Automated data-driven on-line detection and characterization of endoscopic lesions using deep learning

The screening for colorectal cancer during a colonoscopy involves the detection of polyps and subsequently characterizing them in order to determine the required action. Based on the type of polyp, the endoscopist will decide whether to resect or not, whether histopathological analysis is needed or not and if future surveillance colonoscopies are needed. Due to the high polyp miss rate and difficult differentiation of polyp histology, an accurate automated system could greatly lower mortality rate and cost. The accuracy of current state-of-the-art methods is not yet satisfactory. Hence, the aim of this project is to create an automated system for detection and characterization of polyps that can be used online, in real-time, and that meets current clinical standards. Reaching this goal will be attempted by adapting pretrained, deep neural networks and by exploring the incorporation of application-specific and temporal information. The developed techniques will be evaluated on realistic clinical data in collaboration with the Department of Gastroenterology of UZ Leuven.