978-3-319-26690-9_20.pdf

June 11, 2018 | Author: Gehad Ismail Sayed | Category: Documents


Comments



Description

Abdominal CT Liver Parenchyma Segmentation Based on Particle Swarm Optimization Gehad Ismail Sayed and Aboul Ella Hassanien

Abstract Image segmentation is an important task in the image processing field. Efficient segmentation of images considered important for further object recognition and classification. This paper presents a novel segmentation approach based on Particle Swarm Optimization (PSO) and an adaptive Watershed algorithm. An application of liver CT imaging has been chosen and PSO approach has been applied to segment abdominal CT images. The experimental results show the efficiency of the proposed approach and it obtains overall accuracy 94 % of good liver extraction.







Keywords Watershed Preprocessing Segmentation Particle swarm optimization Post-processing Morphological operators CT Parenchyma









1 Introduction Image segmentation is the process of partitioning the image into meaningful objects that have same visual characteristic. It’s considered an important basic task in analyzing and understanding of the image [1]. There are many image segmentation techniques. Clustering is one of the most important techniques used for image segmentation. In image segmentation, Clustering is the process of identifying natural groupings based on some similarity measure like Euclidean distance [2]. Clustering algorithms are used in many applications, such as data mining, image G.I. Sayed (✉) ⋅ A.E. Hassanien Faculty of Computers and Information, Cairo University, Giza, Egypt e-mail: [email protected] URL: http://www.egyptscience.net A.E. Hassanien Faculty of Computers and Information, Benisuef University, Beni Suef, Egypt G.I. Sayed ⋅ A.E. Hassanien Scientific Research Group in Egypt (SRGE), Cairo, Egypt © Springer International Publishing Switzerland 2016 T. Gaber et al. (eds.), The 1st International Conference on Advanced Intelligent System and Informatics (AISI2015), November 28–30, 2015, Beni Suef, Egypt, Advances in Intelligent Systems and Computing 407, DOI 10.1007/978-3-319-26690-9_20

219

220

G.I. Sayed and A.E. Hassanien

segmentation, machine learning, etc. [3]. Particle Swarm Optimization (PSO) is a population based stochastic optimization technique proposed by Dr. Eberhart and Dr. Kennedy in 1995. The idea of the algorithm inspired by social behavior of bird flocking or fish schooling [4]. It consists of a number of particles (solutions) that move in the search space (pixels in the image) in order to search for global optima (maximizing the distribution of intensity levels in the image [5]. It has been used to optimize parameters for medical images as in [18]. Many PSO versions have recently proposed like in [19]. In this paper, we will focus on PSO that will used to segment Liver for CT image. A fully automatic liver segmentation approach from abdominal CT images based on PSO and an adaptive watershed is presented in this paper. The remainder of this paper is ordered as follows. Section 2 explains the basic concepts of PSO. Section 3 shows the proposed PSO and watershed approach. In Sect. 4 shows the experimental results and analysis with details of the datasets. Finally, conclusion and future work are discussed in Sect. 5.

2 Preliminaries PSO algorithm originally developed by Eberhart and Kennedy in 1995 [4], which is taking advantage of the swarm intelligence concept, for example bird flocks and fish schools. The usual aim of the particle swarm optimization (PSO) algorithm is to solve continuous and discrete optimization problems. It is a population-based method, that is iteratively changed until a termination criterion is satisfied. The population of feasible solutions P = P1 , P2 . . . , pn , p2 pain in PSO is often called a swarm. And each P is called particle. These particles travel through the search space to find the optimal solution. The PSO segmentation based algorithm has been one of the most recently used. It has been compared with GA-based algorithms [8]. The results show that it gives better results in less time also it needs only few parameters to adjust. As PSO has no parameter of “mutation”, “recombination”, and no notion of the “survival of the fittest”. At the beginning of the PSO algorithm, particle position velocities are set to zero and their positions are randomly set within the boundaries of the search space. Global, local and neighborhood are initialized with small values. Population size and stopping criteria are very important parameters that need to optimize in order to get an overall good solution with acceptable time limit. In each step of PSO algorithm fitness function is evaluated and each particle’s position and velocity are updated according to Eqs. (1) and (2). The fitness function is used to indicate how close a particle to the optimal solution. vni+ 1 = wvni + ϕ1 r1 ðgni − xni Þ + ϕ2 r2 ðpni − xni Þ + ϕ3 r3 ðnni − xni Þ

