Fed-HARM: A Proposed Federated Hierarchical Adversarial Framework with Mixture-of-Experts Personalization for Multimodal Brain Tumor Segmentation

Abstract

Fed-HARM: A Proposed Federated Hierarchical Adversarial Framework with Mixture-of-Experts Personalization for Multimodal Brain Tumor Segmentation presents a conceptual federated learning framework for multimodal brain tumor segmentation using MRI data. The proposed architecture combines a shared Vision Transformer (ViT)-based feature extractor, adversarial domain alignment, client-local Mixture-of-Experts (MoE) segmentation heads, and a gradient-based explanation module. The framework is designed to address cross-institutional data heterogeneity while preserving data locality through federated optimization. The paper defines the proposed architecture, optimization objectives, parameter-sharing strategy, and a comprehensive evaluation protocol. As a conceptual framework paper, no implementation, experimental results, or clinical validation are reported. The work is intended as a research hypothesis and foundation for future implementation and comparative evaluation against existing personalized federated brain tumor segmentation approaches. Keywords Federated Learning Brain Tumor Segmentation Medical Imaging Vision Transformer Domain Adaptation Mixture of Experts Multimodal MRI Explainable AI Personalized Federated Learning Deep Learning