Bleeding Detection In Wireless Capsule Endoscopy Based On Super-Pixel Segmentation

Wireless capsule endoscopy (WCE) is a novel gastrointestinal disease detection technology which has been widely used in many hospitals. It can overcome the defection of the traditional mechanical endoscopy and avoid physical pain to the patients. However, WCE produce fifty to sixty thousand images, which is a time consuming and tedious task for doctors to diagnosis only with human eyes. Therefore, it is the key issues to develop a set of computer aided detection software currently.In order to improve the detection accuracy of small bleeding area images, images are divided into blocks typically. However, it will lose the edge and semantic information when divide images into a fixed number of rectangular pieces in the conventional methods, and the detection accuracy is low. This thesis presents a detection algorithm based on super-pixel segmentation technology, which get homogenous regions together adaptively, it not only maintains the image boundaries, but also reduces the redundancy of the image.Firstly, this thesis analyzes and summarizes seven kinds of super-pixel segmentation algorithms from graph-based and gradient-based theories. By comparing performance of visual result, boundary recall, and run time, we select the best method to segment images.Secondly, due to the small size, irregular shape and poor texture information of super-pixel blocks, we use the color feature describe super-pixel blocks, considered the statistical characteristics which is unrelated to the shape. This thesis analysis color moment, color component features, red purity feature and the improved red purity feature, experiments show that the result of improved method is better than other methods. We trained the local feature classifier to identify the bleeding patient image data using the improved characteristic.Finally, to reduce false positive rate of WCE images, we cite cascade local and global features detection technology, extracting the Contourlet transformation and LBP texture features in different color spaces, and we choose the Contourlet transformation in HSV to extact the texture feature. The normal images which got on the first level classification, can be classified by using global features, and this function can reduce the false positive rate and improve the sensitivity effectively.