Advancing AI in Healthcare Through Federated Continual Learning

This thesis begins by exploring Continual Learning, a research area focused on reducing model forgetting. To address this, various bio-inspired solutions are proposed, leveraging auxiliary knowledge, auxiliary tasks, and pretraining techniques to enhance learning retention. The thesis then shifts focus to Federated Learning, which enables distributed model training while protecting data privacy—a critical requirement in healthcare. In this section, realistic federated medical scenarios are simulated, and methods are introduced to facilitate data sharing within privacy-preserving frameworks. Specifically, a GAN-based latent space aggregation method is proposed, transforming private datasets into an aggregated and shareable form. This aggregation strategy is further refined through a privacy-preserving latent space navigation technique, increasing the generation of shareable samples by a GAN trained on medical data (such as chest X-rays and retinal fundus images). The thesis progresses by integrating Continual Learning strategies into Federated Learning to address challenges within decentralized medical applications. Building on earlier methods, Continual Learning techniques are combined to a novel Privacy-Preserving GAN, effectively tackling specific obstacles in Medical Federated Learning. The effectiveness of this integrated strategy is assessed on two distinct medical federations focused, respectively, on tuberculosis classification and skin lesion classification. Finally, the thesis addresses the more complex scenario of Federated Continual Learning, where data is both spatially distributed across nodes and evolves over time. To meet this challenge, a communication strategy inspired by experience replay is introduced, enabling effective inter-client communication across nodes. This comprehensive exploration lays a strong foundation for advancing AI in healthcare by merging Federated Learning and Continual Learning, demonstrating pathways for effective privacy-preserving models suitable for real-world medical applications.