Scholarlink

<p>Medical image segmentation faces a critical challenge in effectively integrating global and local features. While traditional CNNs are constrained by limited receptive fields, Transformers suffer from high computational costs for global modeling. To address these problems, we propose CMambaFuse, a dual-branch network that synergizes CNN and Mamba architectures. State space models (SSMs), particularly the Mamba architecture, enable efficient long-range modeling with linear complexity. Inspired by Local Binary Pattern (LBP) operators, we develop the Circular Mamba (CM) module that converts 2D images into structurally coherent 1D sequences through circular scanning paths, enabling linear-complexity global context capture. Our RingStarFusion (RSFusion) module strategically integrates attention mechanisms with CM to achieve adaptive cross-scale feature fusion. Experimental results demonstrate that CMambaFuse achieves competitive performance on polyp datasets and a skin lesion segmentation dataset.</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

CMambaFuse: Circular mamba feature fusion for medical image segmentation

  • 期刊论文

  • 发表日期：

    2025年10月3日

查看全文
  • Xudong Wu,
  • Baohua Yuan,
  • Juxiao Zhang,
  • Lin Xiang,
  • Jin Ma,
  • Teng Fang,
  • Dehao Xiao,
  • Ning Li,
  • Gaochao Yang,
  • Lin Shi

摘要

Medical image segmentation faces a critical challenge in effectively integrating global and local features. While traditional CNNs are constrained by limited receptive fields, Transformers suffer from high computational costs for global modeling. To address these problems, we propose CMambaFuse, a dual-branch network that synergizes CNN and Mamba architectures. State space models (SSMs), particularly the Mamba architecture, enable efficient long-range modeling with linear complexity. Inspired by Local Binary Pattern (LBP) operators, we develop the Circular Mamba (CM) module that converts 2D images into structurally coherent 1D sequences through circular scanning paths, enabling linear-complexity global context capture. Our RingStarFusion (RSFusion) module strategically integrates attention mechanisms with CM to achieve adaptive cross-scale feature fusion. Experimental results demonstrate that CMambaFuse achieves competitive performance on polyp datasets and a skin lesion segmentation dataset.