Scholarlink

<p>Robust segmentation of thin tubular structures remains challenging in medical image pattern analysis due to low contrast, weak distal branches, complex topology, and domain shifts. Although vision foundation models (VFMs) provide strong transferable representations, standard parameter-efficient adaptation mainly improves semantic transfer and provides limited explicit structural guidance for topology-sensitive vessel segmentation. To address this issue, we propose Skeleton-Aware SAM2 (SA-SAM2), a morphology-guided adaptation framework for angiographic vessel segmentation. The proposed method extracts a label-free distance-aware skeleton prior from the input image to provide continuous spatial proximity cues around vessel centerlines. To inject this prior into decoder features, we introduce a Gated SkeletonSPADE (GS-SPADE) module, which performs spatially adaptive feature modulation while using a learnable gate to suppress unreliable responses in the unsupervised prior. In addition, an auxiliary skeleton prediction head is used during training to further regularize topological consistency, while introducing no extra inference cost. Experiments on one source dataset (XCA-TJ) and two unseen target datasets (XCAD and XCAV) show that SA-SAM2 improves both overlap-based and topology-aware performance over strong baselines and topology-aware loss variants. The proposed framework also introduces only marginal computational overhead (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\sim \)</EquationSource> </InlineEquation>1.45 GFLOPs). These results suggest that explicit morphology-guided conditioning provides useful structural bias for topology-sensitive angiographic vessel segmentation under domain shifts.</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

Skeleton-aware SAM2 adaptation for topology-preserving angiographic vessel segmentation

  • 期刊论文

  • 发表日期：

    2026年7月28日

查看全文
  • Shuang Liang,
  • Chunyi Yang,
  • Zhicheng Liu,
  • Baihua Liu,
  • Haotian Sun,
  • Shengjie Zhao,
  • Peng Qi

摘要

Robust segmentation of thin tubular structures remains challenging in medical image pattern analysis due to low contrast, weak distal branches, complex topology, and domain shifts. Although vision foundation models (VFMs) provide strong transferable representations, standard parameter-efficient adaptation mainly improves semantic transfer and provides limited explicit structural guidance for topology-sensitive vessel segmentation. To address this issue, we propose Skeleton-Aware SAM2 (SA-SAM2), a morphology-guided adaptation framework for angiographic vessel segmentation. The proposed method extracts a label-free distance-aware skeleton prior from the input image to provide continuous spatial proximity cues around vessel centerlines. To inject this prior into decoder features, we introduce a Gated SkeletonSPADE (GS-SPADE) module, which performs spatially adaptive feature modulation while using a learnable gate to suppress unreliable responses in the unsupervised prior. In addition, an auxiliary skeleton prediction head is used during training to further regularize topological consistency, while introducing no extra inference cost. Experiments on one source dataset (XCA-TJ) and two unseen target datasets (XCAD and XCAV) show that SA-SAM2 improves both overlap-based and topology-aware performance over strong baselines and topology-aware loss variants. The proposed framework also introduces only marginal computational overhead ( \(\sim \) 1.45 GFLOPs). These results suggest that explicit morphology-guided conditioning provides useful structural bias for topology-sensitive angiographic vessel segmentation under domain shifts.