u ≤ ρ K ( u) ⎨ a ⎪⎩ 0, otherwise i =1 2 (21) where, K(y-x) is kernel function (nonnegative window function), such that K(y-x) = 0 for x ó Oy . This is called local clustering criterion function that evaluates the classification of the intensities in the neighborhood Oy . The smaller value of gy yields the better classification. So to minimize the gy with respect to y over the image domain S , an energy is defined as: ε ∆ ∫ ⎜ ∑ ∫ K ( y − x ) I ( x ) − b( y )ci dx⎟ dy ⎝ i =1 Ωi ⎠ N where, ei ( x ) = ∫ K ( y − x ) I ( x ) − b( y )ci dy 2 i =1 Ω ε (φ , b, c) = ∫ ∑ ei ( x ) M i (φ ( x ) ) dx (25) For convenience, the constants c1, ..., cN can be represented as a vector c = (c1, ..., cN ). Thus the variables of the energy functional g are the level set function g, the vector c and the bias field b. So the energy functional g can be written as: where, p is a potential or energy density function p: [0,4)6U such that p(s) = (1/2)(s!1)2. With this potential p, the energy term Up(N ) is minimized by maintaining ∇ φ = 1 which is called signed distance property. And so Up(N ) is called distance regularization term (Li et al., 2010). The energy F(N, c, b) is minimized by an iteration process with respect to each of its variables (Li et al., 2011). Thus solution for the energy minimization problem is achieved by minimizing this energy term that results in the image segmentation given by the level set function N and estimation of the bias field b. EXPERIMENTAL RESULTS The performance evaluation of the three IIH level set segmentation methods are demonstrated in this section. The intensity inhomogeneous medical images are employed to compare the ability of these methods. The comparison is based on the segmented result, number of iterations desired to complete the segmentation and the CPU processing time for the contour evolution. These measures are commenced to pick up the suitable model for the IIH medical image segmentation. The segmentation results obtained using RSF, SVMLS and LCVLS methods on the heterogeneous MRI slice image of the brain are shown in Fig. 1. For the level set evolution, the initial contour is located at the same position in the image for all the three methods. The curve evolution of the RSF model is shown in Fig. 1a. From 5419 Res. J. Appl. Sci. Eng. Technol., 4(24): 5416-5422, 2012 Processing time 25 (a) 20 15 10 5 0 RSF SVMLS LCVLS Models (a) Processing time (b) (c) Fig. 1: Shows the segmented results of three models in MRI brain image. Contour evolution of RSF, SVMLS and LCVLS methods are shown in, (a), (b) and (c), respectively. 50 45 40 35 30 25 20 15 10 5 0 SVMLS RSF LCVLS Models (b) Fig. 3: Shows the processing time and number of iterations charts for MRI image (a) Table 1:Shows the performance analysis of three models using MRI brain image Parameters ---------------------------------------------------------Models Processing Time Iteration RSF 160.975 400 SVML S72.493 70 LCVLS 9.563 30 (b) Table 2: Shows the performance analysis of three models using human vertebra phantom image Models Parameters ---------------------------------------------------------Processing Time Iteration RSF 22464 50 SVMLS 23.7434 20 LCVLS 0.7176 10 (c) Fig. 2: Shows the segmented results of three models in cross sectional image of a human vertebra phantom. Contour evolution of RSF, SVMLS and LCVLS methods are shown in (a), (b) and (c) respectively. the Fig. 1, we analyze that the RSF model can’t complete its contour evolution even after 2000 iterations. Also, it does not segment the inner regions of the image. The SVMLS segmentation result shown in the Fig. 1b is not perfect since it fails to capture the interior parts of the MRI image. The satisfied and accurate segmented result of the LCVLS model is shown in Fig. 1c. Figure 2 is another example to validate the effectiveness of three models using the cross sectional 5420 Res. J. Appl. Sci. Eng. Technol., 4(24): 5416-5422, 2012 ACKNOWLEDGMENT Iteration 180 160 The authors thank Dr. B. Shanthi, Professor of CSE Department, SASTRA University for her imputing discussions on this manuscript. We thank Dr. N. Sairam, Professor of CSE Department, SASTRA University for his valuable suggestions. Also we wish to thank Prof. T. S. Varadharajan of English Department, SASTRA University for his linguistic support. 