Deep Learning-Based Segmentation of Coronary Arteries and Stenosis Detection in X-Ray Coronary Angiography

Deep learning applications may assist in automatically detecting coronary arteries on invasive coronary angiography (ICA).

The authors aimed to train deep learning models for the segmentation of coronary arteries and the detection of significant stenoses on ICA, conduct external validation, and compare the performance with expert variabilities.

ICA studies from Amsterdam University Medical Centers (center 1) and Emory University Hospital (center 2) were retrospectively collected. Contours of the main coronary arteries and their ≥50% stenoses were manually segmented using dedicated software. Deep learning-based models were created using data from center 1, center 2, and both centers. The performance of the models was assessed on unseen data and compared to expert variability.

A total of 10,573 ICA images were used to train models: 9,065 from center 1 (n = 2,624) and 1,508 (n = 456) from center 2. Validation was done on 186 center 1 images and 123 center 2 images. The segmentation model trained on data sets from both centers had the highest median Dice coefficient (0.86; IQR: 0.81-0.88). The stenoses detection algorithm trained on both centers achieved a detection rate of 0.67 (95% CI: 0.63-0.71), similar to expert agreement (0.65; 95% CI: 0.63-0.68). The model trained on the data with the most stenoses yielded the highest stenosis detection rate (0.67; 95% CI: 0.64-0.71). When matched for data set size and proportion of stenoses, the models trained on both centers performed similarly.

The models achieved performance levels on par with experts in coronary artery segmentation and detection of significant stenoses in the main arteries.