Random forests based WCE frames classification

Wireless Capsule Endoscopy is a commonly used diagnostic technique to explore intestinal regions which are difficult to reach with traditional endoscopy. The large number of images produced by this technology requires the use of computer-aided tools to select only meaningful frames to speed up the analysis time by the expert. This paper proposes a methodology to identify in an ensemble of WCE frames the images that clearly show the narrowing of the intestinal lumen. The proposed technique uses a custom set of Haar features extracted from the images. These are used for the growth of different binary decision trees. Each tree assigns a label. One image is eventually associated with the class that has the majority vote in the forest. Experiments conducted on real WCE images have proved the effectiveness of the proposal and are reported and discussed.