Abstract
Colorectal cancer (CRC) is the third leading cause of cancer deaths in the world, which mostly stems from precancerous polyps. Early detection and accurate classification of polyps play a vital role in colonoscopy. It makes sense to automatically detect the polyp and give a real-time classification feedback according to popular Yamada classification guidance during colonoscopy progress. We propose an interpretable deep neural network method, called multi-task real-time deep neural network with Shapley additive explanations, for polyp detection, polyp classification and polyp segmentation under colonoscopy. To the best of our knowledge, this is the first time to perform polyp classification according to Yamada classification guidance under colonoscopy with a deep learning method. To validate the performance of our proposed method, we conduct various comparative experiments on popular CVC-CLINIC and CVC-COLON datasets. We adopt various performance indicators, including area under receiver operating characteristics curve (AUC), precision, recall, F1 score, accuracy, and mean intersection over union (mIoU). The proposed method achieves satisfactory real-time performance in terms of polyp detection module, polyp classification module and polyp segmentation module. The experimental results show the overwhelming performance of our proposed method compared with other deep learning methods. We have achieved satisfying operating efficiency and interpretable feedback to meet the requirements of the colorectal surgeon, which provides an valuable decision support and reduces the rate of missed diagnosis and misdiagnosis of polyps in the process of colonoscopy. (c) 2021 Published by Elsevier B.V.