Scholarlink

<p>Fine-grained segmentation of intracranial vasculature is critical for early stroke diagnosis, hemodynamic assessment, and surgical planning. However, many existing methods struggle to accurately capture thin vessels and maintain boundary continuity. We propose GDC-Net, a U-Net-based framework that emphasizes segmentation accuracy, especially for micro-vessels. GDC-Net integrates three novel modules: a global–local channel attention (GLCA) residual block, a dilated depthwise multi-head self-attention (DDMHSA) bridge, and a content-aware reassembly upsampling (CARU) decoder. GLCA fuses global context with local texture, heightening sensitivity to tiny perforating vessels. DDMHSA expands the receptive field to capture long-range vascular connectivity. CARU adaptively upsamples features to restore fine vessel edges. On two public cerebral MRA datasets (DMV and MIDAS), GDC-Net achieves substantial gains in Dice and recall (e.g., +2.3% mIoU and +1.5% Dice on DMV) over state-of-the-art models. Ablation studies confirm that each module contributes to precision and completeness: GLCA improves micro-vessel fidelity, DDMHSA enhances continuity, and CARU sharpens boundaries. Catering to large-scale high-resolution cerebral MRA processing and real-time clinical segmentation, GDC-Net features a supercomputing-oriented lightweight, parallelizable design, achieving 0.820 Dice and 97.2 ms inference latency on a 32GB vGPU with 30–50% fewer parameters and lower memory than peers while seamlessly scaling to HPC clusters for multi-center studies and intraoperative navigation, aligning with supercomputing’s core goals of resource optimization and scalable deployment. Our results demonstrate that GDC-Net achieves notable improvements in vessel segmentation accuracy and the detection of microstructures, supporting more reliable vascular analysis. Code: https://github.com/mayeer123/GDC-Net.git.</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

GDC-Net: a U-Net for precise brain vessel segmentation with global–local and depthwise attention plus content-aware upsampling

  • 期刊论文

  • 发表日期：

    2026年3月3日

查看全文
  • Yahui Tian,
  • Yibo Zheng,
  • Yongwei Liu,
  • Yuehan Wang

摘要

Fine-grained segmentation of intracranial vasculature is critical for early stroke diagnosis, hemodynamic assessment, and surgical planning. However, many existing methods struggle to accurately capture thin vessels and maintain boundary continuity. We propose GDC-Net, a U-Net-based framework that emphasizes segmentation accuracy, especially for micro-vessels. GDC-Net integrates three novel modules: a global–local channel attention (GLCA) residual block, a dilated depthwise multi-head self-attention (DDMHSA) bridge, and a content-aware reassembly upsampling (CARU) decoder. GLCA fuses global context with local texture, heightening sensitivity to tiny perforating vessels. DDMHSA expands the receptive field to capture long-range vascular connectivity. CARU adaptively upsamples features to restore fine vessel edges. On two public cerebral MRA datasets (DMV and MIDAS), GDC-Net achieves substantial gains in Dice and recall (e.g., +2.3% mIoU and +1.5% Dice on DMV) over state-of-the-art models. Ablation studies confirm that each module contributes to precision and completeness: GLCA improves micro-vessel fidelity, DDMHSA enhances continuity, and CARU sharpens boundaries. Catering to large-scale high-resolution cerebral MRA processing and real-time clinical segmentation, GDC-Net features a supercomputing-oriented lightweight, parallelizable design, achieving 0.820 Dice and 97.2 ms inference latency on a 32GB vGPU with 30–50% fewer parameters and lower memory than peers while seamlessly scaling to HPC clusters for multi-center studies and intraoperative navigation, aligning with supercomputing’s core goals of resource optimization and scalable deployment. Our results demonstrate that GDC-Net achieves notable improvements in vessel segmentation accuracy and the detection of microstructures, supporting more reliable vascular analysis. Code: https://github.com/mayeer123/GDC-Net.git.