1Universitätsklinikum Schleswig-Holstein Medizinische Klinik II / Kardiologie, Angiologie, Intensivmedizin Lübeck, Deutschland; 2Transilvania University of Brasov (UTVB), , Romania Automation and Information Technology Brasov, Rumänien; 3Universitätsklinikum Mannheim I. Medizinische Klinik Mannheim, Deutschland
Coronary artery disease (CAD) is the dominant cause of death and hospitalization across the globe. Atherosclerosis, an inflammatory condition that gradually narrows arteries and has potentially fatal effects on the coronary circulation, is the most frequent cause of CAD and thus responsible for millions of deaths worldwide. While the onset and progression of CAD can be accurately predicted by taking into consideration classic cardiovascular risk factors, morbidity as a cause of CAD cannot be predicted with the same accuracy. One of the possible factors underlying this fact is that the circulation regularly adapts to vascular narrowing through the formation of collateral arteries, resulting in significant long-term health benefits. However, the factors underlying i.e. coronary collateral circulation (CCC) formation are poorly understood and, therefore, the relevance of CCC remains to be debated.
A major obstacle for the uncovery of the prognostic relevance of CCC is timely and observer-independent detection of CCC. In literature, there is a large heterogeneitiy in CCC scoring systems used and there is no general consensus in the prognostic relevance of each of these scoring systems.
Artificial intelligence may help to overcome these challenges. By analysing a cohort of over 3000 CAD patients in the course of the trans-European collaborative consortiums PROGRESS (PRecisiOn medicine in CAD patients: artificial intelliGence for integRated gEnomic, functional and anatomical aSSessment of the coronary collateral circulation) we were able to generate training data for a neural network and are now able to present a novel deep learning-based method to detect CCC from angiographic images. Our method relies on a convolutional backbone to extract spatial features from each frame of an angiography sequence (DICOM). The features are then concatenated, and subsequently processed by another convolutional layer that processes embeddings temporally. Pretraining the backbone on coronary artery segmentation improved the results consistently. Experimenting with few-shot learning (FSL) further improved performance to over 80% accuracy.
In order to overcome the “black box” concern of artificial intelligence in medicine, we performed numerous subgroup analyses based on i.e. rentrop grading, collateral flow, and collateral grading, which provided valuable insights into how our model takes decisions.
Importantly, we tested the performance of the deep learning-based tool by comparing the inter-observer accuracy in the appraisal of CCC in a blinded fashion by experts in the treatment of chronic total occlusion (CTO) to the accuracy of our deep learning based tool.
Our cohort features a mean follow-up period of 6 years and well as in-depth patient characterization with over 500 data points per patient. Thus, we are able to analyse factors impacting on as well as the prognostic relevance of CCC assessed by both the classical scoring systems as well as our newly developed deep learning-based tool.
Overall, the proposed method shows promising results in detecting CCC and can be further extended to perform landmark based CCC detection and CCC quantification. Our study shows that artificial intelligence-based image analysis has the potential to assist interventionalists in the catheter lab on taking optimal decisions for the treatment of CAD featuring CTO, which remains a highly challenging and highly dynamic field.