Comprehensive Summary
Kim et al. developed and validated a deep learning-based quantitative coronary angiography (AI-QCA) system for automated analysis of major coronary vessels. Three convolutional neural network models (DeepLabV3+, U-Net++, and U2-Net) were trained on 7,658 angiograms from 3,129 patients to segment lumen boundaries and quantify stenoses. Performance was evaluated on 676 angiograms from 370 patients against expert manual QCA. AI-QCA achieved 89% lesion-detection sensitivity and strong correlations with manual QCA for diameter stenosis (R2 = 0.69), minimum lumen diameter (R2 = 0.86), reference lumen diameter (R2 = 0.84), and lesion length (R2 = 0.44). Among 995 matched lesions, 80% exhibited ≤10% difference in stenosis compared with manual assessment. The system accurately identified multiple lesions without manual correction, though performance declined in distal segments and small vessels.
Outcomes and Implications
AI-QCA enables rapid, reproducible, and fully automated quantification of coronary stenoses, offering performance comparable to expert analysis and potential integration into interventional workflows. By minimizing operator bias and analysis time, it could enhance procedural planning, stent sizing, and physiologic assessment. However, clinical implementation requires validation across a more diverse set of imaging conditions, inclusion of side branches and bifurcations, and automation of frame selection to ensure robustness and generalizability.