Scholarlink

<p>Accurate segmentation of complex biological structures, particularly coronary arteries, remains a significant challenge due to their intricate topologies and multi-scale morphological variations. In this paper, we propose GTUNet, a novel hybrid framework that synergistically integrates Graph Attention Networks, Transformers, and UNet for robust 3D medical image segmentation. The architecture features dual parallel encoder-decoder branches utilizing 3D ResUNet and Swin Transformer to capture multi-level representations. To encode geometric priors, a GAT-enhanced module is introduced specifically for the CNN branch to refine topological relationships before cross-stream interaction. Furthermore, a Bi-directional Attention Fusion Module is designed to bridge the two streams: it utilizes dual attention mechanisms to inject high-frequency local details from the CNN into the Transformer, while simultaneously transferring long-range dependencies from the Transformer back to the CNN. Extensive evaluations on the ASOCA and ImageCAS datasets demonstrate that GTUNet achieves state-of-the-art performance in Dice scores, ensuring superior topological continuity. Notably, the model significantly outperforms existing methods in ASSD, highlighting its precision in delineating vascular boundaries.</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

GTUNet: a GNN and transformer enhanced U-Net for coronary artery segmentation

  • 期刊论文

  • 发表日期：

    2026年7月28日

查看全文
  • Mengjun Sheng,
  • Shanshan He,
  • Zhiyu Shi

摘要

Accurate segmentation of complex biological structures, particularly coronary arteries, remains a significant challenge due to their intricate topologies and multi-scale morphological variations. In this paper, we propose GTUNet, a novel hybrid framework that synergistically integrates Graph Attention Networks, Transformers, and UNet for robust 3D medical image segmentation. The architecture features dual parallel encoder-decoder branches utilizing 3D ResUNet and Swin Transformer to capture multi-level representations. To encode geometric priors, a GAT-enhanced module is introduced specifically for the CNN branch to refine topological relationships before cross-stream interaction. Furthermore, a Bi-directional Attention Fusion Module is designed to bridge the two streams: it utilizes dual attention mechanisms to inject high-frequency local details from the CNN into the Transformer, while simultaneously transferring long-range dependencies from the Transformer back to the CNN. Extensive evaluations on the ASOCA and ImageCAS datasets demonstrate that GTUNet achieves state-of-the-art performance in Dice scores, ensuring superior topological continuity. Notably, the model significantly outperforms existing methods in ASSD, highlighting its precision in delineating vascular boundaries.