EchoFM: A Pre-training and Fine-tuning Framework for Echocardiogram Videos Vision Foundation Model

BackgroundEchocardiograms provide essential insights into cardiac health, yet their complex, multidimensional data poses significant challenges for analysis and interpretation. Existing deep learning models for echocardiogram analysis often rely heavily on supervised training, which limits their generalizability and robustness across different datasets and clinical environments. ObjectiveTo develop and evaluate Echo-Vision-FM (Echocardiogram video Vision Foundation Model), a self-supervised video learning framework designed to pre-train a video encoder on large-scale, unlabeled echocardiogram data. Echo-Vision-FM aims to produce robust and transferable video representations, improving downstream performance across diverse echocardiogram datasets and clinical conditions. MethodsThe proposed framework employs advanced self-supervised video learning through a masked auto-encoding technique, which compresses segments of video data and reconstructs the full video by masking non-overlapping video patches. An asymmetric encoder-decoder architecture underpins this approach. To further enhance the learned representations, we introduce STF-Net, a Spatial-Temporal Fusion Net, which integrates spatial and temporal correlations from the video representations. We pre-trained Echo-Vision-FM using the MIMIC-IV-ECHO dataset and fine-tuned it across multiple downstream datasets for specific clinical tasks, including morphological value estimation and the diagnosis of heart function and diseases. ResultsEcho-Vision-FM achieved superior performance in classifying left ventricular ejection fraction (LVEF), with an accuracy of 0.905, an F1 score of 0.941, and an AUC of 0.931. In regression tasks, Echo-Vision-FM outperformed state-of-the-art models, achieving a mean absolute error (MAE) of 3.87% and an r2 of 0.825 for LVEF prediction. The model also demonstrated significant improvements in estimating end-systolic and end-diastolic volumes, with r2 values of 0.782 and 0.742, respectively. Incorporating STF-Net further enhanced performance across all tasks. ConclusionOur results demonstrate that large-scale self-supervised video learning on echocardiogram data enables the extraction of transferable and clinically relevant features, surpassing existing methods. The Echo-Vision-FM framework, particularly with the inclusion of STF-Net, significantly improves the extraction of spatiotemporal features, resulting in enhanced predictive accuracy for a range of cardiac parameters. Echo-Vision-FM offers a scalable and effective solution for echocardiogram analysis, with promising applications in clinical diagnostics and research.