Learn more about Federated Learning
Creating an artificial intelligence (AI) model for a healthcare application which works well at multiple institutions typically requires a large collection of training data acquired from varied sources. Obtaining such large and varied healthcare training datasets can be difficult given the sensitive nature of medical data. FL enables data from multiple institutions to be used for AI model training without data leaving institutional firewalls. This is accomplished by sharing AI models, not sensitive data, with participating institutions.
Participating in AI-LAB FL experiments will require an on-prem ACR Connect installation or a secure ACR-Connect instance in the cloud. To participate in an FL experiment, institutions will be required to create and validate a relevant dataset using ACR-Connect Data Manager. To join an AI-LAB FL experiment, a designated point of contact will select to join the FL experiment on a scheduled date and time. From there, the experiment will be automated for institution participants.
Behind the scenes, an ACR central server will pass a base model to the ACR Connect instance at each participating institution. At each institution, the base model will then be trained on the previously prepared local dataset to create a new model. New models from each institution will be passed back to the ACR central server for aggregation. The process of aggregation will create a single model which will be used as the base model in the next round of training. This process will be repeated many times to create a final FL model on the ACR central server, as well as a local model at each institution finetuned on data from that institution.
If you are interested in participating in AI-LAB FL, please sign up to receive more information on next steps!