Dynamic client selection in federated learning using particle swarm optimization for medical image classification

Federated Learning (FL) enables decentralized medical institutions to collaboratively train predictive models without sharing sensitive patient data. However, conventional FL pipelines commonly rely on random client selection, which ignores the heterogeneous and unequal contributions of local data across institutions. This limitation can slow the convergence and reduce the model performance in non-IID clinical scenarios. To address this problem, we propose a dynamic client selection strategy based on Particle Swarm Optimization (PSO). The proposed method formulates client selection as a combinatorial optimization problem and uses PSO to identify high-utility client subsets at each communication round without modifying local training or the aggregation mechanism. We evaluated the method on three public medical imaging datasets, LC25000, Biglycan, and ALL-IDB, using four CNN architectures under heterogeneous federated settings. The evaluation compares FedAvg, Oort, PyramidFL, and a GA-based baseline, and includes ablation, PSO grid search, Dirichlet non-IID analysis, and statistical validation. The results show that PSO-driven client selection improves model performance mainly in configurations with limited client participation and lightweight architectures, such as SqueezeNet and MobileNet. We demonstrated that the proposed method benefits mainly from empirical client utility, operates reliably within practical PSO parameter ranges, and remains most effective in heterogeneous settings with sufficient client participation. Overall, these findings suggest that optimization-guided client selection can improve federated medical image classification while preserving compatibility with standard FL aggregation pipeline.