AI-assisted echo core lab

Core lab reads with AI enablement, signed by human experts.

InVision's echocardiography core lab pairs peer-reviewed state-of-the-art AI first-pass measurement and interpretation with final adjudication sign-off by expert sonographers and cardiologists — compressing turnaround while tightening the reproducibility that clinical decisions and trial endpoints depend on.

Peer-reviewed in
Nature JACC Nature Medicine Circulation JAMA Cardiology
One study · one pipeline LIVE READ
01 DICOM intake & QCDe-identify · classify views · triage quality Automated
02 AI first-pass18 measurements + holistic draft read EchoNet · EchoPrime
03 Expert adjudicationSonographer verifies · cardiologist over-reads RDCS / RCS + MD
04 Signed & deliveredLocked · versioned · audit-ready Final sign-off
AI pre-read → expert final read Signed
16.8%
AI-initiated reads substantially changed by the cardiologist — vs 27.2% for sonographer reads
Nature 2023 RCT
0.970.99
R² of AI vs expert sonographer measurements across 18 parameters
EchoNet-Measurements · JACC 2025
12M+
Video–report pairs behind the foundation model that drafts each read
EchoPrime · Nature 2025
100%
Reads reviewed and finalized by an expert cardiologist
Human-in-the-loop

Echo reads vary. Patients and trials are impacted by the variance.

Quantitative echocardiography is the workhorse of cardiac imaging, yet the same study read by two qualified humans — or by the same human twice — can yield materially different numbers. For a clinical decision that's noise; for a trial endpoint it's lost statistical power and slipped timelines. Our AI-enabled core lab exists to remove that variance.

FAILURE MODE 01

Inter- and intra-reader drift

Manual LVEF and Doppler quantification carry real reader-to-reader and test–retest variability. In a randomized trial, expert sonographer reads required substantial cardiologist correction in 27.2% of studies while AI reads required adjustment in 16.8% of studies.

Nature 2023 · He et al.
FAILURE MODE 02

Turnaround that gates decisions

Comprehensive interpretation consumes an hour of combined technician and physician time per exam. Backlogs delay cardio-oncology surveillance, heart-failure titration, and trial interim analyses.

> 1 hr clinician time / exam
FAILURE MODE 03

Cost and scale ceilings

Throughput is bounded by scarce expert time. Scaling a traditional core lab means linearly adding readers — and re-introducing the very variance a core lab was meant to eliminate.

Throughput ∝ expert headcount

AI does the first pass. Experts do the deciding.

Every study moves through the same five-stage pipeline. AI handles the repetitive, time-intensive measurement work at machine speed and consistency; credentialed humans own judgment, edge cases, and the signature. No read leaves the lab without a cardiologist's name on it.

01

Intake & quality control

Automated

Studies arrive as standard DICOM from any scanner or PACS. The pipeline auto-de-identifies (HIPAA-safe), classifies every view, and flags inadequate image quality before measurement — the gatekeeping a trial run-in period provides, applied to every study.

02

AI first-pass measurement & draft read

EchoNet-Measurements + EchoPrime

Semantic-segmentation models auto-measure 18 standard B-mode and Doppler parameters; a multi-view vision–language foundation model produces a holistic, study-level draft interpretation across chambers, valves, and function. Deterministic by construction — the same pixels always yield the same numbers.

03

Sonographer adjudication

RDCS / RCS

An expert cardiac sonographer reviews every AI tracing and value against the images, corrects where clinical judgment differs, and confirms protocol adherence. The AI starts the work at a high baseline; the sonographer guarantees it.

04

Cardiologist finalization & sign-off

Level III reader

A board-certified cardiologist over-reads the adjudicated study, finalizes the interpretation, and signs. In randomized testing, cardiologists made fewer substantial changes to AI-initiated reads and over-read them in less time than sonographer-initiated reads.

05

Delivery, locking & audit trail

Final

The signed report is delivered as structured data, then locked and versioned with a complete edit history — designed for 21 CFR Part 11-aligned audit trails, blinded read workflows, and direct integration with sponsor EDC, your PACS, and Epic.

Faster to a finalized read — enabled by AI.

