Federated Learning with Contrastive Pretraining for Retinal OCT Disease Classification on OCT Retinal Images

Abstract

Retinal optical coherence tomography (OCT) is critical for early detection of vision-threatening diseases, yet the development of robust AI models is constrained by data privacy and heterogeneous multi-institutional datasets. We propose a federated contrastive pretraining framework combining MoCo v3 for self-supervised representation learning with SCAFFOLD to mitigate client drift in federated learning. The framework enables collaborative training across multiple institutions without sharing raw patient data. Extensive evaluations were conducted on multiple retinal OCT datasets, including both in-domain and unseen out-of-domain data, and backbone architectures were systematically compared to assess accuracy, efficiency, and communication cost. The proposed approach achieves 96.8% accuracy and an AUC-ROC of 0.994, outperforming centralized ResNet-50 and conventional federated baselines. Cross-domain evaluation demonstrates robust generalization, with +4.2% accuracy gain on an unseen OCT dataset. Scalability results indicate that the system has a consistent performance up to 16 clients, and Efficient-Net-B4 has the optimum trade-off between accuracy, inference efficiency, and communication overhead. The findings reveal that federated contrastive pretraining improves the performance of diagnostics, cross-domain generalization, and training stability, and maintains privacy of the data. The strategy provides a viable means to multi-institutional collaborative learning in ophthalmology, and it has clinical implications of the automated screening of retinal diseases. MoCo v3 + SCAFFOLD provides a privacy-preserving, scalable, and clinically reliable framework for federated retinal OCT classification, supporting robust AI deployment across heterogeneous institutions.