8 - Hybrid feature selection-based feature fusion for liver disease classification on ultrasound images

Abstract

This paper presents a hybrid feature selection (HFS)-based feature fusion system that selects the best features among multiple feature sets to classify liver ultrasound images into four classes: normal, chronic, cirrhosis, and heptocellular carcinomas evolved over cirrhosis. After extracting features by gray-level co-occurrence matrices, gray-level difference matrix and ranklet transform, the system utilizes HFS to select features. Here, HFS method is proposed by combining filter (ReliefF) and wrapper [sequential forward selection (SFS)] methods. Firstly, ReliefF method rank features and preselection are done by discarding low ranked features. Secondly, SFS method finds the optimal feature set. The advantage of proposed method is to make feature selection faster since filter method rapidly reduces the effective number of features under consideration. Thereafter, to take advantage of complementary information from different feature sets, feature fusion schemes are implemented: serial feature combination, serial feature fusion and, hierarchical feature fusion. Experiments are conducted to evaluate the (1) effectiveness of extracted features and proposed HFS method, (2) effectiveness of feature fusion schemes, and (3) performance based on the number of selected features, computational time and accuracy of ReliefF, SFS, sequential backward selection, and proposed method. Finally, the HFS-based hierarchical fusion set obtained accuracy of 95.2% with k-nearest neighbor.