Multi-site brain MRI has a measurement problem

Scanner differences can produce variation comparable in size to the biological change a study is trying to measure. Neuvara works on physics-informed normalisation: reducing scanner-driven variance while preserving biological signal.

UK-based · Research-stage · Two founders

Same subject, same weekTwo scanners
Scanner A
Scanner B

Measured volume

Scanner A
Scanner B

Anatomy unchanged. The measurement is not.

Illustrative only. Not real patient data, and not a Neuvara result.

The problem

The same patient does not produce the same numbers.

Scanned on two different machines, one subject yields two different measurements. That blocks pooling across sites, limits how far a trained model generalises, and makes change over time harder to trust.

What varies

Multi-site brain MRI is affected by scanner manufacturer, field strength, sequence timing, protocol, and software version. Those differences change measured volumes and image-derived features even when the underlying anatomy is unchanged.

Why it matters

In longitudinal studies the signal of interest is often small. Annual brain atrophy in neurodegenerative disease is a subtle change. A scanner change can be larger than that change.

The tension

If scanner variance is left in the measure, it can obscure biology. If correction removes biological signal along with it, the measure looks cleaner while becoming less useful.

Approach

Whether biological signal survives the correction.

The question is not only whether scanner information can be reduced. A method that removes scanner variance and biology together is easy to build and looks tidy. It is not useful.

Physics first

Scanner differences are physical in origin: field inhomogeneity, gradient nonlinearity, vendor reconstruction, noise structure from parallel imaging. We model those mechanisms rather than applying statistical correction after the fact.

Fixed function, not fitted to the cohort

The mapping is intended to be inductive: the same scan produces the same output regardless of which other scans it is processed alongside. Many existing methods refit against the cohort, so a previously reported value can move when new subjects arrive.

A stated operating envelope

We are not trying to build a universal normaliser. The intended scope is written down as a specification — sequence, field strength, vendor — and widened one axis at a time, with validation attached to each step.

Where we are

A UK research-stage company, run by two founders.

We have no product, no deployments, and no published results yet. Everything on this page describes work in progress, not a capability we are offering today.

United Kingdom

One private limited company

Research-stage

Method development and benchmarking

Two founders

MRI physics and clinical translation

Joshua Stromayer

MRI physics, normalisation, infrastructure

joshua@neuvara.org

Baris Ganidagli

Disease biology, clinical translation

baris@neuvara.org

Direction

Where this could go.

Primary direction

Imaging measures for multi-site studies and clinical trials: measures that stay stable when a participant changes scanner, so that site differences do not have to be absorbed as noise.

Longer term

Scanner-invariant representations may support diagnostic research in rare neurological disease, where cases are scattered across sites and scanners and cannot currently be compared.

Research direction, not a current capability

Contact

Get in touch.

We are interested in talking to people working on multi-site imaging, whether or not there is anything to sell yet. The fastest way to reach us is by email.

contact@neuvara.org