What is echocardiography?
How strongly the heart squeezes — ejection fraction.
Leak and narrowing across all four valves.
Chambers, wall thickness, and muscle disease.
Portable ultrasound, anywhere a patient is.
The most common cardiac imaging test — millions of studies each year, every one dozens of videos a clinician must review.
Research direction
Spatiotemporal neural networks that measure structure and function from raw echo video.
Foundation models clinicians can inspect, trust, and correct.
Randomized trials, FDA clearance, and health-system deployment.
Progress, year by year
Increasing the scale of training data · log scale
Per-model training data — from EchoNet-RV to EchoPrime, across 12 studies (2020–2026).
What our models cover — and the open gaps
From open science to the clinic
Open datasets and models that reproduce expert measurements and reveal subclinical disease — proven in randomized trials, FDA-cleared, and deployed in routine care.
OUYANG LAB · development → deployment → implementation · douyang.github.io
A self-playing overview of the Ouyang Lab's cardiovascular-AI research. Bar heights are the number of echocardiogram videos used to train each model — view- and quality-filtered training sets, not full study counts. Task coverage is mapped to the ASE standardized echo report and its assessment & quantification guidelines; colored items are published or released as open code, gray items are open gaps not yet automated.