Diffusion-Guided Structure-Aware Segmentation for Medical Imaging

Medical image segmentation remains pivotal for clinical translation yet faces persistent challenges from ambiguous anatomical boundaries, heterogeneous tissue appearances, and pathological variability. While convolutional neural networks (CNNs) have advanced segmentation capabilities, conventional architectures struggle with structural coherence in complex imaging scenarios. This paper introduces a novel two-stage segmentation framework that synergizes structural prior learning with fine-grained feature reasoning to address these limitations. The first stage employs a conditional diffusion model for structure-preserving network initialization, capturing global anatomical dependencies through iterative denoising of corrupted inputs. The second stage enhances a U-Net backbone with two dedicated modules: (1) Semantic Information Decoupling (SID) at the deepest encoder layer to disentangle foreground/background/uncertainty embeddings, and (2) Contrast-Driven Feature Aggregation (CDFA) modules after each encoder block for contrast-aware feature refinement. A multi-scale uncertainty propagation strategy further injects uncertainty-aware features into decoder stages, enhancing boundary localization in ambiguous regions. Extensive experiments on multiple medical image segmentation benchmarks demonstrate that our method significantly improves the segmentation accuracy, especially for challenging cases with blurred boundaries and complex anatomical variations.