Federated Learning for Radiology

Principal author



Our primary goal is to develop a model- and data-agnostic federated learning toolkit to easily spin up new federated projects with minimal re-implementation of federated components.

We believe federated learning will provide a medium for training deep learning algorithms on large, diverse datasets, to achieve robust performance across institutions worldwide.

Federated Learning for Radiology

We define custom data abstractions for each federated project that handles the full diversity of input and groundtruth data. The implementation details are left to each client at each participating research site.

We develop a design pattern for federated learning that builds on top of local training objects, allowing institution-specific code and data to remain private.

Federated Learning for Radiology flowchart
Desired Outcomes

Improved absolute accuracy and generalization performance of deep learning algorithms across research institutions.

Increased adoption of federated learning using provided components and template code.