ð1Þ

xni+ 1 = xni + vni+ 1

ð2Þ

where w, ϕ1 , ϕ2 and ϕ3 coefficients are assigned weights. w Is inertial weight which represents the memory of previous velocity and ϕ1 , ϕ2 and ϕ3 are acceleration

Abdominal CT Liver Parenchyma Segmentation …

221

coefficients which represent cognitive (personal), social (neighborhood) component usually set between 0 and 4. The symbols r1 , r2 and r3 represent random variables with uniform distribution between 0 and 1. gni is global best information while nni is neghborhood best and pni is local best.

3 The Proposed Segmentation Approach The proposed abdominal CT Liver parenchyma segmentation approach is comprised of five fundamental building phases. These five phases are described in detail in the following section along with the steps involved and the characteristics feature for each phase. The proposed algorithm can be summarized as in Algorithm (1). And the overall architecture of the introduced approach is described in Fig. 1. Algorithm 1 Proposed Approach 1: Read CT image and resize it 256*256 2: Apply median filter with 3*3 window size to 3: Apply PSO clutser-based algorithm with number of level = 3 4: for each clutser image of

do

5: calculate mean 6: end for 7: Select best

with maximum mean value

8: Apply morphology operators like open and close to enhance 9: Apply adaptive watershed on enhanced

for final final liver

region extraction

3.1

Preprocessing Phase

When image transforms from one form to another like scanning some degradation may occur. Image noise is a kind of these degradations. It is the random variation of brightness or color information in images [9]. This noise is always undesirable. So removing noise with preserving edges of the image plays a vital role in image processing. Median filter is one of the simplest and most popular approaches for removing noise like Salt and pepper [10]. It is calculated by sorting all surrounding neighborhood pixel values in numerical order and then replacing the specified pixel being with the middle pixel value. In this paper, CT image will be resized to 256 * 256 in order to reduce computation time. Then, Median filter with window size 3 * 3 is applied to enhance the CT image.

222

G.I. Sayed and A.E. Hassanien

Fig. 1 The proposed automatic CT liver segmentation architecture

3.2

CT Image Clustering Based on PSO Phase

In this phase the modified image from previous phase will be clustered by using PSO. PSO is used in order to search for the n − 1 optimal n level thresholds that maximize the objective function (fitness function). The initial parameters of PSO are shown in Table 1 which showed the best results obtained.

3.3

Post-Processing Phase: Morphological Operators

In this phase the best cluster image produced from PSO will be selected. Selection criteria depend on the maximum mean value obtained from each cluster. Then the image will be converted to binary with thresholding equal to 0.4 (found to be optimal thresholding value), then open morphological operation used to enhance the clustered image and to focus on liver parenchyma, then the maximum region Table 1 Initial PSO Parameters

Population size Number of iteration

150 10

ϕ1 ϕ2 ϕ3 xmax xmin vmax vmin w Number of levels

0.6 0.6 0 255 0 2 −2 0.4 3

Abdominal CT Liver Parenchyma Segmentation …

223

area of the image extracted, then close morphological operation used to fill in holes and small gaps of the image.

