Background:
Coronary artery disease (CAD) accounts for a substantial proportion of global cardiovascular mortality. Coronary collateral circulation (CCC) functions as an endogenous cardioprotective mechanism. Despite its clinical significance, CCC assessment lacks standardization and remains underutilized in risk stratification and treatment planning. The PROGRESS consortium (PRecisiOn medicine in CAD patients: artificial intelliGence for integRated gEnomic, functional and anatomical aSSessment of the coronary collateral circulation) was established as a transnational ERA PerMed initiative to address these critical gaps through integration of artificial intelligence, genomics, and advanced imaging analysis.
Methods:
This multicenter study analyzed a comprehensive registry cohort of 3,151 CAD patients undergoing 4,587 coronary examinations at Universitäres Herzzentrum Lübeck between 2006-2013, with each examination linked to a coronary event as the index procedure. Among these, 369 patients presented with chronic total occlusion (CTO), including 177 CCC-positive and 193 CCC-negative cases.
A specialized annotation tool was developed for systematic angiographic image analysis, enabling expert annotators to quantify collateral vessel parameters including diameter (mm and pixels), Rentrop grading, collateral flow grades, and myocardial blush grades. The deep learning architecture employed a novel spatiotemporal convolutional neural network (CNN) framework processing DICOM coronary angiography sequences. The model backbone extracted spatial features from individual frames, with subsequent convolutional layers performing temporal integration across sequences. To overcome data scarcity limitations, the team implemented transfer learning via pretraining on coronary artery segmentation tasks and applied few-shot learning (FSL) techniques to enhance model generalization in low-data regimes.
Results:
In blinded head-to-head comparative evaluations, the deep learning algorithm achieved 88% sensitivity for collateral detection versus 82% for expert cardiovascular specialists, demonstrating superior detection capability. Human experts maintained higher specificity (85% vs 74%), highlighting complementary diagnostic strengths between AI and clinical expertise. Longitudinal survival analysis over 15 years of follow-up demonstrated that well-developed CCC in CTO patients conferred significant survival benefit. Critically, the AI algorithm independently predicted this mortality advantage without requiring expert interpretation, establishing proof-of-concept for expert-independent prognostic assessment. The integration of genomic analysis with imaging phenotypes enabled identification of genetic variants associated with CCC development, advancing precision medicine approaches for coronary collateralization.
Conclusion:
The PROGRESS study demonstrates that AI-driven spatiotemporal analysis of coronary angiography enables automated, reproducible quantification of collateral circulation with performance exceeding conventional expert assessment in sensitivity. These findings support the clinical implementation of AI-assisted CCC assessment as a standardized tool for optimizing revascularization strategies and identifying candidates for therapeutic arteriogenesis interventions.
