Brain Tumour Segmentation via Simulated Federated Learning

Abstract

Correct segmentation of brain tumours from multi-modal MRI is critical to efficient neuro-oncology, but institutional and data privacy laws hinder centralized deep learning solutions. Federated Learning (FL) facilitates collaborative model learning without raw patient data-sharing, but real-world deployment demands overcoming privacy, clinical usability, and compliance issues. This work presents a federated simulation framework that brings together 3D     

nnU-Net models, state-of-the-art aggregation algorithms (FedAvg, FedProx), and differential privacy, which are evaluated using BraTS 2020–2023 data from 20–100 simulated institutions. The system also includes a web-based clinical user interface and detailed audit trails. Experiments show near centralized segmentation precision (DSC for entire tumour = 89.4%, core = 86.2%, enhancing tumours = 82.7%) within a tight privacy budget (€ = 2.9), with robust privacy being ensured by secure aggregation. Usability experiments exhibit decreased review time and high clinician acceptance. This model provides an audit-ready, privacy-enhancing AI benchmark for real-world neuroimaging.