深度学习与放射科医生对CT急性主动脉夹层检测的系统综述与网络荟萃分析
结果显示,在非增强CT检测中,深度学习汇总敏感度为0.91、特异度为0.90;在5项直接比较研究中,深度学习敏感度和特异度均高于实验条件下的放射科医生对照。研究结论支持深度学习作为辅助分诊工具而非替代医生,但证据确定性低,需前瞻性验证。
要点
- 非增强CT上深度学习检测急性主动脉夹层汇总敏感度0.91、特异度0.90,可用于分诊辅助。
- 与实验条件下放射科医生直接比较,深度学习敏感度和特异度更高,但不应替代医生。
- 真腔与假腔分割精度低于整体主动脉轮廓分割,阅片报告时需注意。
核心结论
- 深度学习在CT上检测急性主动脉夹层显示出有前景的准确性,在5项实验条件下头对头比较中敏感度和特异度均高于放射科医生对照,支持辅助分诊而非替代。
- 腔级分割仍比整体主动脉轮廓勾画更具挑战性。
- 证据确定性低,数据集以病例对照为主,中位患病率50%,异质性显著,发表偏倚明显,估计值为实验性上限,需前瞻性确认。
研究设计与样本
- 检测研究采用双变量随机效应模型汇总,头对头比较采用基于对比的网络荟萃分析,分割Dice系数采用三水平随机效应模型分析。
- 非增强CT检测纳入10项研究、18个队列、54个算法-队列数据集;直接比较纳入5项研究、11个比较;分割合成纳入16项研究、88个数据集。
关键影像表现/征象
- 非增强CT上深度学习检测急性主动脉夹层的汇总敏感度0.91(95%CI 0.87-0.94),特异度0.90(95%CI 0.85-0.94)。
- 放射科医生与深度学习相对敏感度为0.72(95%CI 0.57-0.91),相对特异度为0.88(95%CI 0.83-0.93),反映实验条件下医生对照检测能力低于深度学习。
- 分割汇总Dice系数为0.863(95%CI 0.795-0.931),真腔与假腔分割精度低于整体主动脉轮廓勾画。
诊断标准与关键数值
- 非增强CT检测汇总敏感度0.91(95%CI 0.87-0.94),特异度0.90(95%CI 0.85-0.94)。
- 放射科医生与深度学习直接比较,相对敏感度0.72(95%CI 0.57-0.91),相对特异度0.88(95%CI 0.83-0.93)。
- 分割汇总Dice系数0.863(95%CI 0.795-0.931),真腔与假腔分割低于整体主动脉勾画。
临床应用与阅片要点
- 深度学习可用于非增强CT急性主动脉夹层的辅助分诊,帮助优先处理疑似病例。
- 日常阅片时仍应以放射科医生诊断为主,深度学习结果作为快速筛查参考。
- 关注真腔与假腔分割误差较大的问题,报告中涉及真假腔范围时应结合人工判断。
局限性
- 纳入数据集以病例对照为主,中位患病率50%,可能高估模型实际应用表现。
- 异质性显著,小样本效应明显,证据确定性低。
- 直接比较研究中放射科医生在实验条件下读片,可能未反映临床真实工作流程中的表现。
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
Acute aortic dissection (AAD) is a time-sensitive cardiovascular emergency in which delayed recognition remains associated with poor early outcomes, and deep learning (DL) applied to computed tomography (CT) is increasingly proposed for triage support. We synthesized DL detection accuracy on CT, compared DL with radiologists head-to-head, and secondarily pooled CT segmentation accuracy. This Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA)-compliant review searched PubMed, Embase, and Web of Science through February 18, 2026. Detection studies were pooled with bivariate random-effects models; head-to-head comparisons were synthesized with contrast-based network meta-analysis; segmentation Sørensen-Dice coefficients were analyzed with three-level random-effects models. Twenty-three studies were included. For non-contrast CT detection (10 studies; 18 cohorts; 54 algorithm-cohort datasets), pooled sensitivity was 0.91 (95 % confidence interval [CI], 0.87-0.94) and specificity 0.90 (95 % CI, 0.85-0.94). In direct comparisons (5 studies; 11 comparisons), radiologist-to-DL relative sensitivity was 0.72 (95 % CI, 0.57-0.91) and relative specificity 0.88 (95 % CI, 0.83-0.93). For the secondary segmentation synthesis (16 studies; 88 datasets), pooled Dice was 0.863 (95 % CI, 0.795-0.931), lower for true- and false-lumen than for whole-aorta delineation. DL showed encouraging detection accuracy and higher sensitivity and specificity than the radiologist comparators in five head-to-head studies read under experimental conditions, supporting assistive triage rather than replacement; lumen-level segmentation remains more challenging than whole-aorta contouring. Certainty was low - datasets were predominantly case-control with a median prevalence of 50 %, heterogeneity was substantial and small-study effects pronounced - so these estimates are experimental upper bounds requiring prospective confirmation.
引用
- Wang, Hong, Liu, et al. Deep learning versus radiologists for acute aortic dissection on CT: A systematic review and network meta-analysis. Eur J Radiol. 2026. PMID: 42679415. doi:10.1016/j.ejrad.2026.113189
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。