Scholarlink

<p>Medical image segmentation plays a crucial role in clinical auxiliary diagnosis and intraoperative navigation. However, in scenarios with significant morphological variations, low contrast and strong noise, existing methods often fail to simultaneously balance long-range dependency modeling, boundary detail characterization, and cross-scale semantic consistency. To address these challenges, we propose a novel segmentation network, MDSF-Net, consisting of two key components: the Spatial-Mamba-Frequency Feature Block (SMFB) and the Hierarchical Semantic Link (HSL) module. The SMFB retains Mamba’s strength in capturing long-range dependencies while introducing the Dual-Domain Enhancement Module (DDEM) which combines the spatial and frequency information. This design alleviates the limitations of standard Mamba in capturing detailed structures of medical images. The HSL acts as a bridging structure at the bottleneck. It semantically reorganizes and aligns multi-scale features from different stages, enhancing the reliability of cross-layer information transmission. Experiments on three challenging medical image segmentation benchmarks, including Synapse, ACDC, and AMOS 2022, demonstrate that the proposed method outperforms current state-of-the-art (SOTA) segmentation methods across key quantitative metrics. The code is available at <a href="https://github.com/yjy2001230/MDSF-Net.">https://github.com/yjy2001230/MDSF-Net.</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

MDSF-Net: Mamba-driven spatial–spectral dual-stream fusion and hierarchical semantic linking for medical image segmentation

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Jiayi Yu,
  • Guangyuan Zhang,
  • Kefeng Li,
  • Dianxin Chen,
  • Zhenfang Zhu,
  • Yongpeng Zhao,
  • Guoying Pang,
  • Yufei Peng

摘要

Medical image segmentation plays a crucial role in clinical auxiliary diagnosis and intraoperative navigation. However, in scenarios with significant morphological variations, low contrast and strong noise, existing methods often fail to simultaneously balance long-range dependency modeling, boundary detail characterization, and cross-scale semantic consistency. To address these challenges, we propose a novel segmentation network, MDSF-Net, consisting of two key components: the Spatial-Mamba-Frequency Feature Block (SMFB) and the Hierarchical Semantic Link (HSL) module. The SMFB retains Mamba’s strength in capturing long-range dependencies while introducing the Dual-Domain Enhancement Module (DDEM) which combines the spatial and frequency information. This design alleviates the limitations of standard Mamba in capturing detailed structures of medical images. The HSL acts as a bridging structure at the bottleneck. It semantically reorganizes and aligns multi-scale features from different stages, enhancing the reliability of cross-layer information transmission. Experiments on three challenging medical image segmentation benchmarks, including Synapse, ACDC, and AMOS 2022, demonstrate that the proposed method outperforms current state-of-the-art (SOTA) segmentation methods across key quantitative metrics. The code is available at https://github.com/yjy2001230/MDSF-Net.