BAMCN: Boundary-Assisted Multiscale Contextual Network for Polyp Segmentation

Colorectal cancer is still an important cause of death from cancer. Early polyp detection can play an important role in avoiding colorectal cancer. Colonoscopy has been the most common approach for detecting polyps; however, the results of the examination rely on the expertise of the operator. Computer-aided detection systems are useful for automatic polyp segmentation because they can detect some polyps which might be missed by clinical practitioners. Polyp segmentation is faced with challenges such as the lack of clear boundaries, low background-foreground contrast, and morphological heterogeneity. Although recent deep learning-based methods aim to address the challenges of colorectal polyp segmentation, they still struggle to achieve high accuracy and efficiency. This research introduces a novel encoder-decoder segmentation architecture, the Boundary-Assisted Multiscale Contextual Network for Polyp Segmentation (BAMCN). The proposed network aims to achieve high polyp segmentation accuracy without increasing computational complexity. The suggested architecture integrates the advantages of a pre-trained encoder, our innovative hybrid multiscale feature fusion block (HMFF), multitask learning, depthwise separable convolutions, channel attention, spatial attention, and deep supervision. The hybrid multiscale feature fusion block improves the feature representation by extracting and fusing the features at various scales. This helps the model learn the finer as well as the coarser aspects of the input data, thus improving the robustness of segmentation. Multitasking learning helps in extracting boundary features in order to localize the polyps accurately. Various polyp segmentation datasets have been employed in the assessment of the proposed architecture. The findings indicate that the proposed framework significantly improves polyp segmentation compared to many leading approaches.