ideasIdea
Federated training of pathology and radiology models across hospitals
Train one AI on slides and scans from many hospitals without any hospital ever sharing its images: the model travels, the data stay.
Federated learning has been demonstrated in oncology (for example the multi-national glioblastoma segmentation federation of over 70 sites, and breast-density and pathology consortia), but almost every clinical model is still trained on one or two institutions. The proposal is a persistent federated training infrastructure with secure aggregation, differential privacy options, per-site audit logs and a shared model registry, offered as a public utility to cancer centres.
Hypothesis
Models trained federatedly across 20 or more sites will generalise to unseen hospitals with less than half the performance drop of single-site models, measured on a sequestered multi-site test set.
Rationale
The largest published federated study (Pati et al., Nature Communications 2022) improved out-of-sample glioblastoma segmentation by a third versus a public-data model. Generalisation failure is the main reason cancer AI does not survive deployment.
What would test it
Train a pathology model for a standard task (for example mitotic count or HER2 scoring) both centrally on one large site and federatedly across 10 sites; evaluate both on five held-out hospitals in other countries.
Maturity
early clinical
Who has to act
engineering
Cost to try
Medium ($1M to $50M)
Years to first evidence
3
Bottlenecks it attacks
- Data silos · Records, scans, genomes and outcomes sit in separate systems that cannot talk. Every patient's experience is lost to the next.
- AI that is built but not validated or deployed · Thousands of cancer AI models are published; a handful are in clinical use, and fewer have shown they help patients.