Federated Learning (FL) offers a promising solution for training machine learning models on decentralized data while preserving privacy, making it particularly valuable for sensitive applications such as healthcare. This study implements FL for the classification of Alzheimer’s disease using MRI images, addressing two critical challenges: data heterogeneity and class imbalance. The research evaluates the performance of the FedAdagrad optimization algorithm against the standard FedAvg approach under varying data distribution scenarios. The methodology employs a CNN trained on a dataset of 6,400 MRI images across four severity classes, partitioned non-IID using Dirichlet distributions (α = 0.1, 0.5, 0.9) to simulate real-world heterogeneity. Experiments were conducted using the Flower framework with four clients over ten communication rounds. Results indicate that FedAdagrad achieves a superior F1-score of 50.33% compared to FedAvg’s 48.14%, though both fall short of centralized CNN performance (55%). High data heterogeneity (α = 0.1) leads to a 13.35% accuracy decline, underscoring FL’s sensitivity to uneven data distributions. Class imbalance emerges as the primary bottleneck, affecting all models. The findings contribute to the growing body of research on adaptive optimization in federated settings, offering insights for future improvements in decentralized healthcare AI.
Arini et al. (Thu,) studied this question.