Scholarlink

<p>While vision transformers have advanced medical image segmentation by modeling long-range dependencies, their patch-based tokenization inevitably compromises low-level spatial resolution. Consequently, deep layers in hybrid architectures often suffer from feature discontinuity, leading to structural degra- dation and blurred boundaries. We propose MMRSG-UNet, a hybrid architecture that integrates a CSWin Transformer encoder with two specialized mechanisms to reconcile this trade-off. First, we introduce a symmetric multi-scale state space module (MS-SSM) to address feature discontinuity in the bottleneck. By coupling Mamba’s selective scan mechanism with multi-scale depth-wise convolutions, MS-SSM enables direction-aware contextual aggregation beyond discrete patch partitioning, thereby helping preserve structural integrity in com- pressed latent spaces. Second, we design a channel-spatial gated attention (CSGAT) module to mitigate the feature over-localization and geometric degra- dation of fine anatomical regions induced by combinatorial deep supervision. Distinct from conventional deep-to-shallow guidance, CSGAT employs a reverse semantic gating strategy, in which refined shallow structural details are used to calibrate upsampled deep responses, achieving a better balance between anatomical coverage and boundary fidelity. Experiments on the synapse and ACDC datasets demonstrate that MMRSG-UNet achieves competitive perfor- mance in overlap accuracy, boundary quality, and parameter efficiency. Code: <a href="https://github.com/VelvetDarkChocolate/MMRSG-UNet/tree/main">https://github.com/VelvetDarkChocolate/MMRSG-UNet/tree/main</a></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Mmrsg-unet: integrating multi-scale Mamba and reverse semantic gating for medical image segmentation

  • 期刊论文

  • 发表日期：

    2026年7月11日

查看全文
  • Chuanhao Yang,
  • Tianyu Zhang,
  • Shijia Ge,
  • Yanfeng Cao,
  • Dunke Lu

摘要

While vision transformers have advanced medical image segmentation by modeling long-range dependencies, their patch-based tokenization inevitably compromises low-level spatial resolution. Consequently, deep layers in hybrid architectures often suffer from feature discontinuity, leading to structural degra- dation and blurred boundaries. We propose MMRSG-UNet, a hybrid architecture that integrates a CSWin Transformer encoder with two specialized mechanisms to reconcile this trade-off. First, we introduce a symmetric multi-scale state space module (MS-SSM) to address feature discontinuity in the bottleneck. By coupling Mamba’s selective scan mechanism with multi-scale depth-wise convolutions, MS-SSM enables direction-aware contextual aggregation beyond discrete patch partitioning, thereby helping preserve structural integrity in com- pressed latent spaces. Second, we design a channel-spatial gated attention (CSGAT) module to mitigate the feature over-localization and geometric degra- dation of fine anatomical regions induced by combinatorial deep supervision. Distinct from conventional deep-to-shallow guidance, CSGAT employs a reverse semantic gating strategy, in which refined shallow structural details are used to calibrate upsampled deep responses, achieving a better balance between anatomical coverage and boundary fidelity. Experiments on the synapse and ACDC datasets demonstrate that MMRSG-UNet achieves competitive perfor- mance in overlap accuracy, boundary quality, and parameter efficiency. Code: https://github.com/VelvetDarkChocolate/MMRSG-UNet/tree/main