SAFNet: an effective multiscale interaction method for polyp segmentation

Automated polyp segmentation is essential for early detection of colorectal cancer. However, it remains challenging due to low contrast with surrounding tissue, color similarity, and variations in size. In this study, we propose SAFNet, an effective multiscale interaction framework designed for polyp segmentation. SAFNet is built on a Transformer backbone. It includes three key modules: a self-multiscale fusion module (SMFM), an attention-driven feature fusion module (ADFFM), and a feature collection module (FCM). These modules enhance the discrimination and precise localization capabilities of polyps by working collaboratively, effectively extracting, fusing and refining multiscale features. Extensive experiments conducted on five public datasets (Kvasir-SEG, CVC-ClinicDB, CVC-300, CVC-ColonDB, and ETIS-LaribPolypDB) demonstrate that SAFNet outperforms the current state-of-the-art (SOTA) methods in all evaluation metrics. Additionally, the framework exhibits strong generalization capabilities, achieving mIoU scores of 78.40% and 75.87% on CVC-ColonDB and ETIS-LaribPolypDB, respectively. These results validate the robustness and clinical application potential of SAFNet in achieving precise polyp segmentation during colonoscopy.