"Automated Image Analysis and Motion Tracking in Colonoscopy Videos"
Abstract
Optical colonoscopy is a medical procedure where a flexible probe containing a camera and a fiber optic light source is inserted through the rectum and advanced to the cecum/ileum. Colonoscopy is a critical medical examination used to inspect the mucosal surface and detect abnormalities in the colon. A broad range of disease findings is possible including structural abnormalities pre-cancerous and cancerous lesions, as well as acute and chronic inflammatory. The Mayo score for disease severity estimation in ulcerative colitis has been widely used to evaluate responses to therapies in clinical trials. Despite the common use of colonoscopy in both the diagnosis and longitudinal monitoring of disease, challenges exist regarding the interpretation and assessment of colonoscopy videos. The accuracy in human disease measurement is hampered by the high interobserver variation and poor disease feature localization. Besides, the lack of well-trained human reviewer to apply analysis in a standard manner also limit the use of colonoscopy video interpretation in routine practice.
In this work, we aim at overcoming these challenges by using the power of deep learning and computer vision. A novel computer-aided colonoscopy video analysis system will be presented. The proposed system is able to perform reproducible and human-level disease severity estimation for patients with ulcerative colitis. A combination of contextual analysis and camera motion tracking is performed to detect the frame-level location of disease findings in the colon.
