Automatic bleeding detection scheme based on spatio-temporal feature extraction from wireless capsule endoscopy videos

Abstract

Wireless capsule endoscopy (WCE) is the most advanced and non-invasive video technology to detect small intestine diseases, such as bleeding. Recently, automatic bleeding detection methods have received much attention by several researchers be- cause of its huge diagnostic demand. In this research, e cient bleeding detection schemes are developed to detect bleeding frames and zones in WCE video. In or- der to detect bleeding frames, both image-based and block-based feature extraction methods are investigated. In the image-based method, instead of using conventional color spaces, a composite color plane is introduced and various statistical features are computed in that plane. One major advantage of this method is its low com- putational burden. However, the detection performance strongly depends on the amount of bleeding zones. Next, a color histogram based feature extraction method is proposed for block-based analysis where block statistical features are utilized. Here the e ect of number of histogram bins, block size and amount of block over- lap on overall performance is investigated. In order to use the advantages of both block-based and region-based method, a cluster speci c feature extraction method is proposed, which introduces an unsupervised clustering step to segment the image into two classes prior to global feature extraction. It is found that instead of ex- tracting features from the entire image if features from each cluster along with the di erential cluster features are used signi cantly better detection performance can be achieved. For the purpose of classi cation, various classi ers have been tested, such as support vector machine, k-nearest neighbor and linear discriminant analysis. Once a bleeding image is detected in a WCE video, automatic marking of the bleed- ing zone is very much supportive for the reviewer to diagnose the diseases. Thus, in this research, based on the features extracted for frame classi cation, bleeding zone detection schemes are also developed. Finally, in order to continuously track bleed- ing frames in a WCE video, a post-processing algorithm is introduced considering the decisions made on neighboring frames. Extensive experimentation is carried out on several WCE videos and a very satisfactory performance in comparison to some of the recent methods is achieved in terms of accuracy, sensitivity, and speci city for both bleeding frame and zone detection.