EchoJEPA: Latent Predictive Echocardiography Model

• EchoJEPA is a latent predictive foundation model for echocardiography that uses JEPA and a ViT backbone to disentangle clinically relevant anatomical signals from ultrasound noise.
• It introduces a novel factorized multi-view probing strategy with stream embeddings to improve accuracy in tasks like view classification and left ventricular function estimation.
• Empirical results demonstrate superior sample efficiency, up to 19% MAE reduction, and enhanced robustness against acoustic and artifact perturbations compared to previous models.

EchoJEPA is a latent predictive foundation model for echocardiography, leveraging Joint-Embedding Predictive Architectures (JEPA) and a vision transformer (ViT) backbone to learn generalizable, anatomy-centric representations from large, heterogeneous ultrasound video archives. It directly targets the separation of clinically relevant anatomical signal from common confounders in ultrasound imaging, such as stochastic speckle and acquisition artifacts. EchoJEPA is pretrained on the largest available corpus of echocardiograms—18.1 million studies across 300,000 patients—and introduces a novel multi-view probing methodology utilizing factorized stream embeddings. Empirically, EchoJEPA exhibits superior sample efficiency, robustness, and zero-shot transfer capabilities relative to prior baselines, establishing latent prediction as a dominant paradigm for foundation modeling in clinical ultrasound (Munim et al., 2 Feb 2026).

1. Latent Predictive Architecture

EchoJEPA employs a masked spatio-temporal prediction framework, operating exclusively in embedding space and avoiding pixel-level reconstruction decoders. The input echocardiogram video $V$ is partitioned into $N$ tubelets (localized spatio-temporal subvolumes), which are divided into context tubelets $x$ (visible to the model) and target tubelets $y$ (masked). The network architecture comprises three modules:

1. Context Encoder ($E_\theta$): A ViT backbone processes $x$ to produce context-specific latent tokens.
2. Predictor ($P_\phi$): Consumes the context embeddings $E_\theta(x)$ together with learnable mask tokens $\Delta_y$ to output predicted embeddings for the masked regions.
3. EMA Target Encoder ($E_{\bar\theta}$): Provides stable targets for $N$0 by maintaining an exponential-moving-average of $N$1.

The pretraining objective minimizes the $N$2 loss between predicted and target embeddings: $N$3 where $N$4 indicates the stop-gradient operator. Following each update, EMA weights are updated as

$N$5

with approximately $N$6 to suppress the variance from speckle noise (Munim et al., 2 Feb 2026).

2. Factorized Multi-View Probing and Stream Embeddings

To rigorously evaluate clinically relevant representations under frozen backbone constraints, EchoJEPA introduces a multi-view probing protocol. Each 2D video clip is attributed with two learnable vectors: a view embedding $N$7 (specific to view type) and a clip embedding $N$8 (specific to temporal instance). For a token $N$9 originating from stream $x$0, the initial representation is

$x$1

This factorized approach, requiring just $x$2 auxiliary vectors instead of full-rank $x$3, promotes anatomical and geometric features over unpredictable speckle and acquisition idiosyncrasies.

The multi-view probe operates as a transformer stack (depth $x$4, 16 heads, MLP ratio 4) that concatenates representations from all $x$5 streams, applies the stream embeddings, manages missing views via key-padding masks, and uses a learnable query token with cross-attention to aggregate into a compact vector $x$6 for downstream task heads.

3. Pretraining Regimen and Domain-specific Augmentation

EchoJEPA-G, the highest-capacity instance, was pretrained on 18.1 million echocardiograms for 280 epochs at $x$7 resolution, followed by 80 epochs at $x$8. Masking protocols employ 8 tubelet blocks at spatial scale 0.15 and 2 at 0.7, always masking over the full temporal axis to emphasize temporal prediction. The global batch size is 1024, total update count is approximately 84,000 (86 million clips), and EMA momentum is fixed at 0.99925.

Crucially, augmentations are tailored: temporal sampling at 24 fps (compared to the 4 fps default) is used to capture rapid cardiac motion, and geometric transforms (aspect ratio in [0.9, 1.1], crop scale in [0.5, 1.0]) preserve diagnostically meaningful anatomy.

