Investigation into the Categorization of Crohn's Segments from Capsule Endoscopy Videos:: Introducing A Thick Data Categorization Framework

Crohn's disease (CD) is a complex disorder that causes patchy inflammation and ulcerative effects in the gut from mouth to anus, which requires diagnosis for proper treatment. Capsule endoscopy (CE) procedure emerged as important investigation for screening the small bowel and the entire colon. However, examining the CE videos by gastroenterologists is a tedious and time consuming task as the capsule camera takes more than eighty thousand frames, with a frequency rate of 2 frames per second, for one segment of the gastrointestinal tract. However, having an automatic way to categorize the patchy frames that hold the markers of Crohn's disease will help the examiners to provide more reliable diagnosis and treatment. In this article, we are describing a thick data analysis framework that can help categorizing Crohn's disease sequences through the use of triplet-loss Siamese neural network that can learn from few shots provided by the expert to detect frames with markers of the Crohn's disease as well as to eliminate those frames having reduced mucosal view. Additionally, the framework uses a fuzzy filler to produce intermediate sets of frames having similar markers. These sets can be annotated and returned back to their position at the original CE video. The use of our framework showed promising results in re-orienting the original CE video into marked sequences of Crohn's disease frames. We have trained and tested our framework using two notable CE datasets (KVASIR Capsule and CrohnIPI) that contain the expert annotation on related markers of Crohn's disease.