Adaptive aggregation in federated learning for retinal vein occlusion detection: A dynamic approach based on data...

Clinical decision support systems (CDSS) increasingly rely on artificial intelligence (AI) to interpret biomedical images and provide accurate, real-time diagnostics. Federated Learning (FL) has emerged as a promising privacy-preserving solution for training AI models across decentralized healthcare institutions. However, FL performance is significantly affected by data heterogeneity-especially when client data is non-independent and identically distributed (non-IID), a common occurrence in clinical practice. In this study, we propose a dynamic FL framework specifically designed for Retinal Vein Occlusion (RVO) detection that enhances the adaptability and robustness of CDSS. The core contribution of our approach is a server-side aggregation strategy that selects FedAvg for IID data and SCAFFOLD for non-IID data based on monitoring the normalized L2 divergence between successive communication rounds. This intelligent selection mechanism ensures stable convergence, improves diagnostic accuracy, and supports reliable model performance across diverse healthcare environments. Experimental results show that the proposed approach was evaluated on a two-class RVO dataset across five simulated clients under both IID and non-IID data distributions. By addressing real-world data distribution challenges, this work offers a practical path toward integrating FL-based AI systems into clinical workflows for more effective and trustworthy medical image analysis.