A Federated Learning and Explainable AI Framework for Privacy-Preserving Brain Tumor Diagnosis Using Multi-Institutional MRI Data

Diagnosing brain tumors from multi-sequence magnetic resonance imaging remains a critical yet challenging task due to data privacy concerns, heterogeneity across institutions, and the limited interpretability of deep learning models. This study proposes a novel federated learning and explainable artificial intelligence (FL-XAI) framework for privacy-preserving brain tumor classification using multi-institutional MRI data. The framework leverages a modified ResNet-18 backbone within a federated averaging setup, enhanced with differential privacy guarantees (ϵ ≈ 3.4, δ = 10-5). Experiments were conducted on the Kaggle Brain Tumor MRI dataset, partitioned across five simulated institutions with non-i.i.d. data distributions. The results show that the centralized CNN achieved the highest accuracy (92%, AUC 0.95), while federated models maintained strong performance at 90% without differential privacy and 88% with preservation of privacy, highlighting minimal trade-offs in diagnostic accuracy. A class-wise analysis of the results suggested strong recall scores, with No Tumor (0.92) and Pituitary (0.91), while Meningioma showed slightly lower recall at 0.80. Explainability using Grad-CAM and SHAP showed that the models consistently highlighted clinically relevant tumor boundaries with the greatest contribution of FLAIR, followed by the T1CE modalities. Robustness tests under adversarial client simulation showed that robust aggregation strategies recovered accuracy to ~88%, mitigating malicious updates. These findings establish the proposed FL-XAI framework as a scalable, privacy-compliant, and interpretable solution for brain tumor diagnosis, bridging the gap between AI innovation and clinical trust. Additionally, we introduce a Lightweight Federated Explanation Alignment (L-FEA) module that reduces explanation drift across institutions and improves explanation consistency by 15% with negligible overhead.