Scholarlink

<p>Medical image segmentation methods often struggle to strike a balance between capturing fine-grained local structures and modeling global semantic context, which often results in inadequate boundary detail capture and limited capacity to model long-range dependencies. To address these issues, we propose a Parallel multi-scale attention fusion encoder (PMSAFE), which employs a dual-branch architecture to jointly extract local and global features. PMSAFE incorporates a channel and edge-sensitive attention (CESA) module to enhance feature representations in key regions and along boundaries, significantly improving the discriminative power of the feature maps. In addition, a multi-scale fusion module (MSFM) is integrated to facilitate deep fusion of local and global features at various scales using convolutional kernels of different sizes, further enhancing the model’s representation capability. Extensive experiments on three widely used benchmark datasets validate the effectiveness of PMSAFE: it achieves a Dice score of 83.82% and an HD95 of 12.73mm on the Synapse dataset, 92.27% and 1.08mm on ACDC, and 87.76% and 5.68mm on AVT, demonstrating its outstanding segmentation performance.</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

Pmsafe: parallel multi-scale attention fusion encoder for medical image segmentation

  • 期刊论文

  • 发表日期：

    2025年6月3日

查看全文
  • Zhengyan Li,
  • Linxi Li,
  • Mingwei Tang,
  • Jie Hu

摘要

Medical image segmentation methods often struggle to strike a balance between capturing fine-grained local structures and modeling global semantic context, which often results in inadequate boundary detail capture and limited capacity to model long-range dependencies. To address these issues, we propose a Parallel multi-scale attention fusion encoder (PMSAFE), which employs a dual-branch architecture to jointly extract local and global features. PMSAFE incorporates a channel and edge-sensitive attention (CESA) module to enhance feature representations in key regions and along boundaries, significantly improving the discriminative power of the feature maps. In addition, a multi-scale fusion module (MSFM) is integrated to facilitate deep fusion of local and global features at various scales using convolutional kernels of different sizes, further enhancing the model’s representation capability. Extensive experiments on three widely used benchmark datasets validate the effectiveness of PMSAFE: it achieves a Dice score of 83.82% and an HD95 of 12.73mm on the Synapse dataset, 92.27% and 1.08mm on ACDC, and 87.76% and 5.68mm on AVT, demonstrating its outstanding segmentation performance.