Quantitative coronary angiography (QCA) typically employs traditional edge detection algorithms that often require manual correction. This has important implications for the accuracy of downstream 3D coronary reconstructions and computed haemodynamic indices (e.g. angiography-derived fractional flow reserve). We developed AngioPy, a deep-learning model for coronary segmentation that employs user-defined ground-truth points to boost performance and minimise manual correction. We compared its performance without correction with an established QCA system.
Deep learning models integrating user-defined ground-truth points were developed using 2455 images from the Fractional Flow Reserve versus Angiography for Multivessel Evaluation 2 (FAME 2) study. External validation was performed on a dataset of 580 images. Vessel dimensions from 203 images with mild/moderate stenoses segmented by AngioPy (without correction) and an established QCA system (Medis QFR®) were compared (609 diameters).
The top-performing model had an average F1 score of 0.927 (pixel accuracy 0.998, precision 0.925, sensitivity 0.930, specificity 0.999) with 99.2 % of masks exhibiting an F1 score > 0.8. Similar results were seen with external validation (F1 score 0.924, pixel accuracy 0.997, precision 0.921, sensitivity 0.929, specificity 0.999). Vessel dimensions from AngioPy exhibited excellent agreement with QCA (r = 0.96 [95 % CI 0.95-0.96], p < 0.001; mean difference - 0.18 mm [limits of agreement (LOA): -0.84 to 0.49]), including the minimal luminal diameter (r = 0.93 [95 % CI 0.91-0.95], p < 0.001; mean difference - 0.06 mm [LOA: -0.70 to 0.59]).
AngioPy, an open-source tool, performs rapid and accurate coronary segmentation without the need for manual correction. It has the potential to increase the accuracy and efficiency of QCA.