The slowest part of an echo read is the measurement and report generation, not the judgment. By moving measurement and first-draft interpretation upstream to AI, the human starts from a near-complete read and spends their time verifying rather than constructing. The randomized evidence is directional and clear: AI-initiated reads took cardiologists less time to finalize.

Traditional core labManual measure → over-read
measurement-bound · serial human steps
InVision AI-assistedAI pre-read → expert sign-off
verify, don't construct

Illustrative of where time is spent in the workflow, not a published trial endpoint. Service-level turnaround is committed per engagement — typical clinical reads target 48–72 hours with premium processing; trial reads run to protocol, with AI pre-reads accelerating site-query resolution and interim analyses.

>1hr
Combined physician + technician time a full interpretation consumes today — the budget AI first-pass reclaims.
Workflow baseline
18
Standard B-mode and Doppler parameters auto-measured before a human opens the study.
EchoNet-Measurements
Less
Cardiologist over-read time for AI-initiated vs sonographer-initiated reads in randomized testing.
Nature 2023 RCT
48–72h
Target turnaround for routine clinical reads with premium processing — a committed service level, not a published result.
Service commitment

The same study reads the same way, every time.

Reproducibility is the core lab's product. A deterministic AI first-pass dramatically reduces test–retest variability at the measurement step by construction — identical input yields identical output — and anchors every human reader to the same starting point. Across 18 parameters, AI measurements matched expert sonographers with an R² of 0.97–0.99; in randomized testing, AI-initiated reads were also more concordant with prior independent cardiologist reads.

Conventional independent reads

Wider spread · 27.2% substantially changed
± wide reader variance

Reader-to-reader and test–retest drift widen the measurement band — the variance a sponsor's power calculation must absorb.

AI first-pass + expert sign-off

Tighter band · 16.8% substantially changed
tight band · R² 0.97–0.99

A deterministic starting point plus expert sign-off collapses the band — fewer corrections, higher concordance, less noise to power through.

0.97–0.99
R² of AI vs expert sonographer across 18 parameters (CSMC 0.967; Stanford external 0.987).
EchoNet-Measurements
16.8%
AI reads substantially changed by the cardiologist — vs 27.2% for sonographer reads.
Nature 2023 RCT
−10.4pts
Reduction in substantial-change rate (95% CI −13.2 to −7.7; P<0.001 non-inferiority and superiority).
Nature 2023 RCT
Minimal
Test–retest variability at the AI measurement step — identical input returns identical output.
Deterministic inference

The reads rest on a decade of peer-reviewed research.

The core lab is not a wrapper around a closed model. It runs on the published, externally validated, open-source research that defined AI in echocardiography — including the first and only randomized trial of AI in cardiology. The headline numbers are reported exactly as the journals published them.

Nature 2025 Foundation model Open source

EchoPrime

A multi-view, view-informed, video-based vision–language foundation model that synthesizes every view in a comprehensive study into one holistic interpretation — the engine behind the draft read. Contrastive learning plus retrieval-augmented interpretation weight each video by anatomical relevance, the way a cardiologist does.

Training scale
12M+ pairs
Benchmarks
SOTA on 23
External validation
5 health systems
Performance
Mean AUC ≈ 0.92
JACC 2025 Quantification Open source

EchoNet-Measurements

Deep-learning semantic-segmentation models that automate the measurement work itself — 9 B-mode and 9 Doppler parameters, from LV internal diameter and septal thickness to TR Vmax and septal e′. Validated against expert sonographer annotations at two academic centers, with accuracy on par with the humans it accelerates.

Annotations
877,983
Studies / patients
155,215 / 78,037
Parameters
18
Accuracy (R²)
0.967 – 0.987
Nature 2023 Blinded · randomized NCT05140642

The first randomized trial of AI in cardiology

A blinded, randomized, non-inferiority trial of AI versus sonographer initial assessment of LVEF, with final interpretation by a blinded cardiologist. From 3,769 studies screened, 3,495 analyzable transthoracic echocardiograms were independently re-read by 25 sonographers (mean 14.1 years' experience) and 10 cardiologists (mean 12.7 years). The AI-initiated arm was both non-inferior and superior on the primary endpoint — and cardiologists, unable to reliably tell which arm they were reading, changed AI reads less often and finalized them faster.

