Bag-of-Visual-Words Approach to Abnormal Image Detection in Wireless Capsule Endoscopy Videos

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 "Bag-of-Visual-Words" method which has been successfully used in particular for image classification in non-medical domains. Initially the training image patches are represented by color and texture features, and then the bag of words model is constructed by K-means clustering algorithm. Subsequently the document is represented as the histogram of the visual words which is the feature vector of the image. Finally, a SVM classifier is trained using these feature vectors 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.