Important issues in data privacy and security for medical information are being addressed through Federated Learning (FL), by enabling distributed learning without sharing sensitive data. It is our intention to assess the FL principles across various platforms for two medical use cases, using both traditional convolutional neural networks and custom models. Such an approach emphasizes hyperparameter tuning, including learning rates and aggregation methods, to ensure optimal performance. Our results indicate that the two custom models, BFIS/BFNet and Bionnica, preserve their good performance when delivered in an FL-enabled framework, either supported by hosted services or on-premise services. These findings highlight the potential of Federated Learning for medical use cases, although challenges remain in optimizing performance for resource-constrained environments and data heterogeneity.
