EchoApex: A General-Purpose Vision Foundation Model for Echocardiography
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Quantitative evaluation of echocardiography is essential for precise assessment of cardiac condition and guiding treatment decisions. The diverse nature of echo images, including variations in probe types, manufacturers, and pathologies, poses challenges for developing artificial intelligent models that can generalize across different clinical practice. Here, we introduce EchoApex, a general-purpose vision foundation model designed for comprehensive echocardiography analysis. Pretrained on over 20 million images from 11 clinical centers, EchoApex utilizes self-supervised learning and task-specific decoders for diverse applications in echocardiography. It performs sequence view classification, interactive structure segmentation, left ventricle measurement, and ejection fraction estimation. In benchmark evaluations, EchoApex outperforms task-specific models, achieving a mean BACC of 0.976 in classification of 18 common views, DICE of 0.93 in chamber segmentation, MAE of 3.9mm in left ventricle linear measurement and a zero-shot performance improvement over specialist models in all evaluated datasets. For ejection fraction estimation, EchoApex achieves an MAE of 5.6% and an AUC of 0.927 for cardiomyopathy detection. Despite using less than 4% of trainable parameters with frozen encoders, EchoApex with adapter demonstrates strong performance with minimal degradation compared to fully finetuned models. This work establishes EchoApex as a scalable, general-purpose model for echocardiography, enabling efficient adaptation across various clinical tasks.