Towards generalizable Federated Learning in medical imaging: A real-world case study on mammography data

Federated Learning has been rapidly gaining in popularity in medical applications, due to the increased privacy offered, since medical data doesn't need to leave the hospitals' premises for AI model training. However, a direct translation of a classic experiment to a federated one is not always straightforward. In this work, we delve into the intricacies of federated learning for a breast cancer classification tool. We compare classic model training with a federated variant, and highlight the adaptations that need to be taken care of to ensure the equivalence between the two. Specifically, we introduce the Breast Area Detection tool as an essential component of the pre-processing pipeline to enhance the robustness of Federated Learning by offering data harmonization. On top of that, we present an end-to-end Federated Learning framework that is effective for real-world data and scenarios. Among the three real-world hospitals involved in the experimental procedure, the proposed framework significantly improves performance at the first hospital, providing consistent results similar to those achieved in the classic approach. Experimental results demonstrate that the interventions introduced improved model performance by approximately 35%, aligning federated learning and centralized model performance.