Colorectal Polyp Detection and Comparative Evaluation Based on Deep Learning Approaches

Abstract

Colorectal cancer has been one of the leading causes of mortality over the past decade, and colorectal polyps are the leading cause of this disease. Conventional polyp detection techniques are insufficient for proper detection; thus, an efficient detection method is indispensable. In this study, we collected colorectal images from a hospital in Taiwan, annotated the ground truth of polyp locations, and integrated them with a public dataset to create a colonoscopy dataset. Data augmentation techniques are further used to increase the training dataset's diversity to improve the models' detection performance. By developing the comparison system based on the recent state-of-the-art methods (i.e., FasterRCNN, SSD, YOLOv3, and YOLOv4), we compared the measurement metrics and statistically analyzed the performance of the models to identify the significant statistical difference in models' performance. Moreover, we developed and integrated an error handling mechanism with each model to discard the false and null predictions. Finally, our model comparison system selects and proposes the best performing deep learning model to detect and classify colorectal polyps. We expect that the proposed model will accurately locate and classify different types of polyps. Eventually, this approach will ensure a valuable medical aid model.