The George Washington University School of Medicine and Health Sciences, 2150 Pennsylvania Ave NW, Ste 4-417, Washington, DC 20037.
Jefferson Medical Institute, Philadelphia, PA.
AJR Am J Roentgenol. 2022 Sep;219(3):407-419. doi: 10.2214/AJR.21.27289. Epub 2022 Apr 20.
Deep learning frameworks have been applied to interpretation of coronary CTA performed for coronary artery disease (CAD) evaluation. The purpose of our study was to compare the diagnostic performance of myocardial perfusion imaging (MPI) and coronary CTA with artificial intelligence quantitative CT (AI-QCT) interpretation for detection of obstructive CAD on invasive angiography and to assess the downstream impact of including coronary CTA with AI-QCT in diagnostic algorithms. This study entailed a retrospective post hoc analysis of the derivation cohort of the prospective 23-center Computed Tomographic Evaluation of Atherosclerotic Determinants of Myocardial Ischemia (CREDENCE) trial. The study included 301 patients (88 women and 213 men; mean age, 64.4 ± 10.2 [SD] years) recruited from May 2014 to May 2017 with stable symptoms of myocardial ischemia referred for nonemergent invasive angiography. Patients underwent coronary CTA and MPI before angiography with quantitative coronary angiography (QCA) measurements and fractional flow reserve (FFR). CTA examinations were analyzed using an FDA-cleared cloud-based software platform that performs AI-QCT for stenosis determination. Diagnostic performance was evaluated. Diagnostic algorithms were compared. Among 102 patients with no ischemia on MPI, AI-QCT identified obstructive (≥ 50%) stenosis in 54% of patients, including severe (≥ 70%) stenosis in 20%. Among 199 patients with ischemia on MPI, AI-QCT identified nonobstructive (1-49%) stenosis in 23%. AI-QCT had significantly higher AUC (all < .001) than MPI for predicting ≥ 50% stenosis by QCA (0.88 vs 0.66), ≥ 70% stenosis by QCA (0.92 vs 0.81), and FFR < 0.80 (0.90 vs 0.71). An AI-QCT result of ≥ 50% stenosis and ischemia on stress MPI had sensitivity of 95% versus 74% and specificity of 63% versus 43% for detecting ≥ 50% stenosis by QCA measurement. Compared with performing MPI in all patients and those showing ischemia undergoing invasive angiography, a scenario of performing coronary CTA with AIQCT in all patients and those showing ≥ 70% stenosis undergoing invasive angiography would reduce invasive angiography utilization by 39%; a scenario of performing MPI in all patients and those showing ischemia undergoing coronary CTA with AI-QCT and those with ≥ 70% stenosis on AI-QCT undergoing invasive angiography would reduce invasive angiography utilization by 49%. Coronary CTA with AI-QCT had higher diagnostic performance than MPI for detecting obstructive CAD. A diagnostic algorithm incorporating AI-QCT could substantially reduce unnecessary downstream invasive testing and costs. Clinicaltrials.gov NCT02173275.
深度学习框架已被应用于解释用于冠状动脉疾病 (CAD) 评估的冠状动脉 CTA。我们的研究目的是比较心肌灌注成像 (MPI) 和冠状动脉 CTA 与人工智能定量 CT (AI-QCT) 解释对有创血管造影中阻塞性 CAD 的诊断性能，并评估包括冠状动脉 CTA 和 AI-QCT 在诊断算法中的下游影响。本研究是对前瞻性 23 中心计算机断层扫描评估动脉粥样硬化性心肌缺血决定因素 (CREDENCE) 试验推导队列的回顾性事后分析。该研究纳入了 301 名患者（88 名女性和 213 名男性；平均年龄 64.4±10.2[SD]岁），于 2014 年 5 月至 2017 年 5 月期间因稳定的心肌缺血症状接受非紧急有创血管造影检查。患者在有创血管造影前接受了冠状动脉 CTA 和 MPI，并进行了定量冠状动脉造影 (QCA) 测量和血流储备分数 (FFR)。CTA 检查使用获得 FDA 批准的基于云的软件平台进行分析，该平台可用于 AI-QCT 以确定狭窄程度。评估了诊断性能。比较了诊断算法。在 102 名 MPI 无缺血的患者中，AI-QCT 确定 54%的患者存在阻塞性（≥50%）狭窄，其中 20%的患者存在严重（≥70%）狭窄。在 199 名 MPI 有缺血的患者中，AI-QCT 确定 23%的患者存在非阻塞性（1-49%）狭窄。AI-QCT 对预测 QCA 测量的 ≥50%狭窄（0.88 比 0.66）、QCA 测量的 ≥70%狭窄（0.92 比 0.81）和 FFR <0.80（0.90 比 0.71）的 AUC 均显著高于 MPI（所有 P<0.001）。AI-QCT 结果≥50%狭窄和应激 MPI 缺血的敏感性为 95%，特异性为 63%，而 QCA 测量的敏感性为 74%，特异性为 43%。与在所有患者中进行 MPI 检查和对有缺血的患者进行有创血管造影相比，在所有患者中进行冠状动脉 CTA 和 AI-QCT 检查，并对 AI-QCT 检查中≥70%狭窄的患者进行有创血管造影，可使有创血管造影的使用率降低 39%；在所有患者中进行 MPI 检查和对有缺血的患者进行冠状动脉 CTA 和 AI-QCT 检查，并对 AI-QCT 检查中≥70%狭窄的患者进行有创血管造影，可使有创血管造影的使用率降低 49%。AI-QCT 联合冠状动脉 CTA 对检测阻塞性 CAD 的诊断性能优于 MPI。包含 AI-QCT 的诊断算法可显著减少不必要的下游有创检查和费用。ClinicalTrials.gov NCT02173275。
