ADAFED-BRAINGNN: ADAPTIVE FEDERATED HYBRID CNN–GNN FRAMEWORK WITH DIFFERENTIAL PRIVACY FOR PRIVACY-PRESERVING BRAIN TUMOR DETECTION ACROSS MULTI-INSTITUTIONAL MRI REPOSITORIES
TL;DR
AdaFed-BrainGNN is presented, an adaptive federated learning framework extending BrainGNN-Hybrid with three innovations: adaptive client-weighting aggregation via composite quality scores; formal (ε, δ)-differential privacy via DP-SGD with Rényi DP (RDP) accounting; and structured gradient sparsification that reduces communication.
Abstract
Brain tumor detection from multi-institutional MRI datasets faces two compounding challenges: segmenting heterogeneous glioma sub-regions, and privacy regulations (HIPAA, GDPR) that prevent data centralization. This paper presents AdaFed-BrainGNN, an adaptive federated learning framework extending BrainGNN-Hybrid with three innovations: (1) AdaFedAvg — adaptive client-weighting aggregation via composite quality scores; (2) formal (ε, δ)-differential privacy via DP-SGD with Rényi DP (RDP) accounting; and (3) structured gradient sparsification that reduces communication by 73.4%. Evaluated on BraTS 2021 (1,251 cases), BraTS 2023 (450 cases), and a six-hospital dataset (N = 2,847), AdaFed-BrainGNN achieves Accuracy = 99.14%, F1 = 98.84%, AUC = 0.997, and ε = 2.31 (δ = 10⁻⁵) after 100 federation rounds, with AWS SageMaker inference at 38 ms per volume