Neuvara

01Cross-scanner audit

Find out where your model breaks before a new site does.

A model that scores well on its validation set can drop sharply on a scanner it has not seen. We test your brain MRI model on the same people scanned on different scanners, then show you where its measurements shift, on which scans, and what is likely causing it.

Request an audit How it runs

Why it matters

Most validation sets come from a handful of scanners.

Training and test data usually share the same hospitals and the same machines, so a strong test score says little about the next site. The gap shows up later, during a deployment or in front of a regulator, when it is expensive to fix. An audit finds it while you can still act.

When teams ask for one

  • Before a submissionEvidence of how the model generalises across sites and scanners.
  • Before a new siteCheck a hospital's scanners before the model goes live there.
  • After a dropPerformance fell somewhere and you need to know why.
  • Before retrainingLearn which scanners and protocols your next round of data should cover.

Why not do it in-house

Running the model is the easy part.

Any capable engineer, or an AI coding assistant, can run a model over a folder of scans and print a number for each scanner. Whether that number means anything depends on everything around it.

Quick in-house check Neuvara audit
The scans
Whatever multi-site data is to hand, often one public set.
Curated multi-vendor and travelling-subject data, matched across scanners and checked scan by scan.
What normal looks like
Your model's numbers on their own.
Your model next to widely used open models on the same scans, so a real weakness stands out from ordinary scanner drift.
Why it moved
"Performance dropped on GE 1.5T."
Each shift linked to the scan's acquisition settings and image quality, and explained by an MRI physicist, so you know what is likely responsible.
The traps
Easy to miss: preprocessing mismatches, orientation and header errors, subjects shared between sets.
Checked on every audit, because each one can quietly invalidate a result.
The output
A notebook that makes sense to the person who ran it.
A report your clinical and regulatory leads can use as it is.

What you get

Answers you can act on, not one pooled score.

Every result is broken down so your team can see which scanners are the problem and what to do about them.

  • Consistency by scannerBroken down by brain structure, vendor and scanner, with uncertainty intervals. This measures consistency: whether the same brain gets the same result on every scanner. With your own labelled data, accuracy too.
  • Reference comparisonWhere your model sits against widely used open models, scanner by scanner.
  • The failure casesThe individual scans where the model goes wrong, with the images, so your team can look.
  • Likely physical causesBias field, contrast, resolution, reconstruction. We explain which are the likely drivers of each shift.
  • RecommendationsWhat to change: data to collect, preprocessing to add, sites to watch.
  • A written reportClear enough to hand to your clinical or regulatory lead as it is, plus a recorded walkthrough.

Rankings

Every result, in context.

Your model is compared with reference models tested on the same scans, so you can tell an unusual weakness from drift every model shares, and see where you are ahead.

Headline score · top 5

Preview · placeholder values

See the full rankings

How it runs

Five steps, priced up front.

  1. 1

    Brief

    You fill in a short brief: the model, the task, the metric, and what you want answered.

  2. 2

    Fixed quote

    We reply in writing with scope, price, timeline and an NDA. Questions are handled by email.

  3. 3

    Access

    You send a container, after checking it with our validator script, or we run inside your environment. No patient data comes to us.

  4. 4

    Testing

    We run the model on public multi-vendor and travelling-subject data, and your held-out data if you have it.

  5. 5

    Report

    The written report, the results tables, and a recorded walkthrough. Follow-up questions by email.

What we need from you

Less than you might think.

We do not need your training data, and your model does not have to leave your control.

  • The modelAs a container that takes a scan and returns a label map, or access to run it inside your environment.
  • Its inputsThe preprocessing it expects, so we feed it scans the way you do.
  • The taskWhat it predicts and the metric you report it on.
  • OptionalHeld-out data from your own sites, kept in your environment.

Current scopeT1-weighted brain MRI · 3T · Siemens, GE, Philips. Segmentation and volumetric models. 1.5T in development; other sequences by discussion.

Confidentiality

Your model and your data stay yours.

Every audit runs under an NDA. We do not keep your model, your data or anything derived from your own scans once the work is finished.

With your written permission, we keep results from the public-dataset part of an audit, anonymised, for our own research and as anonymised reference points when comparing other models. We never publish anything that identifies you or your model, and you can say no.

  • Under NDAEvery audit, from the first brief.
  • Nothing keptYour model, your data and anything from your own scans are deleted when the work ends.
  • RankingsYour model is only listed publicly if you choose, under your name or anonymised.
  • Not a certificationTechnical evidence your regulatory team can use. Neuvara does not provide medical devices or regulatory approval.

Next step

Start with a short brief.

Tell us about the model and what you need to know. We reply in writing with a scope and a fixed price. No calls needed.

Send a brief