Automated stenosis estimation of coronary angiographies using end-to-end learning

Abstract

The initial evaluation of stenosis during coronary angiography is typically performed by visual assessment. Visual assessment has limited accuracy compared to fractional flow reserve and quantitative coronary angiography, which are more time-consuming and costly. Applying deep learning might yield a faster and more accurate stenosis assessment. We developed a deep learning model to classify cine loops into left or right coronary artery (LCA/RCA) or "other". Data were obtained by manual annotation. Using these classifications, cine loops before revascularization were identified and curated automatically. Separate deep learning models for LCA and RCA were developed to estimate stenosis using these identified cine loops. From a cohort of 19,414 patients and 332,582 cine loops, we identified cine loops for 13,480 patients for model development and 5056 for internal testing. External testing was conducted using automated identified cine loops from 608 patients. For identification of significant stenosis (visual assessment of diameter stenosis > 70%), our model obtained a receiver operator characteristic (ROC) area under the curve (ROC-AUC) of 0.903 (95% CI: 0.900-0.906) on the internal test. The performance was evaluated on the external test set against visual assessment, 3D quantitative coronary angiography, and fractional flow reserve (≤ 0.80), obtaining ROC AUC values of 0.833 (95% CI: 0.814-0.852), 0.798 (95% CI: 0.741-0.842), and 0.780 (95% CI: 0.743-0.817), respectively. The deep-learning-based stenosis estimation models showed promising results for predicting stenosis. Compared to previous work, our approach demonstrates performance increase, includes all 16 segments, does not exclude revascularized patients, is externally tested, and is simpler using fewer steps.

Keywords: Coronary angiography; Coronary artery disease; Deep learning; Ischemic heart disease; Myocardial infarction; Quantitative coronary angiography.

Fig. 1

Overview of the approach developed for stenosis estimation

Fig. 2

Overview of the different data partitions used for model development and testing. The curated dataset for model development included 13,840 patients, which was the union of the LCA and the RCA cine loops for training and validation. The curated internal test set included cine loops from 5056 patients (the union of LCA and RCA cine loops). Similarly, the curated external test set contained cine loops from 608 patients

Fig. 3

ROC curve for significant stenosis detection for each segment on the internal test set (visual assessment of diameter stenosis > 70%)

Fig. 4

Saliency maps highlighting regions influencing the predictions obtained using Gradient-based Class Activation Maps (Grad-CAM) [31]. The top left image highlights the saliency regions in the cine loop for predicting stenosis in the proximal RCA, where the model predicted 81.6% stenosis, and the cardiologist reported 70%. In the top right image, the model predicted 61.0% stenosis, while the cardiologist reported 80% stenosis in the middle LAD segment. Stenoses were also present in the middle RCA, distal RCA, and left main, however, their corresponding saliency maps are not shown. The two bottom images represent cases with no stenosis