文献速递神经技术与安全CT

CT自动评估颅内出血的GPT增强决策支持系统

Integrated CT pipeline for automatic intracranial hemorrhage evaluation with GPT-enhanced clinical decision support.
PubMed 文献 · Eur Radiol · 2026-01-26

本研究开发了一个深度学习流水线,用于颅内出血的CT自动检测、分割及定位,并利用GPT-4o大语言模型生成临床建议。模型在RSNA等数据集上训练,在CQ500外部测试集验证。结果显示检测AUC达0.96,分割Dice值0.71–0.93,临床决策评价较高。

要点

  • 该流水线可自动检测、分割并定位颅内出血,辅助急诊CT快速判读。
  • 临床决策模块基于已发表指南和试验生成建议,神经外科医生评分较高。
  • 模型在外部验证集上表现良好,体积估计与人工测量一致性高。

核心结论

  • 该深度学习流水线将影像分析与可操作的临床决策相结合,在急诊颅内出血评估中具有显著应用潜力。
  • 系统在颅内出血检测、分割和定位上均表现出较高精度,并生成了可靠的临床建议。

研究设计与样本

  • 分割模型基于HS数据集1226例扫描训练,用于勾画6种颅内出血亚型。
  • 脑积水和中线移位模型基于HS数据集中507例专门子集训练。
  • 出血和水肿位置配准到标准脑区以提高可解释性。
  • 使用CQ500外部数据集含491例患者作为外部验证和测试队列。

关键影像表现/征象

  • 出血和水肿位置被配准至标准脑区,可直观显示出血累及区域。
  • 系统输出包含脑积水和中线移位评估结果,可识别继发改变。

诊断标准与关键数值

  • 检测模型在测试集上AUC为0.96,95%置信区间0.94–0.98。
  • 分割模型Dice值范围0.71–0.93,对应95%置信区间为0.61–0.76至0.90–0.96。
  • 体积估计与人工测量一致性高,一致性相关系数CCC为0.820–0.996。
  • 脑实质出血定位在不同脑区的κ值为0.85–1.00,一致性较强。
  • 临床决策评价:一位神经外科医生对检查和治疗评分中位数分别为4分和5分,另一位均为5分。

临床应用与阅片要点

  • 该流水线可为急诊颅内出血CT检查提供快速自动检测和量化支持,减少诊断延迟。
  • 报告时可参考自动分割的体积和标准脑区定位结果,提升描述一致性和可重复性。
  • 系统生成的临床建议可作为辅助参考,但需结合神经外科医生的最终判断。

局限性

  • 研究仅采用CQ500单一外部数据集进行验证,未报告多中心前瞻性评估结果。
  • 临床决策评价仅由两名神经外科医生完成评分,评估者数量有限。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

英文原文(摘要)· 点击展开

OBJECTIVES: Intracranial hemorrhage (ICH) is a time-critical neurological emergency in which rapid CT-based assessment directly informs treatment decisions. This study aimed to develop an automated deep-learning pipeline to enhance ICH detection, segmentation, and localization, complemented by clinical decision-making support through a large language model.

MATERIALS AND METHODS: The detection model was trained on 21,784 labeled and 3528 unlabeled CT scans from the RSNA dataset using semi-supervised learning. The segmentation model was trained on 1226 scans from the HS dataset to delineate six ICH subtypes. Hydrocephalus and midline-shift models were trained on a dedicated 507-scan subset of the HS dataset. Hemorrhage and edema locations were registered to standard brain regions to improve interpretability. For evaluation, the CQ500 dataset (491 patients) was used as an external validation and test cohort. Clinical recommendations were generated using the GPT-4o Assistants API based on published guidelines and trials.

RESULTS: On the test set, detection achieved an AUC of 0.96 (95% CI: 0.94-0.98), and segmentation yielded Dice values ranging from 0.71 to 0.93 with corresponding 95% CIs from 0.61-0.76 to 0.90-0.96, while volume estimation showed high concordance (CCC 0.820-0.996). Intraparenchymal hemorrhage (IPH) localization demonstrated strong agreement with κ values of 0.85-1.00 across brain regions. Clinical decisions generated by the pipeline were highly rated, with one neurosurgeon assigning median scores of 4 and 5 for examination and treatment, and the other assigning 5 for both.

CONCLUSIONS: This deep learning pipeline combines imaging analysis with actionable clinical decisions, demonstrating significant potential as a valuable tool for emergency care.

KEY POINTS: Question Rapid and accurate identification of ICH on CT is critical for guiding treatment, yet remains difficult using standard emergency radiological evaluation. Findings The end-to-end artificial intelligence pipeline achieved high accuracy in ICH detection, segmentation, and localization, with strong concordance to manual measurements and reliable clinical recommendations. Clinical relevance By automating image analysis and clinical decision-making, the pipeline demonstrated significant potential to reduce diagnostic delays, improve treatment guidance, and enhance patient outcomes in emergency care settings.

来源:Eur Radiol,查看原文。版权归原作者所有。

引用

  1. Zhao, Yan, Wang, et al. Integrated CT pipeline for automatic intracranial hemorrhage evaluation with GPT-enhanced clinical decision support. Eur Radiol. 2026. PMID: 41586845. doi:10.1007/s00330-025-12316-3

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。