4. Quantitative Performance and Generalization

EchoJEPA-G demonstrates substantial improvements across multiple clinical benchmarks compared to echocardiography foundation model baselines such as EchoPrime, PanEcho, and EchoMAE-L:

• Left Ventricular Ejection Fraction (LVEF) Estimation: Across sites, EchoJEPA-G achieves mean absolute error (MAE) reductions of approximately 19% over the previous state of the art; for example: Toronto 4.26% MAE (EchoJEPA-G) vs 5.33% (EchoPrime), Chicago 5.44% vs 6.71%.
• View Classification: EchoJEPA-G attains 87.4% accuracy under frozen probes, significantly outperforming EchoPrime (≈42%) and PanEcho (≈42%).
• Sample Efficiency: With only 1% of labeled data (≈220 views), EchoJEPA-G achieves 78.6% accuracy, while baselines trained with all labels saturate at approximately 42%.
• Robustness to Acoustic Perturbations: Under increasing acoustic artifact severity, EchoJEPA-G's LVEF MAE degrades by only 2.3%, compared to 16.8% for EchoPrime, confirming reduced sensitivity to non-anatomical confounds.
• Zero-shot Pediatric Transfer: On EchoNet-Pediatric, EchoJEPA-G exhibits a 15.3% lower MAE (4.32%) than EchoPrime (5.10%) without any pediatric-specific fine-tuning; after fine-tuning, the MAE further drops to 3.88%.

Task/Setting	EchoJEPA-G Accuracy/MAE	Best Baseline/Notes
LVEF Estimation (Toronto)	4.26% MAE	5.33% (EchoPrime)
View Classification (frozen)	87.4%	≈42% (EchoPrime)
1% Label Sample Efficiency	78.6%	<22% (others)
Acoustic Perturbation (Avg. Deg.)	+2.3% degradation	16.8% (EchoPrime)
Zero-shot Pediatric LVEF MAE	4.32%	5.10% (EchoPrime)

These results confirm that the latent predictive approach with multi-view factorization enhances both clinical accuracy and generalization capacity (Munim et al., 2 Feb 2026).

5. Ablations and Model Component Analysis

Ablation studies clarify the contribution of key innovations:

• Stream Embeddings: Removing these increases RVSP estimation MAE from 4.54 to 4.63 (+2.0%).
• Early vs Late Fusion: Replacing the attention-based late fusion with late averaging increases the RVSP MAE to 5.09 (+12.1%).
• View Dropout: Disabling view dropout (probability $x$9) increases MAE to 5.37 (+18.3%).

Hyperparameter selection is performed via grid sweeps: learning rates $y$0, weight decay $y$1. The probe head is identical (4 layers, 16 heads, MLP ratio 4) across all comparative experiments to isolate representation quality.

This analysis supports that both architectural innovations (factorization, attention-based fusion) and sampling/augmentation decisions are critical to EchoJEPA's observed performance.

6. Context, Comparative Approaches, and Implications

EchoJEPA is positioned as a step change in the foundation modeling of clinical ultrasound, providing the first demonstration that latent-predictive pretraining over large-scale, real-world echocardiography data can disentangle anatomical structure from complex noise processes intrinsic to ultrasound imaging. Prior efforts—including EchoPrime (contrastive), PanEcho (supervised), and EchoMAE (reconstruction)—are shown to be less sample-efficient, more brittle under artifact perturbations, and less capable of transferring to out-of-domain patient populations.

A plausible implication is that the latent-prediction paradigm's reduced reliance on pixel-level details results in greater invariance to nuisance signals, whereas factorized multi-view probing aligns model evaluation with multi-camera, multi-view clinical practice. EchoJEPA's architecture and pretraining framework may generalize to other medical imaging modalities where structured spatial and temporal redundancy, as well as artifact confounds, are central challenges. The methodology advances standardized evaluation for frozen foundation model backbones and demonstrates new state-of-the-art metrics in clinical and pediatric tasks (Munim et al., 2 Feb 2026).