140 120 100 80 60 40 20 0 RSF SVMLS REFERENCES LCVLS Models (a) 400 350 Iteration 300 250 200 150 100 50 0 SVMLS RSF LCVLS Models (b) Fig. 4: Shows the processing time and number of iterations charts for human vertebra phantom image image of a human vertebra phantom. As shown in Fig. 2a, the RSF model segments the image but it also detects the uninterested regions of the image. Though the SVMLS model segments the image, it is not satisfied and not smooth as shown in Fig. 2b One can see the satisfied and desired segmented result of the LCVLS model in Fig. 2c. The total time and the number of iterations taken for the curve evolution of MRI image and vertebra image are reported in the Table 1 and 2 and graphically depicted in the Fig. 3 and 4, respectively. CONCLUSION This study has presented the comparison review on the recent level set segmentation methods that have been developed for addressing the challenge of the heterogeneous medical image segmentation. The goal of this common approach is to identify the suitable level set method for segmenting the medical images with intensity inhomogeneous. The advantage and the effectiveness of each method is evaluated and the experimental results show that the LCVLS model is the promising one for the IIH medical image segmentation. Blake and M. Isard, 1998. Active Contours. Springer, Cambridge, MA. Caselles, V., F. Catte, T. Coli and F. Dibos, 1993. A geometric model for active contours in image processing. Numer. Math., 66(I): 1-31. Caselles, V., R. Kimmel and G. Sapiro, 1997. Geodesic active contour. Int. 1. Comput. Vis., 22(I): 61-79. Chan, T. and L. Vese, 2001. Active contours without edges. IEEE T. Image Proc., 10(2): 266-277. Chiu, Y., V. Pham, T. Tran and K. Shyu, 2010. Evaluation of active contour on medical inhomogeneous image segmentation. Paper Presented at the Proceedings-2010 3rd IEEE International Conference on Computer Science and Information Technology, ICCSIT 2010, 1: 311-314. Li, C., C. Kao, J. Gore and Z. Ding, 2007. Implicit active contours driven by local binary fitting energy. Presented at the IEEE Conference on Computer Vision and Pattern Recognition. Li, C., C. Kao, J. Gore and Z. Ding, 2008. Minimization of region-scalable fitting energy for image segmentation. IEEE T. Image Proc., 17(10): 19401949. Li, C., C. Xu, C. Gui and M.D. Fox, 2010. Distance regularized level set evolution and its application to image segmentation. IEEE T. Image Proc., 19(12): 3243-3254. Li, C., R. Huang, Z. Ding, C. Gatenby, D.N. Metaxas and J.C. Gore, 2011. A level set method for image segmentation in the presence of intensity inhomogeneities with application to MRI. IEEE T. Image Proc., 20(7): 2007-2016. Kass, M., A. Witkin and D. Terzopoulos, 1988. Snakes: Active contour models. Int. J. Comput. Vis., l(4): 321-331. Kichenassamy, S., A. Kumar, P. Olver, A. Tannenbaum and A.Y. Jr, 1996. Conformal curvature flows: From phase transitions to active vision. Arch. Ration. Mech. Anal., 134(3): 275-301. Mumford, D. and J. Shah, 1989. Optimal approximations by piecewise smooth functions and associated variational problems. Comrnun. Pure AppI. Math., 42(5): 577-685. 5421 Res. J. Appl. Sci. Eng. Technol., 4(24): 5416-5422, 2012 Osher, S. and J. Sethian, 1988. Fronts propagating with curvature-dependent speed: Algorithms based on Hamilton-Jacobi formulations. J. Comput. Phys., 79(1): 12-49. Paragios, N. and R. Deriche, 2000. Geodesic active contours and level sets for the detection and tracking of moving objects. IEEE T. Pattern Anal. Mach. Intell., 22(3): 226-280. Paragios, N. and R. Deriche, 2002. Geodesic active regions and level set methods for supervised texture segmentation. Int. J. Comput. Vis., 46(3): 223-247. Ronfard, R., 1994. Region-based strategies for active contour models. Int. 1.Comp. Vis., 13(2): 229-25I. Sethian, J., 1999. Level Set Methods and Fast Marching Methods. 2nd Edn., Springer, New York. Tsai, A., Yezzi and A.S. Willsky, 2001. Curve evolution implementation of the Mumford-Shah functional for image segmentation, denoising, interpolation and magnification. IEEE T. Image Proc., 10(8): 1169-1186. Vese, L. and T. Chan, 2002. A multiphase level set framework for image segmentation using the Mumford and Shah model. Int. J. Compo Vis., 50(3): 271-293. Xu, C. and Prince, 1998. Snakes, shapes and gradient vector flow. IEEE T. Image Process., 7(3): 359-369. Zhang, K., H. Song and L. Zhang, 2010a. Active contours driven by local image fitting energy. Pattern Recognition. Zhang, K., L. Zhang and S. Zhang, 2010b. A variational multiphase level set approach to simultaneous segmentation and bias correction. Paper presented at the Proceedings-International Conference on Image Processing, ICIP, pp: 4105-4108. 5422