Multi-stage feature fusion network for polyp segmentation

Abstract

With the rising incidence and mortality of colorectal cancer, automatic polyp segmentation has gained significant attention. To address the limitations of existing pyramid-based transformer methods in polyp segmentation, specifically their challenges with feature scale diversity and feature fusion, we propose a transformer-based multi-stage feature fusion network (MSFFNet). First, the Contextual Dilation Fusion (CDF) module fuses adjacent multi-layer features and extracts multi-receptive field features, improving adaptability to polyps of different scales and enhancing feature diversity. Second, the Attention-Driven Feature Enhancement (AFE) module suppresses irrelevant background information and strengthens feature representation. Finally, the Dual-path Feature Fusion (DPF) module effectively integrates multi-level features using concatenation and point-wise addition. Extensive experiments on five datasets using four metrics demonstrate the effectiveness and strong generalization ability of the proposed method.