Scheme And Dataset For Evaluating Computer-Aided Polyp Detection System In Colonoscopy

With the development of artificial intelligence, computer-aid polyp detection (CAD) systems are proposed as auxiliary detectors. However, CAD systems are generally evaluated from a computer vision perspective, rather than from a practical perspective to evaluate their performance in assisting physicians. One of the reasons is that clinical tests are expensive, and most of the existing available datasets are images, which cannot be used to evaluate the assistance performance. To this end, in this paper, we propose a new colorectal examination video dataset for evaluation, named ToPV, which contains 360 short videos collected during real colonoscopy examination. 204 of the videos contain one polyp and are labeled by experienced endoscopic physicians.Based on ToPV, we also present an experiment scheme for evaluating the assistance performance of CAD systems.The result of the initial experiment shows potential benefits of CAD systems in improving physicians' performance.