Neuvara

Neuvara — MRI physics & machine learning

Your model was trained on some scanners. It will be used on others.

We test how brain MRI models behave across different scanners, and we do the multi-site data work that has to happen before training. For imaging AI teams with more data than time.

Discuss a project See the audit
Scanner A · 3T Scanner B · 1.5T

Intensity along one row

Scanner A Scanner B
Same anatomy, different numbers. Synthetic illustration, not patient data.

The problem

MRI is not a calibrated measurement.

Scan the same person on two machines and you get two different images. Brightness, contrast, noise and fine detail all shift with the scanner. A model trained on one hospital's machines learns those quirks along with the anatomy, and its accuracy can drop when it meets a scanner it has not seen.

What changes between scanners

  • VendorSiemens, GE and Philips reconstruct images differently, before anyone sees them.
  • Field strength1.5T and 3T produce different contrast and noise for the same tissue.
  • ProtocolSequence timing, resolution and acceleration vary from site to site.
  • The machine itselfField inhomogeneity, gradient nonlinearity and drift over time.

01Cross-scanner audit Flagship

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

We run your model on travelling-subject brain MRI, the same people scanned on several scanners, plus any held-out data you have, and show you where its measurements shift, on which scans, and why.

See the audit

In short

  • ForImaging AI teams preparing a regulatory submission or a new deployment site.
  • You getConsistency by scanner and structure, the failure cases, their likely physical causes, and a written report.
  • ContextYour model compared with widely used open models on the same scans, so you know what is normal.
  • Your modelSent as a container that runs offline and is deleted afterwards, or run inside your environment on public data.
  • PricingFixed price, agreed in writing before we start.

02Multi-site data preparation

Get messy multi-site imaging into a state you can train on.

Data from several hospitals rarely arrives consistent. Sequences are mislabelled, protocols drift, and some scans should never have been included. We sort it out and leave you with a pipeline you can rerun.

Discuss data work

What we handle

  • Sorting and conversionDICOM sorted by sequence and protocol, converted to NIfTI or BIDS.
  • Quality controlMotion, artefacts, wrong sequences, duplicates and label errors, flagged per scan.
  • Dataset auditWhich scanners, sites and protocols you actually have, and where the gaps are.
  • Preprocessing pipelinesReproducible, documented, and handed over as code your team owns.

Why physics

Scanner differences have physical causes.

Field strength, B1 inhomogeneity, gradient nonlinearity, vendor reconstruction, noise from parallel imaging. Each one changes the image in a specific way. We look for those causes instead of treating every shift as noise, which is why our reports say what went wrong rather than only how much.

Alongside client work we are building tooling for scanner-robust MRI measurement.

Who we are

  • Baris Ganidagli
    Co-founder Background in pharmacy
  • Joshua Stromayer
    Co-founder Background in physics
  • Emaan Kamran
    Research Engineer

Contact

Tell us what you are working on.

A few lines on the model or dataset and what you need is enough to start. Everything runs in writing, no calls needed.

Get in touch