A hybrid PSO-SFS-SBS algorithm in feature selection for liver cancer data

Abstract

Feature selection is an essential one in building high performance classification systems with the maximum classification accuracy. In this paper Particle Swarm Optimization (PSO) hybridized with Sequential Forward Selection (SFS) and Sequential Backward Selection (SBS) algorithm is proposed for improving the performance of the classification system. The feature subsets are extracted from the pattern under classification using First Order Statistics (FOS) combined with the Co-occurrence based features for different distance and degrees. Binary Particle Swarm Optimization (BPSO) is applied to the feature subset. After some iteration the 30% of the worst particles in PSO is replaced by the best feature subset of SFS and SBS algorithm. The proposed algorithm improves search ability and investigates two types of hybridization (1) PSO-SFS and (2) PSO-SFS-SBS with two options (1) velocity reset of all particles and (2) velocity reset of only worst particles. This hybrid system is applied to liver cancer data to reduce the features and to classify the liver disease as benign or malignant. Liver diseases like Hepatic Cellular Carcinoma (HCC), hemangioma, Focal Nodular Hyperplasia (FNH) and cholangiocarcinoma are classified. The Region of Interest (ROI) is cropped from an abdominal CT. The results obtained from different hybridized feature selection methods are examined. Experimental results show that the proposed methods select the 40% of features as best features to train the Probabilistic Neural Network (PNN) classifier with insignificant time to categorize the disease to give the accuracy of 96.4% for data set-1 and 92.6% for data set-II.