Scholarlink

<p>Medical image classification requires intelligent integration of both local details and global patterns. Most existing classification models are based on convolutional neural networks (CNNs), Transformers, or their hybrid variants. Although standard CNNs effectively capture fine-grained features such as edges and textures, they are less effective than Transformers in capturing global structures. However, Transformers are computationally expensive and limited spatial awareness. The emerging Mamba model offers efficient sequence processing, but its potential for medical imaging scenarios remains underutilized. To address these challenges, we propose CMFuse—a three-branch network that combines CNNs for local details, Mamba for global context, and an adaptive fusion block (CMF). The CMF block uses dynamic attention mechanisms to automatically balance local and global features while maintaining overall awareness of the lesion shape. Experiments on five medical datasets show that CMFuse achieves superior classification accuracy with lower computational complexity. Notably, it improves accuracy by 2.38% on PAD-UFES-20 and 1.89% on SMAD, demonstrating its robustness and potential in medical imaging.</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

CMFuse: a hierarchical feature fusion model combining convolution and Mamba for medical image classification

  • 期刊论文

  • 发表日期：

    2025年9月3日

查看全文
  • Xu Chen,
  • Xuesong Yin,
  • Qi Huang,
  • Ting Shu,
  • Jianhao Ding,
  • Yigang Wang

摘要

Medical image classification requires intelligent integration of both local details and global patterns. Most existing classification models are based on convolutional neural networks (CNNs), Transformers, or their hybrid variants. Although standard CNNs effectively capture fine-grained features such as edges and textures, they are less effective than Transformers in capturing global structures. However, Transformers are computationally expensive and limited spatial awareness. The emerging Mamba model offers efficient sequence processing, but its potential for medical imaging scenarios remains underutilized. To address these challenges, we propose CMFuse—a three-branch network that combines CNNs for local details, Mamba for global context, and an adaptive fusion block (CMF). The CMF block uses dynamic attention mechanisms to automatically balance local and global features while maintaining overall awareness of the lesion shape. Experiments on five medical datasets show that CMFuse achieves superior classification accuracy with lower computational complexity. Notably, it improves accuracy by 2.38% on PAD-UFES-20 and 1.89% on SMAD, demonstrating its robustness and potential in medical imaging.