Deep Learning in Colonoscopies: Improving Small Polyps Recognition Rate

Deep Learning in automatic video colonoscopy processing may result in missing small polyps or detecting them with low confidence. We conducted a study to demonstrate that we can overcome such problems by fitting the dimensions of the input image to the dimensions of the input layer of a trained Mobilenet Deep Learning network and by grouping polyps in classes according to their size. The results show that it is possible to substantially improve the confidence rates and we propose a scheme to increase small polyps detection in real-time video colonoscopies based on two trained Mobilenet networks.