Deep learning-assisted aortic stenosis detection and grading based on multiview versus single-view echocardiography.

Background: Advances in deep learning (DL) have shown promise in automating echocardiogram interpretation, thereby enhancing accuracy and efficiency in clinical practice. However, a fully automated pipeline for aortic stenosis (AS) analysis remains largely unexplored. This study aimed to develop a DL framework to streamline clinical AS assessment. Methods: A total of 499 AS studies (1,996 echocardiographic view) were selected from 17,436 cases of patients with valvular heart diseases (VHDs) obtained from three hospitals to form training (n=302), validation (n=76), and internal testing (n=121) datasets, while a prospectively collected set of 3,278 consecutive echocardiograms served as a real-world test data set. The DL framework automatically classified echocardiographic views, detected the presence of AS, and employed two algorithms to assess severity: multiview and single-view. Results: The DL model achieved high performance in AS detection in the prospective test dataset, with an area under the curve (AUC) of 0.942. The correlation between DL-graded metrics and manual measurements was excellent for aortic valve (AV) peak velocity (r=0.94; P<0.001), mean peak gradient (r=0.91; P<0.001), left ventricular outflow tract diameter (LVOTd) (r=0.81; P<0.001), AV velocity-time integral (VTI) (r=0.94; P<0.001), LVOT VTI (r=0.88; P<0.001), and AV area (r=0.87; P<0.001). Based on these metrics, the AUC of severe AS was the highest at 0.976 [95% confidence interval (CI): 0.953-1.0], significantly superior to those for moderate AS (AUC =0.907) and mild AS (AUC =0.874). The two-dimensional parasternal long-axis view method yielded comparable AUCs for all AS severities (AUC: 0.869-0.920). Conclusions: The proposed DL algorithm has the potential to automate and enhance the efficiency of clinical workflows for AS screening and grading in echocardiography.