Federated learning has an enormous impact on research.

Unified learning, where according to FAIR PRINCIPLES Data applications rather than the other way around was still an unexplored area for Erasmus MC. ride to join in 2020 In a large-scale international study on the contribution of federated learning to medical scientific research based on artificial intelligence.

Neuroradiologist and professor of neuroradiology Marion Smits knew the principal investigator of Penn’s study, Spyridon Bakas, among others from a joint consortium trying to map brain tumors in various ways (for the research part of Smits’ careers, the focus is on brain tumors). “When Bacas was creating federal education research on these tumors two years ago, he also asked us to join,” says Smits.

Now you see more studies in which federated learning is embedded, but when Erasmus MC started it, it was still new, adds biomedical researcher Sebastian van der Voort (active in the Smits research group). His doctoral research focused on applying AI technology to obtain information about potential brain tumors that would normally be obtained through biopsies.

Advantages of federated learning

Access to large datasets is often a major problem for medical researchers, especially due to privacy regulations. According to Van der Voort, federated learning by itself does not provide access to more data. But since everyone can keep the data under their own management, people are more inclined to give access to their data. “They no longer have to hand over this data, which is always privacy-sensitive, but they can put an AI application under their control over their own data. This makes it much easier to access the data.”

Data access is also often limited as it relates to older data, especially for rare tumours. Nowadays, it is very normal to ask people for permission when using data in medical scientific research, but this has not always been the case,” Smits adds. “Permission not obtained to share data is not a problem if that data has not left the private servers with it. Federated learning thus avoids the problem that privacy regulation sometimes poses.”

Special results

Who is he Research article Reading about the study, he may not immediately realize how special the results are, Smits knows. “It makes sense that we developed a robust, generalized application of AI. But they are special, because they often don’t work. 71 centers were involved, resulting in a very large and diverse amount of data. They all learned how to create a unified learning environment. These types of datasets are really the future. , and no longer self-collected datasets using the self-developed algorithm which are then sent out into the world. Federated learning sets a new standard.”

Read the extensive interview with Marion Smits and Sebastian van der Voort in ICT and Health Issue 1, which will be published on 17 February.