Adaboost learning for small ulcer detection from Wireless Capsule Endoscopy (WCE) images

Abstract

Wireless Capsule Endoscopy (WCE) is getting popular as a non-invasive procedure to view the gastrointestinal tract. Many efforts have been devoted to computer-based bleeding or ulcer detection in WCE images. However, none of them has focused on the small ulcer detection in small bowel. Small ulcers are the small obscure light spots with similar colors of normal tissues as small intestine. During the 1-hour reading time of image frames, i.e. at the speed of 12~15 frame per second, the small ulcers are usually missed in human reading. In this paper, we present a novel approach using AdaBoost learning for small ulcer detection. This approach exploits simple RGB values as feature vectors and does not require any sophisticated routines for extracting high-level features. First, a set of weak classifiers is constructed by using weighted least square regression and AdaBoost learning is utilized to fuse the ensemble of these weak classifiers to a strong classifier for detection. Experiments on real WCE images have shown it can achieve over 80% of accuracy and is very promising in diagnosis applications.