Decentralised Approaches to Training AI Models in Healthcare

In the last decade, Artificial Intelligence (AI) has transformed the healthcare field, particularly medical image analysis, by enabling the automated learning of feature representations from large-scale datasets across a wide range of medical tasks. However, its broader clinical adoption, especially in the imaging context, remains constrained by the difficulty of collecting large, heterogeneous training datasets. Patient data are subject to strict regulatory frameworks, such as the General Data Protection Regulation (GDPR), which impose substantial barriers to multi-centre data sharing and often render large multi-centre centralised training impractical.

Furthermore, AI models trained on data from a single or a limited number of institutions are prone to important performance degradation when deployed in unseen clinical environments, owing to variability in acquisition protocols, scanner hardware, and patient populations.

Federated Learning (FL) is able to address these concerns by enabling collaborative model training without sharing raw data, thereby preserving privacy and improving exposure to diverse data distributions. Despite its demonstrated effectiveness for image classification and segmentation tasks, the application of FL to medical Image-to-Image (I2I) translation was explored only in the field of cross-modality Magnetic Resonance Imaging (MRI) translation prior to this work. ... mehrThis thesis aims to address this gap by focusing on Synthetic Computed Tomography (sCT) generation as a relevant use case and further supporting the FL in healthcare.

In detail, the research work presented in this thesis spanned four interconnected areas: feasibility, efficiency, security, and translational tooling. The feasibility of federated sCT generation was established through multi-centre studies. The first study investigated brain MRI-to-sCT translation across five institutions, giving rise
to the FedSynthCT framework. The second extended this framework to head and neck Cone Beam Computed Tomography (CBCT)-to-sCT translation, using four publicly available institutional datasets. In both studies, federated models trained on heterogeneous distributed data achieved stable convergence and consistent generalisation to independent external datasets, without requiring data centralisation or site-specific adaptations.

These results confirmed that sCT generation is a viable application domain for FL, and that inter-site data heterogeneity does not preclude effective decentralised collaboration. Practical barriers to FL applicability and deployment were subsequently investigated. Within the FedSynthCT framework, an adaptive parameter-freezing strategy was proposed to substantially reduce the computational demand, and therefore the time and energy required for federated training, in line with the increasingly important principles of Green AI. Furthermore, a security analysis of the FedSynthCT framework revealed that gradient inversion attacks did not represent a viable threat in the evaluated setting, while residual vulnerabilities to membership inference and poisoning attacks were identified, motivating the future adoption of system-level safeguards beyond the federated algorithm itself.

The implications of communication-efficiency techniques, including model compression, sparsification, and weight clustering, were further characterised in the context of federated brain tumour segmentation setting, demonstrating that only moderate and well-balanced levels of these techniques can substantially reduce communication overhead while preserving segmentation accuracy.

Open-source software tools were developed and released to bridge the gap between research and its application in healthcare. ImageAugmenter, a medical image augmentation extension for the 3D Slicer Imaging Computing Platform, was published to support data preparation for both centralised and federated pipelines. Subsequently, the 3D Slicer extension named ModalityConverter was released to integrate the pre-trained federated I2I models developed throughout this research alongside existing centralised models, enabling their use and further development within a widely adopted medical research environment. The extension also includes an automated quality assessment module, based on a Deep Learning (DL) framework that predicts quantitative sCT Mean Absolute Error (MAE) without requiring ground-truth Computed Tomography (CT) data, thereby supporting direct output validation in a single environment, without requiring specialised infrastructure or programming expertise.
The contributions of this thesis confirmed that decentralised learning can be a practically viable paradigm for collaborative medical image analysis and established the methodological, analytical, and software foundation for future multi-centre clinical validation and broader deployment of federated I2I translation in healthcare.