Open Collections

Abstract

Aortic stenosis (AS) is a life-threatening heart valve disease affecting approximately 5% of individuals over the age of 65, with rapid progression making timely detection critical for effective treatment. However, current diagnostic procedures rely on comprehensive echocardiographic evaluations, which are typically available only in well-resourced hospitals, limiting accessibility to patients and increasing costs for healthcare systems. To address this gap, we propose a series of deep learning–based techniques for automated AS severity assessment from standardized echocardiogram views. Key challenges in automating AS assessment include robust classification of standard-plane views, interpretable prediction of disease severity, and managing the noise and heterogeneity present in echocardiographic data. We introduce methods which address these challenges while preserving the strong predictive performance and flexibility of deep neural networks. We present a lightweight classifier for detecting relevant standard-plane views and filtering out nonstandard views. We also propose an interpretable prototypical part neural network that follows a transparent decision-making process based on similarity with existing examples. Furthermore, we propose architecture-agnostic training strategies that mitigate the limitations of partial anatomical information, which is inherent to 2-D imaging of the 3-D patient anatomy. Additionally, we introduce a calibration approach to improve performance in under-represented subgroups. We validate our methods on both public and private datasets, demonstrating competitive performance relative to existing state-of-the-art approaches.