Federated learning across healthcare institutions
We investigate collaborative model training when medical images remain at their originating institutions.
How can institutions learn a useful shared model while accounting for differences in their local data?
Method overview
Local data, collaborative training
Each site updates a model using data held within its own institution.
Schematic of the research approach; image patterns and measurements are illustrative.
Research overview
Multi-institutional imaging research must contend with differences in scanners, acquisition protocols, patient populations, and labels. Federated learning allows sites to train locally and exchange model updates instead of pooling their raw images. Keeping data local changes the research workflow, but does not eliminate privacy, security, or statistical challenges.
Our work examines these challenges through methodological analysis and medical imaging applications. Studies in pancreatic cyst risk assessment compare federated and centralized learning across institutions. The evaluation asks how site differences affect the model, and whether shared training improves performance beyond an individual institution’s data.
Local data, collaborative training
- Train local models using each institution’s imaging and reference labels.
- Aggregate model updates and return a shared model to participating sites.
- Evaluate cross-site performance, heterogeneity, and relevant privacy assumptions.
Selected research sources
Research enquiries
Contact the lab