Scholarlink

<p>In recent years, Mamba, an innovative state-space model (SSM), has demonstrated significant potential in medical image segmentation by achieving linear computational complexity relative to sequence length. While Mamba efficiently captures long-range contextual dependencies, precise medical image segmentation requires not only global context but also the preservation of fine lesion details and boundary accuracy. Traditional CNN models are limited by their local receptive fields, whereas the Transformer architecture, despite enhancing global modeling, introduces artifacts at pixel boundaries due to its patch-based image tokenization. To address these limitations, we propose a Hybrid Multi-scale Spatial-State Mamba with Frequency-Adaptive Boundary Refinement (MSSMamba) for medical image segmentation, synergizing CNN-based local feature extraction with Mamba’s global context modeling capability. Specifically, we design a three-layer pyramid structure, where the CNN is responsible for extracting detailed local features, and Mamba captures the global context information effectively through state space modeling, avoiding the structural breakage problem caused by the transformer chunking operation. Furthermore, the framework effectively fuses local and global features and enhances the segmentation accuracy by designing a Multi-scale Attention Convolutional Block (MSACB) and a High-Low Frequency Adaptive Enhancement Module (HLAEM). Extensive experiments on the Synapse, ISIC17/18, and AVT datasets demonstrate that MSSMamba outperforms existing methods, particularly in segmenting small lesions and complex anatomical backgrounds. Our code is available at <a href="https://github.com/JSJ515-Group/MSSMamba">https://github.com/JSJ515-Group/MSSMamba</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

MSSMamba: hybrid multi-scale spatial-state mamba with frequency-adaptive boundary refinement for medical image segmentation

  • 期刊论文

  • 发表日期：

    2026年1月5日

查看全文
  • Xingzhu Liang,
  • Guoao Ge

摘要

In recent years, Mamba, an innovative state-space model (SSM), has demonstrated significant potential in medical image segmentation by achieving linear computational complexity relative to sequence length. While Mamba efficiently captures long-range contextual dependencies, precise medical image segmentation requires not only global context but also the preservation of fine lesion details and boundary accuracy. Traditional CNN models are limited by their local receptive fields, whereas the Transformer architecture, despite enhancing global modeling, introduces artifacts at pixel boundaries due to its patch-based image tokenization. To address these limitations, we propose a Hybrid Multi-scale Spatial-State Mamba with Frequency-Adaptive Boundary Refinement (MSSMamba) for medical image segmentation, synergizing CNN-based local feature extraction with Mamba’s global context modeling capability. Specifically, we design a three-layer pyramid structure, where the CNN is responsible for extracting detailed local features, and Mamba captures the global context information effectively through state space modeling, avoiding the structural breakage problem caused by the transformer chunking operation. Furthermore, the framework effectively fuses local and global features and enhances the segmentation accuracy by designing a Multi-scale Attention Convolutional Block (MSACB) and a High-Low Frequency Adaptive Enhancement Module (HLAEM). Extensive experiments on the Synapse, ISIC17/18, and AVT datasets demonstrate that MSSMamba outperforms existing methods, particularly in segmenting small lesions and complex anatomical backgrounds. Our code is available at https://github.com/JSJ515-Group/MSSMamba.