The rarest cases arethe ones nobody can pool.

Cancer intelligence needs scale that no single centre has, built from records no centre is allowed to send anywhere. Oncology AI trains across registries without moving one case out of the institution that holds it.

Talk to us about a pilot The protocol underneath→

01The problem

Scale is the treatment, and scale is what the rules forbid.

What holds the data in place

Registries do not poolCannot pool

A cancer registry is built from identifiable records under a consent framework specific to the institution that collected them. Merging two of them is a legal act, not an engineering one.

Rare is the whole problemNeeds scale

The subtypes that most need a model are the ones no single centre sees enough of. The signal only exists across sites.

Imaging is heavy and localStays put

Whole-slide and volumetric imaging is measured in gigabytes per case, held in systems that were never designed to export.

None of these are solved by moving the records. They are solved by not moving them.

02The approach

Four steps, none of which export a case.

Ingest in placeScoreTrainAttestServe

Each centre keeps its own archive and runs its own node. What crosses the boundary is a weight update and a signed record of what that centre contributed.

01PLANNED

Score in placeEvery case is evaluated on the centre’s own hardware, against its own ground truth.

02PLANNED

Train across centresRounds of federated training so rare subtypes get the scale they need.

03PLANNED

Attribute contributionA signed record of which centre moved the model, and by how much.

04PLANNED

Serve locallyInference runs beside the scanner. The prediction never becomes an export.

03What crosses

Stays in the centre

Slides and volumes

Pathology reports

Patient identifiers

Treatment records

Nothing in this group is copied, pooled or shipped.

Travels out

Model weights

Aggregate evaluation scores

Signed contribution record

Only these cross the boundary, and only outward.

The record is measured, indexed and annotated where it lies. Only the right-hand column ever leaves.

04Where it applies

Multi-site registriesBuild one model across centres that are not permitted to share a single record.

Tumour boardsSurface comparable prior cases from the centre’s own history, not from a public corpus.

Trial matchingScreen against protocol criteria without moving the record to the sponsor.

Quality assuranceA second read on every case, scored the same way at every participating site.

05Close

Learn from every centre. Move nothing.

Talk to us about a pilot