16.8%
AI reads substantially changed by cardiologist (primary endpoint)
27.2%
Sonographer reads substantially changed
P<0.001
For both non-inferiority and superiority
3,495
Analyzable TTEs · single individually-randomized trial

Selected publications

References

  1. He B, Kwan AC, Cho JH, et al. Blinded, randomized trial of sonographer versus AI cardiac function assessment. Nature. 2023;616:520–524. ClinicalTrials.gov NCT05140642. PubMed 37020027.
  2. Vukadinovic M, Chiu I-M, Tang X, et al. Comprehensive echocardiogram evaluation with view-primed vision–language AI. Nature. 2025. PubMed 41219498.
  3. Sahashi Y, Ieki H, Yuan V, et al. Artificial intelligence automation of echocardiographic measurements. J Am Coll Cardiol. 2025;86(13):964–978. PubMed 40914895.
  4. Christensen M, Vukadinovic M, Yuan N, Ouyang D. Vision–language foundation model for echocardiogram interpretation. Nat Med. 2024;30:1481–1488. PubMed 38689062.
  5. Ouyang D, He B, Ghorbani A, et al. Video-based AI for beat-to-beat assessment of cardiac function. Nature. 2020;580:252–256. PubMed 32269341.

A core lab built for endpoints.

In a cardiovascular trial, echo endpoints are only as good as their measurement variance. Lower variance is not a nicety — it directly raises the signal-to-noise ratio for detecting a treatment effect, and because required sample size scales with the square of the endpoint's standard deviation, tightening the read can lower enrollment or raise power. A deterministic AI first-pass, adjudicated by blinded experts, is a direct lever on that variance.

ENDPOINTS

The measures trials live on

LVEF and GLS for cardiotoxicity and heart-failure trials; diastolic parameters; valve severity; chamber dimensions; RV function; estimated pulmonary pressures — quantified consistently across every site and every visit.

STATISTICAL POWER

Variance you can bank

A deterministic first-pass removes test–retest noise at the measurement step and anchors blinded adjudicators to one baseline, shrinking endpoint variance — fewer patients for the same power, or more power for the same N.

SPEED

Reads that don't gate the DSMB

AI pre-reads compress the measurement bottleneck, accelerating site-query resolution, interim analyses, and database lock — without trading away the blinded, expert-finalized rigor a regulator expects.

INTEGRITY

Blinded, locked, traceable

Blinded read workflows, locked and versioned outputs, and a complete edit trail behind every signed value — designed for 21 CFR Part 11-aligned audit trails and direct delivery to sponsor EDC.

REPRODUCIBILITY

Read it the same in year three

The same version-locked model and protocol re-read a baseline study identically two years later — eliminating drift across long enrollment windows and between interim and final analyses.

PROVENANCE

Evidence a reviewer recognizes

The underlying models are published in Nature, JACC, and Nature Medicine and open-sourced — so the methods behind your endpoints are inspectable, citable, and externally validated rather than a black box.

AI assists. Experts decide.

The model accelerates the read; it never owns it. Accountability stays with the credentialed humans whose names sign the report.

AI never signs a read

Every measurement is verified by an expert cardiac sonographer and every interpretation is finalized by a board-certified cardiologist. The AI output is a draft and a starting point — never the deliverable.

Credentialed readers, defined competency

Reads are performed by RDCS/RCS-credentialed sonographers and Level III-equivalent cardiologist readers, with documented training and ongoing concordance monitoring.

The human can always overrule

When clinical judgment differs from the AI, the human value wins and the divergence is logged. The edit trail captures exactly what the model proposed and what the expert changed.

Quality control on the loop

Sampled over-reads, inter-reader concordance tracking, and model-drift monitoring keep both the AI and the humans inside agreed performance bounds over time.

“We aspire to democratize expert knowledge to more clinicians, and ensure that excellent care is available to every patient — in every clinical setting.”

Dr. David Ouyang, MD
Co-founder, InVision Medical Technology · practicing cardiologist

Bring your studies to a faster, more reproducible core lab.

Schedule a 30-minute briefing with our clinical and operations teams. We'll walk through read protocols, turnaround commitments, the evidence base, and how blinded, expert-finalized AI reads fit your service line or trial — tailored to your endpoints.

Request a briefing

Your email client will open a message to our team — tell us whether this is for clinical reads, a clinical trial, or both.