Abnormal image detection using Texton method in wireless capsule endoscopy videos

Abstract

One of the main goals of Wireless Capsule Endoscopy (WCE) is to detect the mucosal abnormalities such as blood, ulcer, polyp, and so on in the gastrointestinal tract. Only less than 5% of total 55,000 frames of a WCE video typically have abnormalities, so it is critical to develop a technique to automatically discriminate abnormal findings from normal ones. We introduce "Texton" method which has been successfully used for image texture classification in non-medical domains. A histogram of Textons (exemplar responses occurring after convolving an image with a set of filters called "Filter bank") called a "Texton Histogram" is used to represent an abnormal or a normal region. Then, a classifier (i.e., SVM or K-NN, and etc.) is trained using the Texton Histograms to distinguish images with abnormal regions from ones without them. Experimental results on our current data set show that the proposed method achieves promising performances. © 2010 Springer-Verlag.