BCFNET: Boundary-Guided Semantic Cross Fusion for Polyp Segmentation

Colorectal cancer, ranked as the third most prevalent cancer globally, presents a significant health risk. The early detection of polyps through colonoscopy is pivotal for preventive intervention. However, existing methods have not fully utilized boundary information. In response, we propose BCFNet, a two-branch network that strategically incorporates polyp boundary information to guide semantic segmentation, thereby enhancing polyp detection accuracy. The BCFNet architecture comprises three key modules in each branch: a Multi-scale Feature Fusion (MFF) module, a Double Softmax Cross Guidance (DSCG) module, and a Semantic Cross Fusion (SCF) module. The MFF module adeptly merges multi-scaled semantic and boundary features. Concurrently, the DSCG and SCF modules synergistically enhance feature representations bidirectionally. They guide deep semantic polyp features to enrich shallow boundary features, while also leveraging deep boundary features to refine shallow semantic polyp features. Experimental results reveal that our methods as state-of-the-art compared to existing approaches.