Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
EchoJEPA: A Latent Predictive Foundation Model for Echocardiography
| Date | Stars |
|---|---|
| 2026-07-31 | 328 |
| 2026-08-05 | 328 |
| 2026-08-06 | 328 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
<h1 align="center"><b>EchoJEPA</b></h1>
<h3 align="center">A Latent Predictive Foundation Model for Echocardiography</h3>
<p align="center">
<a href="https://arxiv.org/abs/2602.02603" target="_blank"><img src="https://img.shields.io/badge/arXiv-Paper-B31B1B?style=for-the-badge&logo=arxiv&logoColor=white" alt="arXiv"></a>
<a href="https://github.com/bowang-lab/EchoJEPA"><img src="https://img.shields.io/badge/GitHub-Code-4A90E2?style=for-the-badge&logo=github&logoColor=white" alt="GitHub"></a>
<a href="https://echojepa.com/"><img src="https://img.shields.io/badge/Website-Online-00B89E?style=for-the-badge&logo=internet-explorer&logoColor=white" alt="Website"></a>
</p>
## Abstract
Foundation models for echocardiography often struggle to disentangle anatomical signal from the stochastic speckle and acquisition artifacts inherent to ultrasound. We present EchoJEPA, a foundation model trained on 18 million echocardiograms across 300K patients, representing the largest pretraining corpus for this modality to date. By leveraging a latent predictive objective, EchoJEPA learns robust anatomical representations that ignore speckle noise. We validate this using a novel multi-view probing framework with frozen backbones, where EchoJEPA outperforms state-of-the-art baselines by approximately 20% in left ventricular ejection fraction (LVEF) estimation and 17% in right ventricular systolic pressure (RVSP) estimation. The model also exhibits remarkable sample efficiency, reaching 79% view classification accuracy with only 1% of labeled data versus 42% for the best baseline trained on 100%. Crucially, EchoJEPA demonstrates superior generalization, degrading by only 2% under physics-informed acoustic perturbations compared to 17% for competitors. Most remarkably, its zero-shot performance on pediatric patients surpasses fully fine-tuned baselines, establishing latent prediction as a superior paradigm for robust, generalizable medical AI.
<p align="center">
<img src="assets/echo_fig1a.png" width=100%>
</p>
EchoJEPA models trained on just 1% of labeled data outperform baselines trained on 100%. This efficiency implies that latent prediction yields dense representations capable of defining the view manifold with minimal supervision, as evidenced by the distinct anatomical clustering in the figure below.
<p align="center">
<img src="assets/umap_views.png" width=100%>
</p>
EchoJEPA demonstrates anatomical localization, focusing on the mitral valve leaflets, ventricular walls, and annulus while ignoring sector background. Received attention clusters at Doppler jet edges while given attention localizes on valve structures generating flow. Across the cardiac cycle, focus shifts from valve tips during opening to chamber walls during relaxation, indicating it interprets the echocardiogram as a functional biological system.
<p align="center">
<img src="assets/echo_attention.png" width=100%>
</p>
## Getting Started
### EchoJEPA Checkpoints
Our trained EchoJEPA checkpoints are available for download [here](https://drive.google.com/drive/folders/1RFEXMe8TTcABMBz4H_qtiLB43K9jD_lf?usp=sharing). We provide two main variants:
1. **EchoJEPA (V-JEPA 2)** — initialized from the pretrained V-JEPA 2 weights (VideoMix22M: SSv2, Kinetics, HowTo100M, YT1B, and ImageNet), which we have found works much better for downstream functional tasks compared to training on MIMIC from scratch.
2. **EchoJEPA (V-JEPA 2.1)** — based on [V-JEPA 2.1](https://arxiv.org/abs/2603.14482), which adds dense predictive supervision over all tokens (not just masked ones) plus deep self-supervision across multiple encoder layers. This produces spatially structured, temporally consistent features that are better suited for dense prediction tasks such as segmentation, tracking, depth estimation, and interpretability (e.g., attention map visualization).
Checkpoint naming: `pt` = pretraining epochs, `c` = cooldown/annealing epochs (which help with convergence and resolutExcerpt of 20,339 characters
Read on GitHubWould you bet a product on this? Bounded 0–100 and slow moving.
matched fp:7036b811d6d10a0a, llm:Repository description: "EchoJEPA: A Latent Predictive Foundation Model for Echocardiography" (Python) — indicates a foundation model for medical imaging (echocardiography).
matched fp:7036b811d6d10a0a, llm:Repository description: "EchoJEPA: A Latent Predictive Foundation Model for Echocardiography" (Python) — indicates a foundation model for medical imaging (echocardiography).
matched fp:7036b811d6d10a0a, llm:Repository description: "EchoJEPA: A Latent Predictive Foundation Model for Echocardiography" (Python) — indicates a foundation model for medical imaging (echocardiography).