Morphology-Prior Guided Structural Boundary Refinement Network for Accurate Polyp Segmentation

Accurate polyp segmentation is essential for the early detection and treatment of colorectal cancer. However, this task remains challenging due to imaging artifacts, low contrast between polyps and surrounding tissues, and large appearance variations across datasets. Existing boundary-aware deep learning methods predominantly rely on purely data-driven convolutions or gradient-based operators. These approaches are highly susceptible to local pixel noise and inherently struggle to preserve geometric topology in low-contrast clinical scenarios. To overcome this structural vulnerability, we propose MGBR-Net, a morphology-prior guided structural boundary refinement network for polyp segmentation. Rather than relying on fragile intensity gradients, our framework shifts towards structured geometric modeling by explicitly injecting deterministic morphological priors to guide boundary-aware feature learning. Specifically, multi-scale contextual features are first extracted to capture diverse semantic representations. Based on these features, a Morphological Boundary Extraction Module (MBEM) leverages spatial topology via erosion and dilation operations to generate robust boundary structural priors, providing guidance for ambiguous regions. Furthermore, a Structural Prior Integration Module (SPIM) integrates boundary, interior, and surrounding contextual information under the guidance of structural priors. At the same time, a global and local interaction mechanism further improves contextual representation and structural consistency. Extensive experiments across five benchmark datasets demonstrate that MGBR-Net achieves leading average mDice and mIoU performance and remains competitive across complementary region- and boundary-oriented metrics compared with 14 representative methods. The source code and implementation details are publicly available at: https://github.com/zszuiniubi/MGBR-Net.