Towards Privacy-Preserving Big Data Analysis With Decentralized Learning for ROP Imaging

The adoption of Artificial Intelligence in healthcare has significantly enhanced diagnostic accuracy and operational efficiency, particularly in medical imaging. However, the development of deep learning models often relies on centralized data collection, which raises privacy concerns and logistical barriers, especially in sensitive medical domains such as Retinopathy of Prematurity (ROP). This study investigates the use of Federated Learning (FL) as a privacy-preserving alternative to centralized learning for ROP classification. Through three experimental scenarios—centralized learning, FL, and limited-data learning—we evaluate the trade-offs in performance, privacy, and scalability. Results indicate that FL achieves comparable performance to centralized models, with only a minimal drop in AUC (0.05) while preserving data privacy. In contrast, the limited-data scenario demonstrated a significant performance decline, underscoring the importance of data quantity and diversity. Despite its promise, FL faces challenges related to scalability, data heterogeneity, and real- world deployment. This paper highlights the potential of FL for collaborative medical AI while addressing critical threats to validity and proposing directions for future research.