Internal clinical validation

You should be skeptical of medical AI.

So are we. That's why we publish our numbers, methods and limits, instead of asking you to take “clinically proven” as a fact. A veterinarian still reviews and signs every report.

Explore the data Request the validation report

The numbers, in context

High accuracy, nothing hidden.

On a retrospective audit of 4,513 production cases, diagnoVET reaches 93.9% sensitivity and 99.1% specificity, with ground truth established by board-certified radiologists. We validated on ultrasound and echocardiography, the modalities most veterinary imaging tools still avoid. We treat that as a starting point to scrutinize, not a trophy.

4,513 cases · 40 pathologies · ground truth by board-certified radiologists

Sensitivity

Audited across 4,513 cases

93.9%

Specificity

40 pathologies assessed

99.1%

diagnoVET is a second pair of eyes, not a replacement. The veterinarian reviews every finding and signs every report.

What we publish

Everything most vendors hide.

The black box is the real barrier to AI in the clinic. So every finding is built to be verified by a professional in seconds, and our own caveats are stated out loud.

Natural Language Explanations

Every finding carries a plain-language rationale describing the organ characteristics behind it, not just a label.

Differentials

When the diagnosis is not certain or requires confirmation, diagnoVET provides a list of differentials to evaluate. The qualifier (questionable, probable, presumed, definite) is carried with each finding.

Measurement linking

Cited measurements link back to the original ultrasound stills for immediate anatomic verification.

Show Highlights

One tap surfaces the anomalies behind each observation. Nothing is based on blind faith.

How we measured it

Including how our own study could be wrong.

Disclosed limitation

Our ground truth is the radiologist's signed report, written after they saw the platform's draft. That means the system influenced the benchmark it is judged against. We say so plainly, and we mitigate it: every reviewer is a board-certified radiologist who performs a full primary read of the images before verifying the draft.

Performance compares the platform's preliminary findings against the final signed report, on a validation set with zero overlap with training data. Confidence intervals use the Wilson score method at 95%. The report answers, point by point, the published critique of commercial veterinary imaging AI (Ma et al., JAVMA 2024).

Want the audit-level methodology behind these numbers? Talk to the team.

Talk to the team

Who is in command

The veterinarian signs. Always.

diagnoVET drafts in seconds. The professional decides. The workflow is built so the human has the last word, on the record.

01
Case arrives in the inbox
DICOMs route in directly from your DICOM server, or drag in DICOMs, MP4 cineloops, and standard images. Mixed-source uploads are supported in a single case, up to 100 images and 20 videos.
02
Draft prepared in background
Models trained on veterinary data return a structured draft with findings, measurements, and Natural Language Explanations. The next case starts processing while the veterinarian works on this one.
03
Review and edit
The veterinarian reviews, edits, and confirms every finding. Every drafted field is editable; per-field assistance is opt-in. The vet holds final diagnostic authority.
04
Signed report delivered
The signed PDF is downloaded for the owner and written into the clinic's patient history. A push notification confirms delivery to subscribed devices.

How it can be trusted

Posture, not just performance.

Numbers are one piece of trust. These are the operating principles the platform is built on, today.

Drafts, not decisions: Every report is a draft until a veterinarian signs it. The platform does not deliver an unsigned report to anyone.
Editable at field granularity: Every drafted field can be rewritten. Per-field assistance is opt-in, not automatic.
Conservative on uncertain views: Findings are produced where the images support them. Indeterminate views are surfaced as indeterminate, not pushed into a category they don't belong in.
Auditable: Every change to a report is logged with user, time, and a field-level diff. The history is exportable.
Workspace isolation: Cases, templates, findings libraries, and signed reports belong to the workspace. No cross-workspace leakage.
Data residency posture: Studies are stored in a cloud environment dedicated to clinical workloads. Signed URLs are used for any external share.