H2M-UNet: Hierarchical Memory Mamba-Driven UNet Collaborative Optimization Based on Long-Range Forgetting Mitigation...

In medical image segmentation, conventional state-space-model-based Mamba networks suffer from long-range forgetting due to locally dependent scanning mechanisms, which compromises prediction consistency and segmentation accuracy when processing long sequences. To enhance long-range dependency modeling, this study proposes a Hierarchical Memory Mamba-UNet (H2M-UNet) integrating a novel Hierarchical Memory 2D (HM2D) scanning mechanism. HM2D introduces a four-branch parallel scanning strategy with differentiated functional roles to achieve multi-scale feature co-optimization. Specifically, a global scanning path is designed for low-resolution features obtained from down-sampling to strengthen long-range dependency modeling, a local refinement path is deployed on high-resolution skip-connection features to capture fine details, and the conventional dual-path scanning is retained for original-resolution features to preserve basic representations. Extensive experiments on the Synapse multi-organ and ACDC cardiac segmentation datasets validate the effectiveness of the proposed method. On Synapse, H2M-UNet achieves an average Dice of 81.79% and HD95 of 14.00, outperforming nnUNet, Swin-UNet, and UD-Mamba by 1.76%-4.00% in Dice and reducing HD95 by 3.88-20.14. On ACDC, it attains an average Dice of 91.33% and HD95 of 1.12, improving Dice by 0.12%-0.70% and de-creasing HD95 by 0.05-0.46 compared with baseline models. These results demonstrate that integrating the HM2D mechanism effectively mitigates long-range forgetting and enhances fine-grained feature preservation, leading to superior segmentation accuracy and boundary consistency in complex medical imaging scenarios. Our code is available at: https://github.com/fangxiao-yu/HMMamba .