3.4

Region of Interest Extraction Phase

Watershed is one of image segmentation techniques. It used to segment the image into several homogeneous regions with similar gray levels [11]. The input image treated as a topographic map. Three types of points are used to express a topographic interpretation points that belong to regional minimum “minimum”, points at which a drop of water will fall to a single minimum “catchment basin”, and lines that separate catchment basin “watershed line” [12]. The aim of watershed is to search for regions of high intensity gradients (watersheds) which divide neighbored local minima (basins) [11]. In this paper, watershed is used to final segment liver parenchyma from abdominal CT image. Also watershed has a great advantage that it is very fast as No seed is needed, it has some drawbacks. Some of these are very sensitive to noise and over segmentation problem. So in order to overcome these problems. An adaptive approach of watershed used.

3.5

Evaluation Phase

Three measurements are used to evaluate the performance of the presented approach. These measurements are Dice Coefficient, Correlation and True Positive. The true positive ratio measure calculated by dividing the number of true positive (which that is mean pixels that actually belong to liver region) by the total number of liver region pixels. Dice coefficient is statistical validation metric to evaluate spatial overlap accuracy between two binary images. It is commonly used in reporting the performance of segmentation results. Its values range between 0 and 1, where 0 indicate no overlap and 1 perfect agreement [17]. It is calculated using Eq. (3). Correlation is another measurement to indicate how strong relationship between two binary images. 1 indicates a strong positive relationship, −1 indicates a strong negative relationship and 0 indicates no relationship at all. It is calculated using Eq. (4). Diceðx, yÞ =

2 * j x ∩ yj jx + yj

∑ ∑ ðxðm, nÞ − xÞðyðm, nÞ − yÞ m n ffi Corrðx, yÞ = rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð∑ ∑ ðxðm, nÞ − xÞ2 Þð ∑ ∑ ðyðm, nÞ − yÞ2 Þ m n

where y is mean of y and x is mean of x.

m n

ð3Þ

ð4Þ

224

G.I. Sayed and A.E. Hassanien

4 Experimental Results and Discussion CT scanning is a diagnostic imaging procedure that uses X-rays in order to present cross-sectional images (slices) of the body. The proposed approach will be applied to a complex dataset. The dataset is divided into seven categories, depending on the tumor type: Benign (Cyst (CY), Hemangioma (HG), Hepatic Adenoma (HA), and Focal Nodular Hyperplasia (FNH)); or Malignant (hepatocellular carcinoma (HCC), Cholangiocarcinoma (CC), and Metastases (MS)). Each of these categories has more than 15-patients, each patient has more than one hundred slices, and more than one phase of CT scans (arterial, delayed, portal venous, non-contrast). The dataset includes a diagnosis report for each patient. All images are in JPEG format, selected from a DICOM file, and have dimensions of 630630, with horizontal and vertical resolution of 72 DPI, and bit depth of 24 bits [13]. The proposed approach was tested on 43 abdominal CT images from different patients. The accuracy of the proposed approach is measured using Correlation, Dice Coefficient and True Positive Ratio. The proposed CT image segmentation approach was programmed in MATLABR 2007 on a computer having Intel Core I3 and 2 GB of memory. The following figures will demonstrate the result of each step in the proposed approach. Figure 2 shows the results obtained from preprocessing phase and PSO. Fig. 2a shows the original CT image, Fig. 2b shows CT image after applying median filter in order to remove noise, Fig. 2a shows the first cluster image obtained from PSO, where Fig. 2b shows the second cluster image and Fig. 2c shows the third cluster image. Figure 3 shows post-processing phase results, Fig. 3a shows image after converted to binary image using ostu’ thresholding method, Fig. 3b shows the results obtained from applying opening morphology, Fig. 3c shows image after selecting the largest area in the image, Fig. 3d shows image after applying close morphology and Fig. 3e shows image after filling holes. Figure 4

Fig. 2 Preprocessing and PSO results a Original Image, b Image after applying median filter, c Cluster-1, d Cluster-2 and e Cluster-3

