Automated bleeding detection in wireless capsule endoscopy images based on sparse coding

Abstract

Wireless Capsule Endoscopy (WCE) is a revolutionary technique for screening of gastrointestinal (GI) tract. However, WCE needs automated methods to reduce the time required for viewing its large image data and also to improve the accuracy of inspection. In this work, we propose novel sparse coded features to detect bleeding in WCE images. We acquire Scale-Invariant Feature Transform (SIFT) based key points as regions of interest of an image. Further, we compute SIFT and uniform Local Binary Pattern (LBP) features around the key points from an image. After that sparse coded features are obtained and support vector machine (SVM) is used for image classification. We provide comprehensive experimental results and also comparison with some recent bleeding detection methods and with traditional ways of computing sparse coded features. The best results are obtained with a dictionary size of 300 atoms. The classification accuracy we achieve with the proposed approach is 98.18%. The results presented in the paper indicate that the proposed method is reliable for bleeding detection in WCE images.