Foundation Model Nature 2026 JACC 2025 In development

InVision Precision Reporting

Automated, structured echocardiographic reports built on two complementary foundation models — EchoPrime and EchoNet-Measurements.

Structured Echo Report 18 PARAMS MEASUREMENTS · DRAFT WORDING Left ventricle EF 58% LVIDd 4.8 IVSd 1.0 GLS -19% Normal cavity size and wall thickness with normal systolic function. No regional wall motion abnormality. Right ventricle TAPSE 2.1 S' 12 RVSP 28 Normal RV size and systolic function. No evidence of elevated pulmonary pressures. Mitral valve MR trace E/A 1.2 E/e' 7 Trace mitral regurgitation. Leaflets thin and mobile, no stenosis. Normal diastolic filling pressures. Aortic valve + DOPPLER · IVC AI DRAFT

What it does.

Cardiologists and sonographers spend more than an hour per echocardiography exam on acquisition, measurement, and interpretation. InVision Precision Reporting automates the routine parts of that workflow — producing structured report drafts that integrate quantitative measurements with holistic multi-view interpretation.

Precision Reporting is powered by two complementary open-source foundation models. EchoPrime, published in Nature (2026), is a multi-view, video-based vision-language foundation model trained on over 12 million video-report pairs. It integrates information across all standard echocardiographic views in a comprehensive study and delivers holistic clinical interpretation, achieving state-of-the-art performance across 23 diverse benchmarks of cardiac form and function — validated in five international healthcare systems including Kaiser Permanente Northern California, Cedars-Sinai, Stanford Healthcare, Beth Israel Deaconess, and Chang Gung Memorial Hospital in Taiwan.

EchoNet-Measurements, published in JACC (2025), handles the quantitative substrate. Trained on 877,983 sonographer annotations from 155,000+ studies, it automates 18 standard B-mode and Doppler parameters — ejection fraction, chamber dimensions, wall thickness, Doppler velocities, TAPSE, and more — with accuracy comparable to expert sonographers (R² of 0.97 on internal test data and 0.99 on Stanford external validation).

12M+
Video-report pairs used to train the EchoPrime foundation model
23
State-of-the-art benchmarks of cardiac form and function surpassed
18
Automated B-mode and Doppler measurements from EchoNet-Measurements

Every number is yours to change.

Most automated echo output arrives as pixels. A measurement burned into a DICOM image is a picture of a number — the cardiologist can read it, but cannot correct it, and nothing downstream recomputes when they disagree. Correcting it means measuring again by hand and retyping the result.

InVision measurements are not burned into the image. They are written natively into the structured report in your PACS and reporting platform, in the same fields your sonographers and cardiologists already edit. They arrive as data, not as a rendering of data.

That difference is what makes the final read genuinely the cardiologist's. Adjust a measurement and the derived values recompute from your edit. The draft impression is re-checked against the change and flagged if it no longer agrees with the measurements. Nothing is rewritten behind you, and every report carries a complete edit trail of what the model proposed and what the physician changed.

Delivered as
Discrete values in the structured report fields of your PACS and reporting platform — not burned-in pixel overlays
Editable by
The sonographer and the final reporting cardiologist, in the tool they already use — no second application, no re-measuring by hand
On edit
Derived values recompute from the edited measurement; the draft impression is re-checked and flagged if it no longer agrees
Audit
Full edit trail of the model's proposal and the physician's change, retained with the signed report
The software drafts. The physician authors. A measurement you cannot change is a measurement you have to trust. One you can change — in place, with everything downstream recomputing — is a starting point.

Peer-reviewed research behind the product.

Built for your existing workflow.

Indication
Automated echocardiographic report generation and measurement
Input
Comprehensive multi-view transthoracic echocardiogram studies
Output
Structured interpretive report; 18 automated B-mode and Doppler measurements
Foundation models
EchoPrime (Nature 2026) + EchoNet-Measurements (JACC 2025)
Accuracy
R² 0.97 – 0.99 vs. expert sonographer measurements
Validation
Internal + external validation across 5 international health systems