A Divergence-Aware Personalized Federated Transformer Framework for Privacy-Preserving Multi-Center Medical Image Intelligence

Abstract

In a non-IID medical imaging scenario, the problems that occur in federated learning (FL) include high data variability, data leakage, convergence instability, and suboptimal global aggregation. The number of medical imaging applications is vast, and Federated Learning (FL) has already been used in many of them; some main challenges are the large variability in data, the potential for leakage of privacy, convergence instability, and suboptimal global aggregation in non-IID scenarios. Current adaptive aggregation strategies are heuristic optimizer switching, do not take advantage of representation learning to transform the inputs, and are not able to personalize it in a divergence-aware way. This paper will present DAP-FedTrans, a Divergence-Aware Personalized Federated Transformer framework that will be used to conduct privacy-preserving multi-center medical image intelligence. This framework quantifies the statistical heterogeneity with Jensen-Shannon divergence, gradient similarity, and Wasserstein feature distance for the purpose of dynamically partitioning the clients and providing the aggregation per cluster. The classification heads are specialized for institutions, and the universal Vision Transformer encoder is used to encourage generalization. The privacy guarantees are augmented with secure aggregation and adaptive differential privacy. The accuracy is 97.84%; F1 is 0.968, and the loss of communication is 33% in the extreme non-IID case. The outcomes reveal increased convergence stability, equity, and scalability, making DAP-FedTrans a possible paradigm for the collaborative implementation of AI-based healthcare.