Federated Learning for Brain Tumor Detection

The current global model is broadcast to all 8 real institutions; each fine-tunes it on its own real data only, and just the resulting weight updates come back — never the scans themselves.

A brain-MRI classifier (glioma / meningioma / notumor / pituitary) trained across 8 real hospital and public data sources — two Kaggle collections, a Chinese-hospital dataset, UT Southwestern, an EU GBM cohort, UPenn-GBM, ReMIND, and AOMIC — without ever centralizing a scan. That's the premise federated learning tests: real clinical imaging is usually too sensitive to pool across institutions, so instead of moving the data, the model moves instead. Five of the eight real institutions see only a single class each — genuine non-IID data, not a synthetic partition — and every mechanism is scored against a fixed, leakage-checked 300-scan holdout that no training run ever touched.

Three real questions structure the work. RQ1: does federated training cost accuracy against a centralized baseline trained on the same data? RQ2: on cases the model was never trained to classify, does its attention still land on the real lesion? RQ3: what does adding differential privacy (DP-SGD) actually cost in accuracy? Four aggregation mechanisms (FedAvg, FedProx, SCAFFOLD, FedPer) plus two custom ablations were run for real, full 15-round federated training across all 8 institutions — including a mechanism that didn't work, reported as a real negative result rather than left out.