MSMA-Net: A Multi-scale Multidirectional Adaptation Network for Polyp Segmentation

Polyps are mostly noncancerous tumors that occur in several locations in the digestive tract but are most com-monly found in the colon. But over time, some colon polyps can develop into cancer, especially Adenomas that need to be detected to remove as soon as possible. In recent years, with the variety of modern techniques for polyps detection, image segmentation using deep learning has always been an appreciated method. However, polyp segmentation in this way also has some trouble such as long dependencies, complexity computation, poor local and global context, and lack of multi-scale context. There have been many researches and techniques proposed to overcome these problems, such as attention mechanisms, atrous spatial pyramid pooling (ASPP), Receptive Field Block(RFB), etc. Inspired by those advances in deep learning, in this work, we inherited and proposed an Efficient Attention Receptive Field Block (EA-RFB Block) and Local Global Fusion (LGF) that ensures the network's multi-scale representation, capturing enough information of both local and global context. Our proposed network, namely MSMA-Net has demonstrated improved performance through two metrics of Intersection over Union (IoU) and Dice Coefficient when compared with other state-of-the-art models in Polyp datasets.