An improved automatic system for aiding the detection of colon polyps using deep learning

Colorectal cancer is responsible for the most cancer deaths after lung cancer. It has been well-established that early detection and removal of polyps can prevent colorectal cancer. It is therefore essential that automated polyp detection has the highest sensitivity and precision possible in order to detect the most cases and prevent unnecessary treatment. We present a deep learning model based on YOLOv3 that was trained to detect polyps. Training made use of the 39308 images of 78 polyps and 393 completely healthy images from the SUN database. The model was subsequently validated using both the public CVC-clinic and ETIS-Larib datasets containing both standard defintion (SD) and high definition (HD) images. The per-image polyp detection sensitivity(precision) was calculated as 91.5(96.6)% and 86.5(94.2)% for the CVC-clinic and Etis-Larib datasets, respectively. These results represent the best-known performance in the validation datasets in comparison with the results of a recent review.