CEPerFed: Communication-Efficient Personalized Federated Learning for Multi-Pulse MRI Classification

TL;DR

CEPerFed tackles privacy-preserving, multi-pulse MRI classification by combining a novel local/global collaborative optimization with a layer-aware HSVD compression scheme. It introduces historical risk and historical average gradients to harmonize personalized updates with global learning, while HSVD reduces communication from large CNN models by decomposing gradients and parameters into carefully chosen low-rank factors across three convolutional parts. The approach demonstrates robust improvements over competitive FL baselines across G, S, R pulse configurations and under non-IID data, with ablations confirming the value of both the risk-guided optimization and the hierarchical compression. The work highlights practical gains for distributed medical imaging with limited bandwidth and heterogeneous data, and provides a path toward scalable, privacy-preserving deployment in real clinical settings.

Abstract

Multi-pulse magnetic resonance imaging (MRI) is widely utilized for clinical practice such as Alzheimer's disease diagnosis. To train a robust model for multi-pulse MRI classification, it requires large and diverse data from various medical institutions while protecting privacy by preventing raw data sharing across institutions. Although federated learning (FL) is a feasible solution to address this issue, it poses challenges of model convergence due to the effect of data heterogeneity and substantial communication overhead due to large numbers of parameters transmitted within the model. To address these challenges, we propose CEPerFed, a communication-efficient personalized FL method. It mitigates the effect of data heterogeneity by incorporating client-side historical risk gradients and historical mean gradients to coordinate local and global optimization. The former is used to weight the contributions from other clients, enhancing the reliability of local updates, while the latter enforces consistency between local updates and the global optimization direction to ensure stable convergence across heterogeneous data distributions. To address the high communication overhead, we propose a hierarchical SVD (HSVD) strategy that transmits only the most critical information required for model updates. Experiments on five classification tasks demonstrate the effectiveness of the CEPerFed method. The code will be released upon acceptance at https://github.com/LD0416/CEPerFed.

Figures (6)

• Figure 1: Overview of CEPerFed. Left: Each client updates its local model and historical risk gradient through local and global collaborative optimization module. After local training, clients decompose their gradients and parameters into low-rank factors using the proposed HSVD strategy and upload these factors to the server. Right: The received factors are reconstructed into each client’s gradients and models. The server then computes the historical average gradient and the historical risk gradient via Eq. (\ref{['eq12']}) and Eq. (\ref{['eq13']}), respectively, and updates the global model based on the aggregated parameters. Center: The server sends the updated gradients and the new global model to clients, which then proceed to the next training round.
• Figure 2: Hierarchical partitioning of ResNet-18 into three distinct parts for the HSVD strategy.
• Figure 3: Execution workflow of the proposed HSVD strategy across different network parts. The right arrows indicate the transmission direction of the compressed low-rank factors and do not imply a sequential dependency between the parts.
• Figure 4: Heatmap showing the average classification accuracy across the five diagnostic tasks for each pulse configuration.
• Figure 5: Ablation experiment results of risk matrix $\alpha$ and HSVD strategy.