<![CDATA[Breast cancer remains one of the leading causes of cancer-related mortality among women worldwide, making early detection a critical clinical priority. While deep learning has demonstrated strong performance in mammogram classification, its development is constrained by data privacy regulations that prevent centralized data collection across medical institutions. Federated learning (FL) offers a promising solution by enabling distributed model training without transferring raw patient data. However, FL faces significant challenges under Non-Independent and Identically Distributed (Non-IID) data conditions, which are common in real-world medical settings and can degrade model performance and convergence stability. This study proposes a breast cancer classification system based on FL using FedProx and DenseNet-201, with hyperparameter optimization via Particle Swarm Optimization (PSO) to improve model performance under heterogeneous data distributions. Experiments were conducted across multiple Dirichlet distribution scenarios (α = 0.1, 0.3, 0.5) comparing FL baseline and FL PSO based. Results show that PSO consistently improved performance across all scenarios, with the most significant gain observed under highly Non-IID conditions (α = 0.1), where the F1-score increased from 83.18% to 88.00%. PSO-optimized FL also demonstrated faster convergence, reducing the number of rounds required to reach optimal performance. Furthermore, the optimized configuration yielded lower training time per round compared to the baseline. These findings indicate that PSO-based hyperparameter optimization effectively enhances FL performance under data heterogeneity while preserving patient data privacy.]]>
