AI Evaluation of Stenosis on Coronary CTA, Comparison With Quantitative Coronary Angiography and Fractional Flow Reserve

TLDR

Clinical reads of coronary CTA, particularly by less experienced readers, often overestimate stenosis severity relative to expert interpretation. This study aimed to compare the detection and grading performance of AI‑enabled quantitative coronary CTA (AI‑QCT) with core‑lab CTA, quantitative coronary angiography, and invasive fractional flow reserve. AI‑QCT was performed on 303 stable patients using an FDA‑cleared cloud‑based software that segments coronary arteries, quantifies plaque, and determines stenosis severity. The AI‑QCT analysis took an average of 10.3 minutes per patient, achieved 94% sensitivity and 84% accuracy for ≥50% stenosis, showed strong agreement with core‑lab QCA (ICC = 0.73), and produced a 7.3% false‑positive vessel rate, two‑thirds of which had FFR < 0.8. The study is registered under NCT02173275 (CREDENCE trial).

Abstract

Clinical reads of coronary computed tomography angiography (CTA), especially by less experienced readers, may result in overestimation of coronary artery disease stenosis severity compared with expert interpretation. Artificial intelligence (AI)-based solutions applied to coronary CTA may overcome these limitations.This study compared the performance for detection and grading of coronary stenoses using artificial intelligence-enabled quantitative coronary computed tomography (AI-QCT) angiography analyses to core lab-interpreted coronary CTA, core lab quantitative coronary angiography (QCA), and invasive fractional flow reserve (FFR).Coronary CTA, FFR, and QCA data from 303 stable patients (64 ± 10 years of age, 71% male) from the CREDENCE (Computed TomogRaphic Evaluation of Atherosclerotic DEtermiNants of Myocardial IsChEmia) trial were retrospectively analyzed using an Food and Drug Administration-cleared cloud-based software that performs AI-enabled coronary segmentation, lumen and vessel wall determination, plaque quantification and characterization, and stenosis determination.Disease prevalence was high, with 32.0%, 35.0%, 21.0%, and 13.0% demonstrating ≥50% stenosis in 0, 1, 2, and 3 coronary vessel territories, respectively. Average AI-QCT analysis time was 10.3 ± 2.7 minutes. AI-QCT evaluation demonstrated per-patient sensitivity, specificity, positive predictive value, negative predictive value, and accuracy of 94%, 68%, 81%, 90%, and 84%, respectively, for ≥50% stenosis, and of 94%, 82%, 69%, 97%, and 86%, respectively, for detection of ≥70% stenosis. There was high correlation between stenosis detected on AI-QCT evaluation vs QCA on a per-vessel and per-patient basis (intraclass correlation coefficient = 0.73 and 0.73, respectively; P < 0.001 for both). False positive AI-QCT findings were noted in in 62 of 848 (7.3%) vessels (stenosis of ≥70% by AI-QCT and QCA of <70%); however, 41 (66.1%) of these had an FFR of <0.8.A novel AI-based evaluation of coronary CTA enables rapid and accurate identification and exclusion of high-grade stenosis and with close agreement to blinded, core lab-interpreted quantitative coronary angiography. (Computed TomogRaphic Evaluation of Atherosclerotic DEtermiNants of Myocardial IsChEmia [CREDENCE]; NCT02173275).