Deep Video Networks for Automatic Assessment of Aortic Stenosis in Echocardiography.

Aortic valve stenosis (AS) is the narrowing of the heart’s aortic valve opening, which restricts blood flow from the left ventricle to the aorta. Accurate diagnosis and timely intervention of AS are crucial since the mortality rate of this condition rapidly increases as symptoms begin to develop. Automated AS estimation in echocardiography faces several challenges, including generalization from diverse medical data, access to high-quality Doppler imaging, and noisy training labels. In this paper, we propose a method for automatic Aortic Stenosis assessment in echocardiography, which is, to the best of our knowledge, the first deep learning pipeline to automate the identification and grading of AS using cardiac ultrasound. Trained and evaluated on a large dataset of 9,117 echocardiograms obtained from 2,247 patients, our method achieves a mean \(F_1\) score of 96.5% for the identification of AS and a mean \(F_1\) score of 73% for grading AS. We use a multi-task training scheme to predict AS severity and key parameters used in clinical AS assessment along with their aleatoric uncertainties. Compared to a baseline that only predicts AS severity, our results show that our multi-task uncertainty-aware inference method achieves comparable classification performance while improving the ability to detect out-of-distribution examples. This is crucial for the clinical deployment of our method in point-of-care settings, where ultrasound operators have less experience in acquiring high-quality echocardiograms.