Scholarlink

<p>Aortic dissection is a life-threatening emergency, where automated segmentation faces challenges such as indistinct boundaries. In contrast to generic hybrid architectures like TransUNet, which often struggle with computational burden and fine boundary details, or pure Transformer models like Swin-UNet that may underutilize local multi-scale features, HVPUNet presents a co-designed framework targeting the specific challenges in vascular structure segmentation. Its core contributions include: (1) the SegCSFormer Block, using Cross-Shaped Window Attention for efficient global–local interaction; (2) the Adaptive Multi-Kernel Block to enhance multi-scale feature extraction; and (3) the CA-SaE Block, combining coordinate attention with multi-branch excitation for improved feature fusion. Trained with high-performance computing on multiple GPUs to handle computational demands, HVPUNet achieved Dice coefficient improvements of 3.93% (true lumen) and 7.77% (false lumen) on an aortic dissection dataset. Experimental results demonstrate that the proposed HVPUNet offers a promising approach for supporting clinical diagnosis and surgical planning in aortic dissection.</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

HVPUNet: an automated deep learning model for precise true–false lumen segmentation in aortic dissection

  • 期刊论文

  • 发表日期：

    2026年1月29日

查看全文
  • Xiaojie Duan,
  • Yaqi Lei,
  • Jianming Wang,
  • Yanchao Wang

摘要

Aortic dissection is a life-threatening emergency, where automated segmentation faces challenges such as indistinct boundaries. In contrast to generic hybrid architectures like TransUNet, which often struggle with computational burden and fine boundary details, or pure Transformer models like Swin-UNet that may underutilize local multi-scale features, HVPUNet presents a co-designed framework targeting the specific challenges in vascular structure segmentation. Its core contributions include: (1) the SegCSFormer Block, using Cross-Shaped Window Attention for efficient global–local interaction; (2) the Adaptive Multi-Kernel Block to enhance multi-scale feature extraction; and (3) the CA-SaE Block, combining coordinate attention with multi-branch excitation for improved feature fusion. Trained with high-performance computing on multiple GPUs to handle computational demands, HVPUNet achieved Dice coefficient improvements of 3.93% (true lumen) and 7.77% (false lumen) on an aortic dissection dataset. Experimental results demonstrate that the proposed HVPUNet offers a promising approach for supporting clinical diagnosis and surgical planning in aortic dissection.