Myocardial Scar Assessment Using Artificial Intelligence-Powered Contrast-Free MRI

BACKGROUND: Myocardial scar detection using cardiovascular magnetic resonance (CMR) may now be achieved without contrast agents using artificial intelligence virtual native enhancement (VNE); prospective multicenter clinical validation is needed to establish its feasibility for routine clinical use. OBJECTIVES: The aim of this study was to prospectively investigate the diagnostic performance of VNE in myocardial scar detection using a multicenter, blinded approach in real-world clinical settings. METHODS: CMR data sets were prospectively collected from 2 sites (Leeds and Fuwai). Leeds and Fuwai teams independently determined scar presence on late gadolinium enhancement (LGE) as ground truth. The Oxford team generated VNE images using only precontrast cine and T1 map images from the other 2 sites. The Oxford team scored VNE images blinded to clinical data and LGE. Four additional clinical readers blindly assessed the matched-slice VNE and LGE images. RESULTS: VNE diagnosis of myocardial infarction achieved 94.4% accuracy on the basis of confident images (n = 107) and 87.5% accuracy on all images (n = 136). Myocardial scar quantification by VNE correlated strongly with LGE (R = 0.90) and agreed closely with LGE (mean difference 3.2%; 95% CI: -10.4% to +16.8%). VNE-LGE agreement on infarcted myocardial territories was 90.0%. The 4 clinical readers considered VNE adequate without proceeding to LGE in 69.7% of patients, and in these cases, VNE achieved an average diagnostic accuracy of 93.7%, comparable with that of LGE (93.9%). CONCLUSIONS: Accurate myocardial scar quantification can be achieved without contrast agents in high-quality and clear VNE images. VNE may triage the need for LGE and obviate its use in more than two-thirds of referrals for chronic myocardial infarction scar assessment without compromising diagnostic accuracy.