Investigation of the accuracy of classifying coronary artery disease severity using machine learning with subdomain analysis of Fractional Flow Reserve diagnosis in patients

Coronary artery disease (CAD) is a condition where there is blood-flow reduction in the coronary artery due to plaque build-up. The current standard to diagnose CAD severity is fractional flow reserve (FFR) using the ratio of distal and proximal stenotic pressure measurements. This work investigated the use of a machine-learning classifier of CAD severity. Sixty-four coronary CT angiographies (CCTA) were collected at 70% through the cardiac R-R cycle. Eight straightened curved planar reformations (SCPRs) were reconstructed from each CCTA considering 45° increments around the coronary artery centerline. FFR measurements were considered ground truth to train a convolutional neural network to predict CAD severity based on the 0.80 FFR threshold. Classification accuracy and area under the receiver operating characteristic curve (AUROC) were used to assess the network’s predictive capacity. SCPR data were optimized using class-activation maps, and the network was re-trained and assessed in the same manner. Subgroup analysis of the network’s performance was carried out considering different coronary artery branches and patient FFR measurements in and out of the FFR grey-zone. Different network input conditions were assessed such as SCPR slice-thickness and SCPR reconstruction using the minimum or average value across the vessel centerline. Network for CAD severity prediction was significantly higher (P<0.05) using thicker SCPR slices. No significant difference was found in network performance using SCPRs from different coronary artery branches, or considering SCPR reconstruction using the minimum or average value. This work indicates that a CNN can predict CAD severity using coronary artery SCPRs.