Dual pathway attention over attention network with multi-scale supervision for colonoscopy polyp segmentation

초록

Accurate polyp segmentation in colonoscopy images is essential for early colorectal cancer detection and preven tion. Existing methods struggle with the inherent variability in polyp appearance, scale, and texture. Moreover, these approaches often rely on single-stream encoders with uniform attention mechanisms that fail to capture both local and global features and lack specialized fusion strategies for multi-scale integration. To overcome these lim itations, we introduce a dual-stream architecture named A2SegNet that combines ConvNeXt Large and Pyramid Vision Transformer v2 (PVTv2) encoders to extract complementary local and global features. ConvNeXt processes the early stages to preserve fine spatial details, whereas PVTv2 handles the deeper stages for long-range context modeling. Furthermore, we introduce a scale-adaptive Strip-Pooling Coordinate Attention (SCOA) to enhance positional precision for boundary localization from the ConvNeXt stages. Multi-Headed Self-Attention (MHSA) is applied after the PVTv2 Stage 3 features to capture inter-region dependencies, and Efficient Channel Attention (ECA) in the final stage enhances feature discriminability before refinement via Optimized CBAM (OCBAM). Three fusion modules progressively integrate multi-scale features: the Detail-Preserving Fusion Module (DPFM) for low-level spatial accuracy, the Semantic Fusion Module (SFM) for mid-level structural coherence, and the Attentive Fusion Module (AFM) for high-level semantic integration. Multi-scale deep supervision at four reso lution levels ensures consistent learning across the feature hierarchy. Comprehensive experimental analysis and ablation studies are conducted to evaluate each module and the blocks used. Experiments across five benchmark datasets demonstrate that A2SegNet achieves state-of-the-art performance and handles challenging cases more effectively than existing approaches.

키워드