Research on Edge Feature Separation Attention Network in Colon Polyp Segmentation

Colonoscopy is currently the most effective and widely recognized technique for detecting colorectal polyps, serving as an essential txool for early screening and prevention of colorectal cancer. Despite its efficacy, accurate polyp segmentation remains challenging due to the diverse sizes and complex shapes of polyps, as well as the ambiguity of their boundaries with the mucosa. Addressing this issue, our study introduces an Edge Feature Separation Attention Network designed to enhance performance in polyp segmentation tasks, inspired by strategies used in human object camouflage recognition. This approach, drawing on human strategies for recognizing camouflaged objects in images by analyzing differences between foreground and background, reveals object contours. The core strategy involves implementing a Multi-Scale Residual Feature Extractor (MRFE) to extract features at various scales, and employing a Separation Attention Mechanism to distinguish between foreground and background, thereby accurately identifying polyp regions and minimizing background noise. Additionally, this study explores the integration of multi-scale upsampling with transformer modules to capture global and local features while optimizing computational efficiency. It also employs a Multi-Scale Channel Attention Module (MS-CAM) to adjust the number of feature map channels, enhancing the model's ability to recognize boundaries. Verified across five datasets, including Kvasir-SEG, ClinicDB, ColonDB, CVC-300, and ETIS, our method has demonstrated significant improvements in detail capture and segmentation accuracy.