Abdominal CT Liver Parenchyma Segmentation …

225

Fig. 3 Post-processing phase results. a Binarized clustered image, b Image after applying Open Morphology, c Image after selecting largest region, d Image after applying Close Morphology, e Image after fillling holes

Fig. 4 Results of using modifed watershed. a CT image after enhancement, b Gradient Image, c Gradient image after normalization, d Image after remove local maxima regions, e Image after apply watershed and take maximum region

shows results of using modified watershed. Fig. 4a the modified watershed use CT image after enhancement in post-processing phase as input image, Fig. 4b shows gradient image calculated from the original image, Fig. 4c shows gradient image after normalization, Fig. 4d shows image after remove local maxima regions and Fig. 4e shows the final segmented liver image after applying watershed and take maximum region. Table 2 compares the results of the proposed approach with other previous works. As it can be seen, the proposed approach gives better results and it can extract liver in less than 4 s. Table 3 shows the importance of using watershed in the proposed approach in terms of Dice Coefficient, and Correlation for the used

226

G.I. Sayed and A.E. Hassanien

Table 2 Comparison with existing work on liver segmentation Authors

Year

Accuracy (%)

Jeongjin et al. [16] Ruchaneewan et al. [15] M. Abdallal. [14] Z. Abdallal. [6] M. Anter [13] Nidaa Aldeek [7] Proposed approach

2007 2007 2012 2012 2013 2014 2014

70 86 84 92 93 87 94

Table 3 Comparison between using watershed in proposed approach and without using it in terms of dice coefficient, correlation and true positive Using watershed Without using watershed

Dice (%)

Correlation (%)

True positive (%)

91.89 89.12

90.62 87.94

94.62 90.23

dataset. As we can see, the accuracy increase to almost 3 %. Figure 5 shows the best similarity indices for the used dataset of the proposed approach in terms of Dice Coefficient, and Correlation. Table 4 compares the results obtained from different levels in terms of Dice Coefficient, True Positive and Correlation. From this table we can prove that the optimal level number is equal to three. Table 5 shows the total CPU process time of the whole data set for different levels. From this table, we can see as the number of levels increasing the CPU process time is increasing too. Table 6 compares the

Fig. 5 The best similarity indices obtained for the used dataset of the proposed approach in terms of dice coefficient, and correlation

Abdominal CT Liver Parenchyma Segmentation …

227

Table 4 Comparison between the results obtained from different levels in terms of dice coefficient, true positive and correlation No. of levels

Dice (%)

Correlation (%)

True positive (%)

2 3 4 5

78.80 91.89 87.98 80.48

76.51 90.62 86.83 80.62

71.71 94.62 93.36 88.44

Table 5 CPU process time for different levels No. of levels

Total CPU process time in seconds

2 3 4 5

50.62 56.66 62.39 75.69

Table 6 Comparison between the proposed approach and other methods Active contour Global threshold Proposed approach

Dice (%)

Correlation (%)

True positive (%)

71.87 81.34 91.89

69.22 79.41 90.62

72.43 81.19 94.62

results from the proposed approach and other methods in terms of Dice Coefficient, Correlation and True Positive. From these results, the proposed approach gives better results than other methods.

5 Conclusion and Future Works We proposed an integrated approach based on using PSO and watershed algorithm for automatic extract liver from abdominal CT images. The experimental results show that the proposed approach gives better result and obtained over all accuracy about 94 % of good liver extraction. This results from proposed approach can help for further diagnosis and treatment planning. In the future work, we plan to increase the number of CT images to evaluate the performance of the proposed PSO and watershed algorithm. And test new versions of PSO.

