Federated Learning for Breast Cancer Classification: A Comparative Study of Aggregation Methods

Federated Learning (FL) allows healthcare institutions to collaboratively develop machine learning models while safeguarding patient data, making it ideal for privacy-sensitive medical imaging. This study explores the effects of data heterogeneity on federated breast cancer classification using MobileNetV2 across five simulated clients. Five aggregation methods—FedAvg, FedProx, FedNova, FedDyn, and SCAFFOLD—were assessed under various data distributions, including balanced, imbalanced, non-homogeneous, and non-IID. Results indicate that aggregation performance is significantly affected by data distribution; FedAvg excels in balanced settings but falters in heterogeneity, whereas FedProx shows robustness in extreme non-IID cases, achieving up to 98.466% accuracy. FedDyn and SCAFFOLD also demonstrate adaptability but are less consistent in severe imbalance scenarios. Beyond accuracy, recall and robustness under extreme non-IID conditions were analyzed to assess clinical reliability in cancer detection. These results underscore the necessity of choosing suitable aggregation methods for effective medical federated learning.