Improving an SVM-based Liver Segmentation Strategy by the F-score Feature Selection Method Y. Xu 1 , J. Liu 1 , Q.M. Hu 1 , Z.J. Chen 1 , X.H. Du 1 , P.A. Heng 2 1 Human-Computer Interaction Research Center, Shenzhen Institute of Advanced Integration Technology, Chinese Academy of Sciences, Shenzhen, China 2 The Chinese University of Hong Kong, Shenzhen, China Abstract— A fast and accurate computer-aided liver seg- mentation plays a vital role in the virtual hepatic surgery. Large amount of features yielded in supervised segmentation methods may lead to slow training and classifying processes. Therefore, feature selection is of importance in order to speed up the liver segmentation. Recently, a hybrid method was proposed by Liu et al. combining thresholding, classifier and region growing. However, this method suffers from long process time caused by the large amount of features. F-score is a simple technique to measure the discrimination of different features. We therefore combine F-score to the hybrid method to reduce the time required in the training and testing stage. Four sets of abdominal CT images were obtained from Shan Dong University. The data consists of multiple, serial, axial computed tomography images derived from helical, 64 multi- slice CT and was stored in DICOM format of size 512 by 512 with 12-bit gray level resolution. The hybrid method which we proposed is to segment CT images by support vector machines after supervised thresholding, K means clustering, and texture feature extraction (Gray level co-occurrence Matrix-GLCM). We applied principle component analysis (PCA), forward orthogonal search algorithm by maximizing the overall depen- dency (FOS-MOD) and F-score to select the features from the GLCM. The experiment showed that F-score helps in accelerating training and classifying stage by 50% whilst the PCA-based feature selection method failed to extract the liver contour correctly. This may be explained by the fact that useful infor- mation for classifying may be lost when using PCA. FOS-MOD algorithm is time consuming mainly because its orthogonaliza- tion procedure and the calculation of the correlation matrix are very complex. In conclusion, F-score is a promising feature selection method for the svm-based classification. Our hybrid method with F-score can speed up the segmentation with accu- rate results ensured. Keywords— GLCM, F-score, FOS-MOD, PCA I. INTRODUCTION Advancement in radiological image segmentation is cru- cial in the virtual hepatic surgery. Only precisely segmenta- tion of parenchyma organs such as liver and kidney with computed tomography (CT) or other medical images makes planning of operations possible [1]. However, several prob- lems have to be solved during the segmentation process: a) high shape variation due to natural anatomical variation, b) inhomogeneous grey-value appearance caused by tumor, c) low contrast to neighboring organs like colon [2]. For prac- tical reason, a segmentation scheme has to be capable of handling these problems in a time-efficient manner. Recently, a hybrid method was proposed by Liu et al. combing thresholding, SVM and region growing [3]. Al- though this method provides an accurate solution of the segmentation, it suffers from long processing time yielded from the SVM for the large amount of features. In order to accurate the segmentation process, feature selection is of importance. Feature selection refers to finding the minimally sized feature subset that is necessary and sufficient to the target concept [4]. The aim of feature selection is to choose a subset of features for improving prediction accuracy or decreasing the size of the structure without significantly decreasing prediction accuracy of the classifier built using only the selected features. There are four basic steps in a feature selection method [5]: 1. a generation procedure to generate the candidate subset; 2. an evaluation function to evaluate the subset; 3. a stopping criterion; 4. a validation procedure to check the final subset. Feature methods can be categorized by the two main steps: generation produce and evaluation function. PCA is a classical dimensionality reduction method that has been used in many fields including medical image processing and pattern recognition. FOS-MOD proposed by Wei et al. [6] is a novel heuristic method based on a ranking criterion. F-score proposed by Chen et al [7] is a simple algorithm to measure the discrimination of two sets of real numbers. It is advantageous to limit the number of input features in a support vector machine to have a good predic- tive. We hence applied these algorithms in our supervised method to compare their results in order to find an appropri- ate solution to speed up the segmentation process. The rest of the paper is organized as follows. Section 2 summarizes some feature selection methods. Section 3 presents the experiment condition and the final result with a O. Dössel and . (Eds.): WC 2009, IFMBE Proceedings 25/IV, pp. www.springerlink.com W C. Schlegel 13–16, 2009.