<p>This study presents a comparative analysis of ResNet and EfficientNet in a decentralized, privacy-preserving environment utilizing an innovative federated adaptive aggregation (FAA) framework. The FAA framework introduces a dynamic switching mechanism that alternates between federated averaging (FedAvg) and federated stochastic gradient descent (FedSGD) based on the observed data divergence across clients. Specifically, FedSGD is employed in high-divergence scenarios to enable more refined updates, whereas FedAvg is utilized when divergence is low to optimize communication efficiency. Experiments conducted on tuberculosis chest X-rays, brain MRI scans, and diabetic retinopathy images demonstrate the efficacy of this approach. The FAA framework reduces communication overhead by 23.5% while improving model convergence speed by 18.7% compared to conventional FedAvg-based aggregation. EfficientNet, due to its lightweight architecture, achieves a 5.2% reduction in communication cost while maintaining a classification accuracy exceeding 96% across all datasets. ResNet, in contrast, exhibits greater robustness, particularly in handling complex image features such as microaneurysms in diabetic retinopathy and tumor boundaries in MRI scans. By adapting to the heterogeneity of data distributions, including class imbalance and variations in dataset sizes across institutions, the FAA framework ensures privacy-preserving, scalable, and resource-efficient collaborative learning. This research underscores the significance of the FAA framework as a practical solution for multi-institutional healthcare collaborations, addressing critical challenges in federated learning-based medical image classification.</p>