Comparative Evaluation of Gradient Compression Strategies for Communication-Efficient Federated Learning in Multi-Hospital Medical Image Classification

Federated learning enables privacy-preserving collaborative training across hospitals, yet the communication overhead of exchanging model parameters remains a critical deployment bottleneck. While gradient compression techniques have been extensively studied in distributed training, their effectiveness under the heterogeneous data distributions characteristic of multi-hospital settings is not well understood. This paper presents a controlled empirical comparison of six gradient compression strategies --- stochastic quantization (QSGD), ternary quantization (TernGrad), sign-based compression (signSGD), Top-K sparsification, Random-K sparsification, and a hybrid sparsification-quantization approach --- applied to federated medical image classification. Experiments are conducted on Fed-ISIC2019 with six natural hospital centers and PathMNIST with synthetic non-IID partitioning across five clients. Results indicate that Top-K sparsification with error feedback achieves the strongest accuracy--communication tradeoff, retaining 97.8% of the 200-round baseline accuracy at nominal 100× compression on Fed-ISIC2019. Multi-bit quantization methods remain more stable as data heterogeneity increases. Sign-based compression, evaluated under a different aggregation protocol (majority vote) than the other methods, degrades substantially under natural non-IID conditions. The hybrid approach performs strongly in the low-budget regime but introduces additional implementation complexity. Communication savings are reported as analytical estimates based on nominal compression ratios; protocol-level overhead would moderately reduce actual savings in deployment. These findings provide evidence-based guidance for healthcare institutions selecting compression strategies for bandwidth-constrained federated learning deployments.