In order for machine learning algorithms to work robustly and produce accurate results, large amounts of diverse data are necessary. The amount of data available for radiological applications at any single center is insufficient and thus pooling of data across centers is critical for robust AI. Traditional approaches for sharing imaging data involves each hospital transferring its curated and harmonized data sets to a central system who then releases the trained model back to the participating hospitals. However this approach is burdened by regulatory and privacy concerns, even with de-identification, which is often imperfect. We discuss the promise of federated learning for enabling scalable and robust AI for radiology applications at this year's Bioethics, the Law, and Data-Sharing: AI in Radiology Summit.
August 4, 2021