Federated brain tumor classification via gated dual-stream interaction and data-free distillation

Federated learning (FL) enables multi-institutional brain tumor classification without sharing patient data, yet extreme non-IID distributions, where clients entirely lack certain tumor classes cause severe minority-class degradation that standard aggregation cannot remedy. Existing methods either require auxiliary patient data at the server, apply class-agnostic aggregation, or rely on heuristic attention without principled feature interactions. We propose a privacy-preserving FL framework with three contributions: (i) Gated Dual-Stream Feature Interaction (GDSFI), an agreement gate F=s1⊙s2+(1−s1)⊙(1−s2)≈1−(s1−s2)2 that amplifies concordant features while suppressing conflicts via cosine progressive integration; (ii) data-free multi-teacher distillation using server-generated synthetic probes containing no patient-identifiable content, with per-class reliability weighting that hard-zeros absent-class contributions from non-expert clients; and (iii) an automatic intervention monitor that dynamically detects and corrects per-class accuracy degradation during federation. Evaluated on the Figshare brain tumor dataset across five class-absent heterogeneous clients, the framework yields significant gains in minority-class F1, accuracy, and AUC, recovering minority-class performance to natural non-IID levels while transmitting no patient data. As all three components operate on per-class signals, the framework generalizes to class-imbalanced IID and moderately non-IID settings without architectural modifications.