Comparative Evaluation of PCA-based Feature Transformation Techniques in Classification of Hepatic Focal Lesions with Ultrasound Images

Abstract

Dimensionality reduction is the most imperative stage to extract relevant information from the feature set used in a specific classification problem. Principal component analysis (PCA) is a widely accepted and frequently used dimensionality reduction technique that converts the set of correlated variables into set of uncorrelated variables. Various PCA-based techniques are available in the literature with different motives, but it is still unclear which one will perform well in the problem of detection and classification of focal hepatic lesions. Therefore, there is a need to perform a comparative evaluation to select the best PCA technique for dimensionality reduction of feature set in this specific problem. Consequently, in the present work a comparative evaluation has been performed with four PCA-based techniques, named as Fast PCA, Sparse PCA, Kernel PCA and Nonlinear fuzzy robust PCA. Among them, frequently applicable selected PCA techniques are (i) linear transformation techniques, viz., Fast PCA and Sparse PCA (ii) nonlinear transformation techniques, viz., Kernel PCA and Nonlinear fuzzy robust PCA. These PCA techniques have been applied on the 208 texture features to find out the best possible diagnostically important principal components in order to classify the five liver tissue categories [6]. Subsequently, these principal components are used to train the multi support vector machine classifier. The experimental results reveal that Sparse PCA outperforms the other PCA-based techniques showing the overall classification accuracy of 94%. Results also reveal that kernel PCA along with polynomial kernel outperforms the Fast PCA as it captures the high-order information from the feature space which is not possible by applying Fast PCA.