Small Bowel Mucosa Segmentation for Frame Characterization in Videos of Endoscopic Capsules

Endoscopic capsules are vitamin-sized devices that leverage from a small wireless camera to create 8 to 10 hour videos of the patients' entire digestive tract, still being the leading tool to diagnose small bowel diseases. The revision of the produced videos is a very time-consuming task, currently conducted manually and frame-by-frame by an expert. Since endoscopic videos usually contain a considerable amount of frames where the mucosa is not clearly visible, the segmentation of the informative regions is a vital component to reduce the necessary time to review each exam. In this work, a CNN encoder-decoder architecture is applied to segment informative regions in small bowel frames of videos of endoscopic capsules. The network was trained and tested with a dataset of 2,929 manually annotated images, achieving a 91.2% Dice coefficient and 83.9% IoU. Furthermore, a video-wise analysis based on the amount of informative pixels in each frame is done.