BP neural network classification for bleeding detection in wireless capsule endoscopy.

Bleeding in the digestive tract is one of the most common gastrointestinal tract (GI) diseases, as well as the complication of some fatal diseases. Wireless capsule endoscopy (WCE) allows physicians to noninvasively examine the entire GI tract. However it is very laborious and time-consuming to inspect large numbers of WCE images, which limits the wider application of WCE. It is therefore important to develop an automatic and intelligent computer-aided bleeding detection technique. In this paper, a new method aimed at bleeding detection in WCE images is proposed. Colour texture features distinguishing the bleeding regions from non-bleeding regions are extracted in RGB and HSI colour spaces; then a neural network using the colour texture features as the feature vector inputs is designed to recognize the bleeding regions. The experiments demonstrate that the bleeding regions can be correctly recognized and clearly marked out. The sensitivity of the algorithm is 93% and the specificity is 96%.