Virgo: 1,053,880 US procedures across 148 centers (Sept 2025 audit). Public dataset demographics from each dataset's published documentation. Diversity index = Shannon entropy across White / Black / Asian / Hispanic / Other; higher is more balanced.

Out-of-distribution validation · external datasets

Validated across new sites, scopes, and disease areas

HyperKvasir

Colonoscopy

Mayo Endoscopic Scoring

Bærum Hospital, Norway

Kvasir-Capsule

Capsule endoscopy

Lesion / anatomy classification

Bærum Hospital, Norway

CholecT50

Laparoscopy

Surgical action triplets

IHU Strasbourg, France

SUN

Colonoscopy

Polyp detection

Showa University, Japan

UNIFI Phase 3

Colonoscopy

Endoscopic healing prediction

Janssen multi-site trial

YODA / external UC cohort

Colonoscopy

MES (QWK 0.83)

Independent academic centers

Competitors (e.g. DovaVision UC, Iterative Health) typically train and evaluate on a single internal data source. EndoDINO is pre-trained on Virgo's corpus and validated on independent public benchmarks and external clinical trial cohorts.

148

US medical centers

vs. 1–6 in public datasets

1.05M

US procedures with demographics

procedure-weighted, not patient-weighted

46.6%

non-White representation

12.9% Black · 15.1% Asian · 13.9% Hispanic

0.713

Shannon diversity index

Hyper-Kvasir 0.28 · LDPolypVideo 0.10

HyperKvasir · 4-class Mayo Endoscopic Scoring

State-of-the-art with frozen features.

EndoDINO ViT-g/14 delivers leading performance on Mayo endoscopic scoring with a frozen backbone.

Model: EndoDINO ViT-g/14
Data: 130K+ procedures

Endoscopic lumen view representing Mayo 0: Normal or inactive disease — Mayo 0

Normal or inactive disease
Conf 0.97

Endoscopic lumen view representing Mayo 1: Mild disease — Mayo 1

Mild disease
Conf 0.92

Endoscopic lumen view representing Mayo 2: Moderate disease — Mayo 2

Moderate disease
Conf 0.90

Endoscopic lumen view representing Mayo 3: Severe disease — Mayo 3

Severe disease
Conf 0.91

Frame-level predictions aggregated per procedure.

Scored by EndoDINO • Inference latency 14.8 ms.

Macro F1 0.748 · Linear probe on frozen backbone

How EndoDINO learns

From raw procedure video to a shared representation layer for GI.

One model. Every downstream task: scoring, detection, prediction, biomarker discovery.

Capture

Raw endoscopy video from the procedure stream.

Structure

Frames organized, deduplicated, temporally aligned.

Pretrain

Self-supervised learning at population scale.

Represent

A reusable embedding for any downstream task.

Capabilities

One representation layer. Many downstream tasks.

Proof point · UNIFI, Phase 3 UC

7 mo

Saved

$38M

Avoided

Validated on Stelara Phase 3 UC trial data. UNIFI could have reached the same readout faster and at lower cost using EndoDINO as a covariate.

01

Placebo response

Covariate models that reduce trial size and accelerate enrollment.
02

Subgroup response

Precision enrichment: identify likely responders before randomization.
03

Continuous AI scores

UC and CD efficacy assessment beyond Mayo and SES-CD categories.
04

Bayesian priors

Real-world evidence priors from EndoDINO at population scale.

The data moat

Endoscopy video is the substrate. We capture more of it, from more procedures, than anyone else. The archive grows every day.

Capture is the foundation of everything downstream. Real-world endoscopy video (at population scale, longitudinal, and continuously growing) is what makes a foundation model for GI possible. Models built on smaller datasets plateau. Ours compound.

3M+ Procedures recorded across partner sites through 2025
1M+ New procedures captured each year, and growing
3.5B Video frames already in the EndoDINO training set
24/7 Live capture pipeline across institutional partners

The platform

One model base. Built for the full procedure.

Foundation model

EndoDINO

Virgo's foundation model for endoscopy. One model base for scoring, prediction, detection, and biomarker work, trained on the full procedure, not just the frame.

Build environment

EndoML

The environment for building on top of EndoDINO. A GI-specific model layer for clinical and research workflows.

Request access

See the full evidence package.

Manuscript, UEGW 2025 poster, benchmark results, and partnership models. Sent directly to qualified researchers and partners.

Infrastructure for the future of endoscopy AI.

Contact research Read the platform overview

Virgo builds frontier AI models to solve colorectal cancer.

Press & podcast coverage

The most powerful AI for endoscopy. Trained on the largest dataset.

Largest endoscopy video dataset in the literature

57× more frames than the next largest dataset

State-of-the-art on every benchmark. Validated across sites, scopes, and patient populations.

State-of-the-art with frozen features

Predicting 8-week endoscopic healing from baseline video

The most demographically representative endoscopy dataset

Validated across new sites, scopes, and disease areas

State-of-the-art with frozen features.

From raw procedure video to a shared representation layer for GI.

Capture

Structure

Pretrain

Represent

One representation layer. Many downstream tasks.

Placebo response

Subgroup response

Continuous AI scores

Bayesian priors

Endoscopy video is the substrate. We capture more of it, from more procedures, than anyone else. The archive grows every day.

One model base. Built for the full procedure.

EndoDINO

EndoML

See the full evidence package.

Infrastructure for the future of endoscopy AI.

The most powerful AI for endoscopy.
Trained on the largest dataset.

State-of-the-art on every benchmark.
Validated across sites, scopes, and patient populations.