Severity Classification of Ulcerative Colitis in Colonoscopy Videos by Learning from Confusion

Md Farhad Mokter, Azeez Idris, Jung-Hwan Oh, Wallapak Tavanapong, Piet C. de Groen

Abstract: Endoscopic measurement of ulcerative colitis (UC) severity is important since endoscopic disease severity may better predict future outcomes in UC than symptoms. However, it is difficult to evaluate the endoscopic severity of UC objectively because of the non-uniform nature of endoscopic features associated with UC, and large variations in their patterns. In this paper, we propose a method to classify UC severity in colonoscopy videos by learning from confusion. The similar looking frames from the colonoscopy videos generate similar features, and the Convolutional Neural Network (CNN) model trained using these similar features is confused. Therefore, it cannot provide accurate classification. By isolating these similar frames (features), we potentially reduce model confusion. We propose a new training strategy to isolate these similar frames (features), and a CNN based method for classifying UC severity in colonoscopy videos using the new training strategy. The experiments show that the proposed method for classifying UC severity increases classification effectiveness significantly.