# Federated training of pathology and radiology models across hospitals

Source: https://onco.cc/ideas/idea-data-federated-learning-imaging/  
OnCo record `idea-data-federated-learning-imaging` (Idea). Data CC BY-NC 4.0, attribute "Data from OnCo (onco.cc)"; commercial use needs a licence.

## TL;DR

Train one AI on pathology slides and radiology scans from dozens of hospitals without any hospital sharing its images: the model travels to the data. Federated learning has worked for glioblastoma segmentation across 70-plus sites, yet almost every clinical model is still trained at one or two institutions, so a persistent shared training infrastructure is proposed.

## Summary

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.

## Fields

- Kind: Idea
- Last checked: 2026-09-08
- 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.
- Proposed test: 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
- Actor: engineering

## Sources

- Federated learning for glioblastoma (Pati et al. 2022): https://www.nature.com/articles/s41467-022-33407-5

## Connected records

- fronts: [AI & Computation](https://onco.cc/fronts/ai-computation/)
- technologies: [AI in radiology](https://onco.cc/technologies/radiology-ai-screening/), [Digital pathology & AI](https://onco.cc/technologies/digital-pathology-ai/), [Pathology & radiology foundation models](https://onco.cc/technologies/pathology-foundation-model/)
- companies: [Owkin](https://onco.cc/companies/owkin/)
- bottlenecks: [AI that is built but not validated or deployed](https://onco.cc/bottlenecks/b-ai-validation/), [Data silos](https://onco.cc/bottlenecks/b-data-silos/)
- ideas: [A pre-competitive consortium to train a shared multimodal cancer foundation model](https://onco.cc/ideas/idea-data-precompetitive-cancer-foundation-model/)

---
JSON: https://onco.cc/api/v1/entities/idea-data-federated-learning-imaging.json