An Investigation into Crohn’s Disease Lesions Variability Sensing Using Video Colonoscopy and Machine Learning Techniques

Abstract

Crohn's disease (CD) is a chronic inflammatory disease characterized by transmural inflammation and may affect any part of the gastrointestinal (GI) tract, from the mouth to the perianal area. Crohn's disease most commonly affects the colon and the last part of the small intestine (ileum). Crohn’s disease causes various lesions in the mucosa, which is the inner layer of the gut. CD presents with focal ulcerations, erythema and edema adjacent to areas of normal appearing mucosa resulting in heterogeneous patchy patterns of disease. Knowing the type and the extent of these patterns is important for the clinicians to provide the right treatments. The medical treatment aims at keeping the disease in remission and abating flares, whereas surgical treatment is indicated to address complications that are beyond the efficacy of the medical treatment. The videocolonoscopy is considered the gold standard in examining the colon and the terminal ileum and the video capsule endoscopy (VCE) to examine the entire small bowel. Examination of the video of both viewing procedures can be enhanced by using computer vision and machine learning techniques. In this paper, we have conducted our first investigation to cluster capsule endoscopy video frames from the small bowel into five CD clusters. We call our approach the CD lesion variability sensing as the uses pipeline of variability recognition utilizing thick data image augmentation techniques and deep learning that have the ability to learn such variability from few samples using Siamese neural network (SSN) with triple loss and fuzzy filter that uses structural similarity index (SSIM). The accuracy of our SSN with the triple loss function reached 68% and our added fuzzy filter increased it to reach over 75%. This is only the start of our investigation in this complex field.