Comparative evaluation of feature extraction techniques for automatic disease detection in capsule endoscopy images

Abstract

A capsule endoscopy abnormal detection system is used to classify the abnormal images in wireless capsule endoscopy videos. Wireless Capsule Endoscopy is a relatively new technology allowing doctor to view most of the small intestine. The research has been attempted to automatically find abnormal regions to reduce doctors` time needed to analyze the videos. The proposed system uses multi-resolution features, which is compatible with various disease sizes of capsule endoscopy. ICA, PCA and NMF are applied to four different window size based approach to extract the part-based and holistic feature representation of the abnormal pattern. The classification of abnormal windows for each resolution is carried out by means of a SVM classifier, and Single-Layer Perceptron combines the 4 SVM results. After that, feature extraction method with maximum performance is selected and frame based decision is brought out. The ROC curves are used to exhibit the true positive rate (sensitivity) versus false positive rate (1-specificity) of the classfier. The proposed algorithm shows high accuracy of abnormality detection with signi_cant time reduction to inspect videos of WCE with a minimal loss of performance.