https://doi.org/10.1007/s00392-025-02737-x
1Internal Medicine V, Department of Angiology University Hospital Schleswig Holstein 24105, Deutschland; 2University of Lübeck Institute for Cardio Genetics Lübeck, Deutschland; 3Transilvania University of Brasov Automation and Information Technology Brasov, Rumänien; 4Universitätsklinikum Mannheim GmbH I. Medizinische Klinik Mannheim, Deutschland
Background:
Coronary artery disease (CAD) remains a leading cause of global mortality, largely due to challenges in predicting ischemic damage and optimizing therapeutic interventions. Coronary collateral circulation (CCC) has emerged as a naturally protective mechanism in some patients, mitigating the effects of arterial obstruction through alternative vascular pathways1. However, CCC's clinical and therapeutic relevance is often underappreciated, and its quantification lacks standardization and reproducibility across clinical practice. To address these critical gaps, the PROGRESS consortium (PRecisiOn medicine in CAD patients: artificial intelliGence for integRated gEnomic, functional and anatomical aSSessment of the coronary collateral circulation) was established.
Methods:
The study utilized a comprehensive cohort of 3,151 CAD patients with 4,587 coronary examinations, each linked to a coronary event as the index procedure4. Among these, 369 patients presented with chronic total occlusion (CTO), with 177 demonstrating CCC-positive status and 193 classified as CCC-negative.
Data Annotation and Preparation: Angiographic images underwent manual annotation using a specialized tool developed specifically for this study. Expert annotators marked collateral vessels within angiograms while recording quantifiable parameters including collateral vessel diameter (in both millimeters and pixels), Rentrop grading scores, collateral flow grades, and blush grades5.
AI Architecture Development: The deep learning framework employed a novel architecture extracting both spatial and temporal features from coronary angiographic sequences (DICOMs). The model backbone utilized a convolutional neural network (CNN) processing individual angiography frames to capture spatial characteristics, with features subsequently concatenated and processed through additional convolutional layers for temporal analysis. To address data scarcity challenges, the team implemented transfer learning by pretraining the backbone on coronary artery segmentation tasks, which consistently enhanced performance. Few-shot learning (FSL) techniques were further applied to improve model generalization capabilities in low-data regimes.
Results:
The deep learning tool demonstrated superior sensitivity for collateral detection compared to expert human assessment (88% vs 82%) in blinded comparative evaluations where cardiovascular specialists competed directly against the algorithm1. However, human experts maintained higher specificity (85% vs 74%), indicating complementary strengths between AI and clinical expertise. Longitudinal analysis over 15 years of follow-up demonstrated that well-developed CCC in CTO patients significantly improved survival outcomes. Critically, the AI algorithm independently predicted this survival benefit without requiring expert interpretation, establishing its potential for expert-independent clinical decision support.
Conclusion:
The PROGRESS study demonstrates the transformative potential of AI-driven image analysis for advancing both scientific understanding and clinical application of coronary collateral circulation. The integration of automated detection capabilities with long-term outcome prediction establishes a foundation for precision medicine approaches, potentially improving risk stratification and treatment selection for patients with complex coronary anatomy.