EchoJEPA: A Latent Predictive Foundation Model for Echocardiography

TL;DR

EchoJEPA introduces a latent-predictive foundation model for echocardiography that suppresses stochastic speckle and acquisition artifacts to learn anatomy-focused representations. Trained on a massive corpus of 18 million videos, it demonstrates state-of-the-art performance on $LVEF$ estimation, $RVSP$ prediction, and view classification, while delivering exceptional sample efficiency and robustness to acoustic degradation. The work further presents a standardized multi-view probing framework for fair evaluation and generalization across adult and pediatric cohorts, and provides a public release of EchoJEPA-L for community validation. By showing that objective-domain alignment via latent prediction outperforms pixel-reconstruction baselines, it highlights a practical path toward more reliable, data-efficient ultrasound foundation models with broad clinical impact.

Abstract

Foundation models for echocardiography promise to reduce annotation burden and improve diagnostic consistency by learning generalizable representations from large unlabeled video archives. However, current approaches fail to disentangle anatomical signal from the stochastic speckle and acquisition artifacts that dominate ultrasound imagery. We present EchoJEPA, a foundation model for echocardiography trained on 18 million echocardiograms across 300K patients, the largest pretraining corpus for this modality to date. We also introduce a novel multi-view probing framework with factorized stream embeddings that standardizes evaluation under frozen backbones. Compared to prior methods, EchoJEPA reduces left ventricular ejection fraction estimation error by 19% and achieves 87.4% view classification accuracy. EchoJEPA exhibits strong sample efficiency, reaching 78.6% accuracy with only 1% of labeled data versus 42.1% for the best baseline trained on 100%. Under acoustic perturbations, EchoJEPA degrades by only 2.3% compared to 16.8% for the next best model, and transfers zero-shot to pediatric patients with 15% lower error than the next best model, outperforming all fine-tuned baselines. These results establish latent prediction as a superior paradigm for ultrasound foundation models.

Figures (4)

• Figure 1: EchoJEPA architecture. Echocardiograms from multiple views are partitioned into spatio-temporal tubelets and split into masked and unmasked sets. The encoder $E_\theta$ processes masked video frames, and the predictor $P_\phi$ infers embeddings for masked regions conditioned on learnable mask tokens. The EMA encoder $E_{\bar{\theta}}$ processes unmasked frames to provide prediction targets. The $L_1$ loss is computed between predicted and target embeddings, with no gradients flowing into the EMA encoder.
• Figure 2: Multi-view probing framework. The frozen EchoJEPA encoder extracts video embeddings from multiple echocardiographic views. Each embedding is augmented with learnable view and clip stream embeddings encoding position in the study. During training, view dropout randomly masks views to improve robustness to variable study composition. The concatenated tokens are passed to a lightweight attentive probe that outputs study-level predictions.
• Figure 3: Downstream evaluation. EchoJEPA pretrained on 300K patients and 18M videos is evaluated on three clinical tasks with frozen backbones and lightweight probes. RVSP estimation, LVEF regression, and view classification.
• Figure 4: Attention visualization comparing VideoMAE and V-JEPA. Columns show three frames from an apical four-chamber echocardiogram under pretrained and finetuned conditions. Rows display received attention and given attention for each model. Finetuned V-JEPA in the bottom right demonstrates precise localization on valve leaflets and ventricular walls synchronized to cardiac motion.