基于AI的CTA工具检测前循环大血管闭塞的多中心验证
本研究为回顾性多中心研究,纳入4家美国及欧洲机构共379例急性缺血性卒中患者的CTA检查,验证Methinks CTA-LVO软件自动检测前循环大血管闭塞的性能。该AI工具敏感性为95.8%,特异性为88.6%,AUC为97.6%,平均通知时间为3.30分钟。结论支持其作为分诊辅助工具用于加速卒中救治流程。
要点
- AI软件对前循环大血管闭塞检测敏感性95.8%、特异性88.6%,可用于快速分诊。
- M1段闭塞敏感性最高(98.5%),M2段相对较低(89.7%),阅片时需注意复核M2区域。
- 平均通知时间3.30分钟,假阳性/假阴性可能与严重狭窄或诊断标准界定相关。
核心结论
- Methinks CTA-LVO软件对前循环大血管闭塞检测具有高准确性、快速通知和良好的跨机构泛化能力。
- 该工具适用于辅助急性卒中分诊流程,最终诊断仍需专家审核确认。
研究设计与样本
- 回顾性多中心研究,纳入美国及欧洲4家外部机构连续CTA扫描数据。
- 经质量评估后共纳入379例患者,其中大血管闭塞组142例,非大血管闭塞组237例。
- 参考标准由独立神经放射学专家审阅并经裁决确定。
关键影像表现/征象
- 错误分析显示部分假阳性和假阴性源于参考标准模糊、严重狭窄或超出研究定义的闭塞。
- AI性能在纳入机构间及不同CT设备供应商间保持稳健。
诊断标准与关键数值
- 总体检测敏感性为95.8%(95%CI:91.0-98.4%),特异性为88.6%(95%CI:84.0-92.4%),AUC为97.6%。
- 按闭塞亚型:M1段敏感性98.5%,M2段敏感性89.7%,颈内动脉闭塞敏感性97.1%。
- 平均通知时间为3.30分钟。
临床应用与阅片要点
- 该AI工具可作为快速分诊辅助手段,提高前循环大血管闭塞CTA筛查效率和诊断一致性。
- 对M2段闭塞或存在严重狭窄的病例,AI检测敏感性相对降低,需结合原始CTA图像复核。
- 最终影像诊断仍应在专家监督下完成,AI结果不应单独作为治疗决策依据。
局限性
- 研究为回顾性设计,未报告是否纳入连续入组时间窗口以外的偏倚控制细节。
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
BACKGROUND: Rapid identification of anterior circulation large vessel occlusions (LVOs) is critical for timely mechanical thrombectomy in acute ischemic stroke. Computed tomography angiography (CTA) interpretation can be challenging, particularly in settings without continuous subspecialty expertise. Artificial intelligence (AI) based decision support tools may improve workflow efficiency and diagnostic consistency. The purpose of this study is to evaluate the diagnostic performance, processing time, and generalizability of the Methinks CTA-LVO software for automated detection of anterior circulation LVOs.
METHODS: This retrospective multicenter study included consecutive CTA scans from four external institutions in the United States and Europe. After quality assessment, 379 patients were analyzed (142 LVO, 237 non-LVO). Ground truth was established by review with adjudication by independent expert neuroradiologists. Primary endpoints were sensitivity and specificity for LVO detection. Secondary analyses included performance by occlusion subtype (ICA, M1, M2), time-to-notification, false positive/negative characterization, and institution-level generalization.
RESULTS: The algorithm achieved a sensitivity of 95.8% (95% CI: 91.0-98.4%) and specificity of 88.6% (95% CI: 84.0-92.4%), with an AUC of 97.6%. Sensitivity by subtype was 98.5% for M1, 89.7% for M2, and 97.1% for ICA occlusions. The mean time-to-notification was 3.30 minutes. Error analysis showed that several apparent false positives and negatives reflected ground-truth ambiguity, severe stenosis, or occlusions outside the study definition. Performance remained robust across institutions and CT vendors.
CONCLUSION: The Methinks CTA-LVO software demonstrated high accuracy, rapid notification, and good generalizability for anterior circulation LVO detection, supporting its use as a triage tool to assist timely stroke care, with final interpretation remaining under expert supervision.
引用
- Cendrero, Tanzi, Hernández-Pérez, et al. Multicenter Validation of an AI-Based CTA Tool for Anterior Circulation Large Vessel Occlusion Detection. AJNR Am J Neuroradiol. 2026. PMID: 42580857. doi:10.3174/ajnr.A9575
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。