Invasive methods for diagnosing coronary artery disease through catheterization, such as invasive coronary angiography (ICA) and fractional flow reserve (FFR), are considered the gold standard for measuring the severity of stenosis in patients with coronary artery disease. Coronary CT angiography (CCTA) is a non-invasive imaging technique that enables the detection of both stenosis and plaque in the coronary vessels. Thanks to its high sensitivity, it is used as a first-line diagnostic tool for identifying which patients require catheterization for further invasive treatment. Nevertheless, studies have shown that 20 to 50% of patients identified using CCTA did not, in fact, have significant functional stenosis. In view of this, numerous AI image analysis models have been proposed that combine CCTA with functional stenosis measurements through computed tomography-derived FFR (CT-FFR), to increase specificity and thus reduce unnecessary admissions to the cath lab. Despite their success, they are prone to errors when presented with suboptimal image quality and heterogeneity in the image data.

The goal of this work is to extend the applicability of AI-based analysis to this use case by developing methods robust to artifacts and outliers. This will facilitate the accurate identification of patients requiring invasive treatment, thus making diagnosis more accessible to patients in hospitals with less advanced imaging equipment. Finally, to enable autonomous employment of our proposed methods, special attention will be paid to ensuring their interpretability by clinicians and medical staff.

This project is part of the Artificial Intelligence for Accessible Medical Imaging (AI4AI) consortium.