NeuroFLAME: A Scalable, Privacy-Preserving Federated Framework for Secure, Reproducible, and Multi-Site Neuroimaging Analysis
Journal:
bioRxiv
Published Date:
Aug 27, 2026
Abstract
Federated analysis offers a scalable approach to multi-site neuroimaging research by enabling distributed statistical modeling, machine learning, decomposition, harmonization, and validation without exchanging sensitive individual-level data. However, a significant implementation gap exists, as the majority of federated clinical neuroimaging studies remain technical proofs of concept, struggling to fully navigate the rigorous data-sharing regulations and security constraints of real-world clinical and research environments. Compounding this, existing federated analysis frameworks are typically exposed as command-line tools and configuration files, imposing a considerable setup burden that demands specialized technical expertise from neuroscience researchers. To address these barriers, we present NeuroFLAME, an enterprise-grade, open-source federated neuroimaging platform built upon the NVIDIA FLARE (NVFlare) framework. NeuroFLAME couples a client-outbound-only communication framework with a graphical user interface specifically tailored for neuroscientists. Using certificate-based trust and containerized execution, the platform ensures reproducible analyses with baseline privacy guarantees through data localization and mTLS encryption; advanced protections such as differential privacy and homomorphic encryption are available as optional configurations. By ensuring that clients only need outbound communication and that raw data never leaves each site, NeuroFLAME is designed to support institutional data governance requirements and regulatory frameworks such as HIPAA and GDPR. We demonstrate the platform's utility through three federated analysis workflows, each implemented as a NeuroFLAME computation: multi-site federated closed form voxel-wise regression, decentralized constrained joint independent component analysis (dcjICA) and federated label-based dimensional prediction. Empirical validation of both workflows shows that NeuroFLAME achieves high consistency with established centralized approaches, effectively bridging the gap between experimental federated learning prototypes and production-ready collaborative tools for privacy-preserving, large-scale federated neuroimaging analysis.