228

G.I. Sayed and A.E. Hassanien

References 1. Aly, A., Deris, S., Zaki, N.: Research review for digital image segmentation techniques. Int. J. Comput. Sci. Inf. Technol. 3(5), 99–106 (2011) 2. Islam, S., Ahmed, M.: Implementation of image segmentation for natural images using clustering methods. Int. J. Emerg. Technol. Adv. Eng. 3(3), 175–180 (2013) 3. Jain, A., Duin, R., Mao, J.: Statistical pattern recognition: a review. IEEE Trans. Pattern Anal. Mach. Intell. 22(1), 4–37 (2000) 4. Kennedy, J., Eberhart, R.: Particle swarm optimization. Proc. IEEE Int. Conf. Neural Netw. 4, 1942–1948 (1995) 5. Ghamisi, P., Couceiro, M., Benediktsson, J., Ferreira, N.: An efficient method for segmentation of images based on fractional calculus and natural selection. Expert Syst. Appl. Elsevier 39(16), 12407–12417 (2012) 6. Abdalla, Z., Neveen, I.G., Aboul Ella, H., Hesham, A.H.: Level set-based CT liver image segmentation with watershed and artificial neural networks, IEEE, pp 96–102 (2012) 7. Aldeek, N., Alomari, R., Al-Zoubi, M., Hiary, H.: Liver segmentation from abdomen ct images with bayesian model. J. Theor. Appl. Inform. Technol. 60(3), 483–490 (2014) 8. Hammouche, K., Diaf, M., Siarry, P.: A comparative study of various metaheuristic techniques applied to the multilevel thresholding problem. Eng. Appl. Artif. Intell. 23, 676–688 (2010) 9. Sakthivel, N., Prabhu2, L.: Mean median filtering for impulsive noise removal. Int. J. Basics Appl. Sci. 2(4), 47–57 (2014) 10. Leavline, EJ., Singh, DAAG.: Salt and pepper noise detection and removal in gray scale images: an experimental analysis. Int. J. Signal Process. Image Process. Pattern Recognit. 6(5), 343–352 (2013) 11. Salman, N.: Image segmentation based on watershed and edge detection techniques. Int. Arab J. Inf. Tech. 3(4), 104–110 (2006) 12. Kaur, A., Aayushi.: Image segmentation using watershed transform. Int. J. Soft Comput. Eng. (IJSCE) 4(1), 5–8 (2014) 13. Anter, A., Azar, A., Hassanien, A., El-Bendary, N., El-Soud, M.: Automatic computer aided segmentation for liver and hepatic lesions using hybrid segmentations techniques. In: IEEE Proceedings of Federated Conference on Computer Science and Information Systems, pp.193–198 (2013) 14. Abdalla, M., Hesham, H., Neven, I.G., Aboul Ella, H., Gerald, S.: Evaluating the effects of image filters in CT liver CAD system. In: proceeding of IEEE-EMBS International Conference on Biomedical and Health Informatics, The Chinese University of Hong Kong, Hong Kong (2012) 15. Ruchaneewan, S., Daniela, S., Jacob, F.: A hybrid approach for liver segmentation, intelligent multimedia processing laboratory, 3d segmentation in the clinic: a grand challenge, pp 151–160 (2007) 16. Jeongjin, L., Namkug, K., Ho, L., Joon, B., Hyung, J., Yong, M., Yeong, S., Soo Hong, K.: Efficient liver segmentation using a level-set method with optimal detection of the initial liver boundary from level-set speed images, Elsevier, computer 1 methods and programs in biomedicine, vol. 88, pp. 26–38 (2007) 17. Dice, L.R.: Measures of the amount of ecologic association between species. Ecology 26, 297–302 (1945) 18. Chakraborty, S., Samanta, S., Biswas, D., Dey, N., Chaudhuri, S.S.: Particle swarm optimization based parameter optimization technique in medical information hiding. In: IEEE International Conference on Computational Intelligence and Computing Research (ICCIC) pp. 1–6 (2013) 19. Hu, W., Yen, G.G.: Adaptive multiobjective particle swarm optimization based on parallel cell coordinate system, IEEE Trans. Evol. Comput. 19(1), 1–18 (2013)

Copyright © 2024 DOKUMEN.SITE Inc.