Scholarlink

<p>Precise localization and complete segmentation of entire lesions are crucial in medical image segmentation. However, existing methods focus on capturing spatial dimensional information while losing sight of effectively modeling of channel dependencies. The substantial semantic discrepancies among features at various levels in UNet-based image segmentation models poses a challenge for capturing edge features. To tackle these issues, a dual-branch encoder-based medical image segmentation model, named MCFNet, is proposed. In global feature extraction branch of the encoder, a channel attention-enhanced Mamba structure (CEMamba) designed employs channel attention and depth-wise convolution to capture multi-channel information, achieving a collaborative modeling of global spatial and channel dimensions. The local feature extraction branch adopts ResNet34 to extract local information. Furthermore, a Multi-level Features Aggregation Decoding module (MFAD) is proposed, where large kernel convolutions are employed to capture multi-scale spatial features, while the Dynamic Channel Attention module (DCA) is designed to achieve cross-stage features semantic alignment. Extensive experiments were conducted on four medical image segmentation datasets-ISIC2017, ISIC2018, Kvasir-SEG, and CVC-ClinicDB. Compared to H-vmunet, our method achieves significant improvements in mDice by 1.09, 0.69, 4.08 and 4.30 percentage points, respectively.</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

A Mamba-based channel attention-enhanced model for medical image segmentation

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Yazhi Liu,
  • Jiamin Li,
  • Yating Zhao,
  • Zhigang Liu,
  • Wei Li

摘要

Precise localization and complete segmentation of entire lesions are crucial in medical image segmentation. However, existing methods focus on capturing spatial dimensional information while losing sight of effectively modeling of channel dependencies. The substantial semantic discrepancies among features at various levels in UNet-based image segmentation models poses a challenge for capturing edge features. To tackle these issues, a dual-branch encoder-based medical image segmentation model, named MCFNet, is proposed. In global feature extraction branch of the encoder, a channel attention-enhanced Mamba structure (CEMamba) designed employs channel attention and depth-wise convolution to capture multi-channel information, achieving a collaborative modeling of global spatial and channel dimensions. The local feature extraction branch adopts ResNet34 to extract local information. Furthermore, a Multi-level Features Aggregation Decoding module (MFAD) is proposed, where large kernel convolutions are employed to capture multi-scale spatial features, while the Dynamic Channel Attention module (DCA) is designed to achieve cross-stage features semantic alignment. Extensive experiments were conducted on four medical image segmentation datasets-ISIC2017, ISIC2018, Kvasir-SEG, and CVC-ClinicDB. Compared to H-vmunet, our method achieves significant improvements in mDice by 1.09, 0.69, 4.08 and 4.30 percentage points, respectively.