Scholarlink

<p>Fast and accurate segmentation of intracranial hemorrhage (ICH) on non-contrast head CT (NCCT) is important for emergency assessment, hematoma volume estimation, and clinical decision support. However, tiny lesions, ambiguous boundaries, and complex backgrounds still lead to missed lesions and false positives. To address these challenges, we propose SAM2-ICHNet, a detection-guided cascaded framework for ICH segmentation. The method keeps the Hiera encoder of SAM2 as the backbone, introduces a lightweight Detail Adapter for tiny hemorrhagic foci, adopts an AG-BAE decoder for boundary recovery, and uses a YOLO branch only at test time to provide soft spatial priors and consistency correction. Experiments on an in-house clinical dataset and two public datasets, BCIHM and MBH, show that SAM2-ICHNet achieves Dice scores of 89.03%, 58.39%, and 55.58%, respectively, with HD95 values of 13.76, 64.80, and 65.40. The method performs best among the compared approaches, while the lower external-dataset scores indicate that cross-domain generalization remains challenging.</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

SAM2-ICHNet: A detection-guided intracranial hemorrhage segmentation framework for tiny lesions and complex backgrounds

  • 期刊论文

  • 发表日期：

    2026年6月25日

查看全文
  • Wanying Xie,
  • Ronghui Ju,
  • He Li,
  • Wei Guo,
  • Zhaoxuan Gong,
  • Guodong Zhang

摘要

Fast and accurate segmentation of intracranial hemorrhage (ICH) on non-contrast head CT (NCCT) is important for emergency assessment, hematoma volume estimation, and clinical decision support. However, tiny lesions, ambiguous boundaries, and complex backgrounds still lead to missed lesions and false positives. To address these challenges, we propose SAM2-ICHNet, a detection-guided cascaded framework for ICH segmentation. The method keeps the Hiera encoder of SAM2 as the backbone, introduces a lightweight Detail Adapter for tiny hemorrhagic foci, adopts an AG-BAE decoder for boundary recovery, and uses a YOLO branch only at test time to provide soft spatial priors and consistency correction. Experiments on an in-house clinical dataset and two public datasets, BCIHM and MBH, show that SAM2-ICHNet achieves Dice scores of 89.03%, 58.39%, and 55.58%, respectively, with HD95 values of 13.76, 64.80, and 65.40. The method performs best among the compared approaches, while the lower external-dataset scores indicate that cross-domain generalization remains challenging.