A Polyp Segmentation Method Combining Cross-Layer Attention and Boundary Refinement

Accurate segmentation of polyps provides an important basis for detecting and preventing colorectal cancer. However, due to the notable variations in the position, size, and number of polyps across different images, and certain images depict polyps with indistinct boundaries, a higher false negative rate was observed in existing algorithms for colorectal polyp segmentation. In this study, we propose a novel colorectal polyp segmentation algorithm that leverages cross-layer attention and boundary refinement techniques. Firstly, a multi-scale feature enhancement module targeting deep features has been designed to accurately locate polyp targets within deep feature maps. In addition, during the decoding phase, spatial attention is extracted from deep features and utilized to weight shallow features, allowing for a progressive fusion from deep to shallow layers, thereby refining the segmentation instances. Finally, the merged features undergo boundary refinement to produce the segmentation map. We performed comprehensive experiments on five widely utilized public datasets. Our algorithm demonstrates superior capabilities in reducing false negative in challenging polyp images compared to existing mainstream polyp segmentation networks.