Scholarlink

<p>Accurate medical image segmentation is fundamental for diagnosis, treatment planning, and surgical guidance, yet existing deep learning methods face persistent challenges. Convolutional networks excel at extracting local details but struggle with global context, while Transformers capture long-range dependencies at the cost of efficiency and heavy data requirements. To address this gap, we propose MSMamba, a hybrid architecture that integrates convolutional operations with state space modeling, achieving both local precision and global contextual understanding. MSMamba incorporates a Wavelet Transform Stream (WTStream) for frequency-aware multi-scale representation and a Multi-Scanning Vision State Space (MSVSS) block that applies multi-directional scanning and merging to capture dependencies across spatial dimensions while preserving structural continuity. Extensive experiments on four representative benchmarks, including AbdomenMRI, Endoscopy, Microscopy, and BUSI datasets, demonstrate the superior performance of MSMamba compared with CNN, Transformer, and Mamba-based state-of-the-art methods. On AbdomenMRI, MSMamba achieves an average Dice score of 79.55% and NSD of 84.95%. For Endoscopy instrument segmentation, it obtains a Dice score of 56.50% and NSD of 58.73%, outperforming strong Mamba-based competitors. On Microscopy cell segmentation, MSMamba attains an F1-score of 59.35%, exceeding SwinUMamba by 2.38%. On the BUSI dataset, MSMamba delivers an mIoU of 71.42%. Ablation studies confirm the complementary contributions of WTStream and MSVSS, with multi-directional scanning consistently delivering performance gains. By combining local feature fidelity, global dependency modeling, and computational efficiency, MSMamba introduces a new paradigm for medical image segmentation with significant potential to advance clinical applications and inspire future research in state space-based architectures. Code and models are available at: <a href="https://github.com/30Liu/MSMamba">https://github.com/30Liu/MSMamba</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

MSMamba: enhancing medical image segmentation with a multi-scanning Mamba hybrid network

  • 期刊论文

  • 发表日期：

    2025年12月3日

查看全文
  • Ruoyun Liu,
  • Jianshu Chao,
  • Jiahua Lai,
  • Qingwei Guo,
  • Ke Sun,
  • Zeyu Zhang

摘要

Accurate medical image segmentation is fundamental for diagnosis, treatment planning, and surgical guidance, yet existing deep learning methods face persistent challenges. Convolutional networks excel at extracting local details but struggle with global context, while Transformers capture long-range dependencies at the cost of efficiency and heavy data requirements. To address this gap, we propose MSMamba, a hybrid architecture that integrates convolutional operations with state space modeling, achieving both local precision and global contextual understanding. MSMamba incorporates a Wavelet Transform Stream (WTStream) for frequency-aware multi-scale representation and a Multi-Scanning Vision State Space (MSVSS) block that applies multi-directional scanning and merging to capture dependencies across spatial dimensions while preserving structural continuity. Extensive experiments on four representative benchmarks, including AbdomenMRI, Endoscopy, Microscopy, and BUSI datasets, demonstrate the superior performance of MSMamba compared with CNN, Transformer, and Mamba-based state-of-the-art methods. On AbdomenMRI, MSMamba achieves an average Dice score of 79.55% and NSD of 84.95%. For Endoscopy instrument segmentation, it obtains a Dice score of 56.50% and NSD of 58.73%, outperforming strong Mamba-based competitors. On Microscopy cell segmentation, MSMamba attains an F1-score of 59.35%, exceeding SwinUMamba by 2.38%. On the BUSI dataset, MSMamba delivers an mIoU of 71.42%. Ablation studies confirm the complementary contributions of WTStream and MSVSS, with multi-directional scanning consistently delivering performance gains. By combining local feature fidelity, global dependency modeling, and computational efficiency, MSMamba introduces a new paradigm for medical image segmentation with significant potential to advance clinical applications and inspire future research in state space-based architectures. Code and models are available at: https://github.com/30Liu/MSMamba.