PSO-based Client Selection to Improve Medical Image Classification in Federated Learning

Federated Learning (FL) offers an effective approach to enable collaborative model training across decentralized medical institutions while preserving patient privacy. However, FL faces challenges related to data heterogeneity and inefficient client participation issues. In this study, we introduce a novel client selection strategy based on Particle Swarm Optimization (PSO) to improve medical image classification in federated settings. Our method dynamically selects informative clients by evaluating their contributions to the global model, which reduces communication overhead and supports faster convergence. We performed experiments using three publicly available medical imaging datasets, Biglycan, ALL-IDB, and LC25000, focusing on binary classification tasks. The results show that the PSO-based approach improves performance over random client selection across all evaluated scenarios, with gains in accuracy, recall, precision, and F1-Score, even when selecting fewer clients. These findings indicate that incorporating PSO into FL frameworks helps address challenges related to data heterogeneity and communication efficiency while maintaining classification performance and privacy preservation.