Federated Learning Achieves Privacy-Preserving Brain Tumor Localization with CAFEIN\textsuperscript{\textregistered}

Brain tumor localisation presents a significant challenge, often hampered by the difficulty of accessing sufficiently large and diverse patient datasets due to stringent privacy regulations. Researchers Andrea Protani, Riccardo Taiello, and Marc Molina Van Den Bosch, alongside Luigi Serio, from institutions including the European Organization for Nuclear Research and BCN Medtech, have tackled this problem by developing a novel federated learning framework. Their work, utilising a Transformer-Graph Neural Network within CERN’s CAFEIN platform, allows multiple healthcare institutions to collaborate on model training without directly sharing sensitive patient data. Crucially, this research demonstrates that federated learning not only overcomes data silo limitations , achieving performance comparable to centralised training , but also provides valuable insight into how the model arrives at its conclusions through modality-level explainability, confirming the clinical relevance of MRI modalities like T2 and FLAIR.

This breakthrough demonstrates the power of collaborative learning in scenarios where data access is restricted, paving the way for more generalizable and accurate brain tumor analysis tools. Furthermore, the team incorporated explainability analysis via Transformer attention mechanisms, providing insights into which MRI modalities most influence the model’s predictions, a crucial step towards building trust and facilitating clinical adoption. This observation aligns with radiological expertise, where these modalities are known to provide complementary information essential for accurate tumor delineation.

This work opens new avenues for multi-institutional research and the development of advanced diagnostic tools for brain tumors, ultimately improving patient outcomes. The team achieved this by deploying their Transformer-GNN architecture using CAFEIN®, a platform specifically designed for federated learning in healthcare environments. This deployment showcases the practical feasibility of the proposed framework and its compatibility with existing healthcare infrastructure. By exchanging only model parameters, rather than raw patient data, the system ensures compliance with stringent privacy regulations while enabling collaborative model training. The study’s findings provide compelling evidence that federated learning is not merely a theoretical possibility, but a viable and effective solution for addressing the challenges of data fragmentation in medical imaging, and the results demonstrate a significant benefit from aggregating knowledge across multiple institutions.

Federated Brain Tumor Localization via CAFEIN® offers promising

The team engineered a system where multiple institutions collaboratively train a model without directly sharing sensitive patient data, exchanging only model parameters to preserve privacy. Experiments employed the BraTS dataset to rigorously evaluate the performance of this federated approach. Researchers meticulously designed experiments to compare isolated training with federated learning, revealing a critical performance bottleneck. Conversely, federated learning enabled continued model refinement by leveraging the collective knowledge distributed across multiple institutions, ultimately achieving performance levels equivalent to centralized training.

This finding strongly justifies the adoption of federated learning for complex tasks involving high-dimensional data, demonstrating the benefits of aggregated institutional knowledge. To understand the model’s decision-making process, the study pioneered an explainability analysis using Transformer attention mechanisms. Scientists harnessed these mechanisms to reveal which MRI modalities, T1-weighted, T1ce, T2-weighted, and T2-FLAIR, most influenced the model’s predictions. The innovative methodology not only addresses privacy concerns but also enhances model performance and interpretability. This approach enables multi-institutional collaboration, unlocking the potential of diverse datasets to improve brain tumor localization and ultimately, patient care. The team’s work showcases the power of combining advanced deep learning architectures with federated learning principles to overcome data limitations and build more robust and transparent AI systems for medical imaging.

Federated Learning Matches Centralised Brain Tumour Performance

The team meticulously constructed supervoxel graphs from BraTS volumes, partitioning each normalized T1 volume into 4000 locally uniform regions using 3D SLIC clustering. Background supervoxels with low T1 intensity were pruned, and a sparse adjacency structure was established connecting each retained supervoxel to its eight nearest neighbours based on Euclidean distance. Each supervoxel was then represented by a tensor of 360×48 dimensions, derived from 90 patches per modality across the four MRI modalities, T1, T1ce, T2, and FLAIR, and assigned a binary label based on a threshold of 0.20 for tumor presence. This innovative approach focuses on tumor localization at the supervoxel level, providing an anatomically-informed shape that better delineates irregular tumor boundaries, with Dice score reported as an auxiliary metric.