Polyp Detection in Colonoscopy Video

Colonoscopy is the only screening that can detect polyps in the colon that can mutate into colorectal cancer. The task of detecting polyps is a job done by the doctors themselves who perform the procedure. It often happens that polyps are not recognized for reasons such as fatigue of the doctor himself, preparation in the previous days for the procedure not done correctly by the patient, confusion between polyp or mucosa, probe lights, or water injected. The work in this thesis was carried out thanks to a grant from Linkverse s.r.l., a company based in Rome, Italy, which believes that Machine Learning can help physicians in the detection of polyps and thus can help them identify the most difficult ones. During the Ph.D. course, we tried to solve the problems related to this topic by using known Object Detection architectures. The first contribution is related to the specialization of a known object detection network through clustered features and fine-tuning steps. The second contribution concerns an attention mechanism integrated into an object detection network to focus on specific regions of each image used in the training phase. Finally, the last contribution is about a framework created for Data Augmentation by exploiting a known Inpainting network. This contribution is useful to provide other researchers with a more extensive and variable dataset with realistic data.