Comprehensive Summary
In this study, Tu et al. systematically reviewed 18 studies including 3,568 patients and 13,362 vascular images in order to evaluate deep learning accuracy for coronary artery stenosis detection across CCTA and CAG platforms. Binary classification models achieved pooled accuracies of 0.79 for <50% stenosis and 0.73 for >50% stenosis at vessel level. In addition, multiclass grading demonstrated peak performance in intermediate severity bands, with accuracies of 0.86 for 25-50% stenosis and 0.83 for 50-70% stenosis. However, reduced precision and wide confidence intervals were observed in the ≥70% range. Because of methodological heterogeneity, including non-uniform stenosis thresholds, variable imaging inputs, and incomplete diagnostic contingency reporting, the planned bivariate diagnostic meta-analysis could not be performed. Instead, the authors pooled reported accuracy estimates using conventional meta-analytic aggregation.
Outcomes and Implications
The findings demonstrate that deep learning can reach acceptable diagnostic performance for coronary stenosis categorization, particularly within intermediate severity ranges commonly relied upon for risk stratification and treatment decisions. Accuracy instability at both minimal and high-grade stenosis highlights the need for standardized stenosis thresholds as well as consistent image annotation practices. The lack of prospective implementation studies and clinical outcome linkage limits claims regarding readiness for integration into diagnostic workflows, supporting the need for coordinated multicenter validation and reporting frameworks before routine deployment.