Evaluation of a deep learning-based automated algorithm for labeling coronary arteries in computed tomography angiography images

Abstract Objective To evaluate the performance of a fully automatic algorithm for labeling coronary arteries in CCTA images using deep learning based on the two 3-dimensional (3D) U-Net architectures for myocardium structure extraction. Methods In total, 157 patients who underwent CCTA scanning were retrospectively included. An automatic coronary artery labeling algorithm based on the distance transformation algorithm was proposed to identify the anatomical segments of the centerlines extracted from CCTA images. Sixteen segments were identified and labeled. The results obtained via the algorithm were recorded and reviewed by three experts. The performance of segment detection and labeling of each segment was evaluated, and the proportion of agreement between the two experts on the manually labeled segments was also calculated. Results Compared with the labels of the experts, 117 labels (5.4%) (2180 segments) from the algorithm needed to be changed or removed. The overall accuracy of label presence was 96.2%. The average overlap between the expert reference and algorithm labels was 94.0%. The average agreement between the two experts was 95.0%. Conclusions The proposed deep learning algorithm provided a high accuracy of the automatic labeling with respect to the labels from the clinical experts. This method is promising for labeling coronary arteries automatically and alleviating the workload of radiologists in the near future.