Federated Transformer-GNN for Privacy-Preserving Brain Tumor Localization with Modality-Level Explainability

TL;DR

The paper tackles privacy-preserving brain tumor localization from multimodal MRI by deploying a federated Transformer–GNN on a supervoxel graph representation within CERN's CAFEIN platform. It introduces modality-level explainability via Transformer attention, showing that deeper layers emphasize T2 and FLAIR modalities, validated with rigorous statistics ($p<0.001$, $d=1.50$). Experiments on BraTS demonstrate that federated training matches centralized performance and outperforms isolated training, demonstrating the value of cross-institutional data sharing without data leakage. Overall, the work advances clinically relevant, privacy-preserving AI for brain tumor analysis and supports broader adoption of federated learning in healthcare.

Abstract

Deep learning models for brain tumor analysis require large and diverse datasets that are often siloed across healthcare institutions due to privacy regulations. We present a federated learning framework for brain tumor localization that enables multi-institutional collaboration without sharing sensitive patient data. Our method extends a hybrid Transformer-Graph Neural Network architecture derived from prior decoder-free supervoxel GNNs and is deployed within CAFEIN\textsuperscript{\textregistered}, CERN's federated learning platform designed for healthcare environments. We provide an explainability analysis through Transformer attention mechanisms that reveals which MRI modalities drive the model predictions. Experiments on the BraTS dataset demonstrate a key finding: while isolated training on individual client data triggers early stopping well before reaching full training capacity, federated learning enables continued model improvement by leveraging distributed data, ultimately matching centralized performance. This result provides strong justification for federated learning when dealing with complex tasks and high-dimensional input data, as aggregating knowledge from multiple institutions significantly benefits the learning process. Our explainability analysis, validated through rigorous statistical testing on the full test set (paired t-tests with Bonferroni correction), reveals that deeper network layers significantly increase attention to T2 and FLAIR modalities ($p<0.001$, Cohen's $d$=1.50), aligning with clinical practice.

Figures (3)

• Figure 1: Supervoxel graph preprocessing pipeline: (a) Input MRI modalities; (b) 3D SLIC supervoxel generation applied to T1; (c) Graph construction connecting each supervoxel to its k-nearest neighbors; (d) Patch extraction within each supervoxel for multimodal feature encoding.
• Figure 2: Training dynamics comparison between centralized, federated, and isolated (average across clients) learning paradigms. Evolution of loss, Dice score, recall, and F1 score over training epochs/rounds. Shaded bands represent variability across clients. For visualization purposes, centralized and isolated curves are extended beyond their training duration using the last recorded value (dashed lines with reduced opacity) to facilitate comparison with federated training. Key observation: isolated training plateaus early due to limited local data, while federated learning continues improving and ultimately matches centralized performance.
• Figure 3: CLS token attention distribution across modalities and Transformer layers (full test set, error bars $\pm$1 SD). Statistical analysis reveals that deeper layers significantly increase attention to T2 and FLAIR modalities compared to early layers (trend test: $p<0.001$, Cohen's $d$=1.50).