Scholarlink

<p>Medical image segmentation is vital for clinical diagnosis but remains challenging due to organ shape variability, ambiguous boundaries, and complex backgrounds; convolutional neural networks (CNNs) like U-Net capture local features well but struggle with long-range dependencies, whereas Vision Transformers (ViTs) model global context effectively but suffer from high computational costs and poor local detail preservation in high-resolution medical images. To address these issues, we propose ECCA-UNet, a Transformer-based segmentation model that integrates the Cross-Shaped Window (CSWin) mechanism for linear-complexity long-range dependency modeling, Squeeze-and-Excitation (SE) channel attention for enhanced feature discrimination, and an auxiliary edge-aware branch for precise boundary retention. The novelty lies in the synergistic combination of spatial-level and channel-level attention with explicit edge supervision, thereby enabling accurate delineation of fine anatomical structures while maintaining strong semantic consistency. Extensive experiments on two widely used public benchmarks demonstrate that ECCA-UNet achieves a Dice Similarity Coefficient of 81.90% and a Hausdorff Distance of 20.05 mm on the Synapse multi-organ CT dataset, as well as a Dice score of 91.10% on the ACDC cardiac MRI dataset, outperforming most state-of-the-art methods and highlighting its strong potential for challenging medical image segmentation tasks.</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

Ecca-unet: edge-aware and channel-enhanced cross-attention network for medical image segmentation

  • 期刊论文

  • 发表日期：

    2025年10月17日

查看全文
  • Tao Liu,
  • JianLong Hu,
  • MengYu Zhao

摘要

Medical image segmentation is vital for clinical diagnosis but remains challenging due to organ shape variability, ambiguous boundaries, and complex backgrounds; convolutional neural networks (CNNs) like U-Net capture local features well but struggle with long-range dependencies, whereas Vision Transformers (ViTs) model global context effectively but suffer from high computational costs and poor local detail preservation in high-resolution medical images. To address these issues, we propose ECCA-UNet, a Transformer-based segmentation model that integrates the Cross-Shaped Window (CSWin) mechanism for linear-complexity long-range dependency modeling, Squeeze-and-Excitation (SE) channel attention for enhanced feature discrimination, and an auxiliary edge-aware branch for precise boundary retention. The novelty lies in the synergistic combination of spatial-level and channel-level attention with explicit edge supervision, thereby enabling accurate delineation of fine anatomical structures while maintaining strong semantic consistency. Extensive experiments on two widely used public benchmarks demonstrate that ECCA-UNet achieves a Dice Similarity Coefficient of 81.90% and a Hausdorff Distance of 20.05 mm on the Synapse multi-organ CT dataset, as well as a Dice score of 91.10% on the ACDC cardiac MRI dataset, outperforming most state-of-the-art methods and highlighting its strong potential for challenging medical image segmentation tasks.