Developing an efficient polyp segmentation using residual learning in asymmetric non-local neural networks

Colorectal cancer is among the leading causes of death by cancer worldwide. Thus, colorectal cancer needs early diagnosis to help in the management of the disease. Colonoscopy remains the main tool for polyp detection, yet the procedure sometimes fails because of the problem of vastly different structures and of very tiny polyps, rendering a possible overlook. To overcome the limitation, we propose a new image segmentation framework based on an Asymmetric Non-local Neural Network (ANN). Having considerably enhanced the accuracy of polyp detection in colonoscopy images, the ANN presents the best features of the non-local operation that considers subtle local features and long-range dependencies of an image for its segmentation and hence better polyp identifications. Upon evaluation of the framework over the Kvasir-SEG dataset, higher results were obtained with a Dice score of 0.947 and an average IoU (Intersection over Union) of 0.901. These metrics are beyond any known methods available today. It is believed that this breakthrough will provide major support for colorectal-cancer screening by reducing misdiagnoses and improving overall patient care, thereby ushering new possibilities for the early detection and effective treatment of this devastating disease.