Hierarchical spatial perception network and SAM-assisted uncertainty suppression for medical image segmentation.

Accurate medical image segmentation is critical for image-guided clinical procedures, yet the complexity and variability of anatomical structures make it highly challenging. To address this, we propose a novel segmentation framework that integrates a Hierarchical Spatial Perception network (HSP-Net) and a foundation model (SAM: Segment Anything Model)-assisted uncertainty suppression (SUS) strategy. HSP-Net enhances the classical U-Net architecture using a Group Pyramid Attention (GPA) module, which dynamically emphasizes discriminative features at multiple scales through a Channel Pooling Attention (CPA) block and a Spatial Hierarchical Attention (SHA) block. The SUS strategy leverages the SAM to identify hard-to-segment pixels and reduce prediction uncertainty. By combining semantic feature learning with foundation model-assisted refinement, our framework achieves promising performance across five publicly available datasets (i.e., BUSI, CVC, ISIC2017, ISIC2018, and Synapse). Extensive experiments show that it consistently outperforms U-Net and its multiple variants (e.g., UNeXt and UTNet), achieving an average Dice score of 0.8815 and a 95% Hausdorff distance of 7.3867, while demonstrating robust generalization across multimodal medical images. The source code is publicly available at https://github.com/wmuLei/hsp-net.