Federated Learning for Multi-Disease Ophthalmic Diagnostics Using OCT Angiography

To conduct a comprehensive systematic evaluation of federated learning (FL) strategies for multi-disease retinal classification using OCT angiography (OCTA), implementing a 2-part experimental framework to establish foundational feasibility and optimize performance under realistic heterogeneous conditions while ensuring privacy preservation.

Retrospective multi-center FL study using a systematic 2-part experimental design: (1) foundational feasibility evaluation under controlled homogeneous conditions, and (2) comprehensive optimization under realistic heterogeneous conditions using Dirichlet distribution partitioning (α = 0.5).

A total of 456 OCTA images from patients with 7 retinal pathologies, with diabetic retinopathy (31.1%) and normal cases (25.2%) comprising the majority, sourced from the public OCTA-500 data set (n = 300) and a private collection from the University of Illinois Chicago (n = 156).

Five FL aggregation strategies (federated averaging [FedAvg], federated proximal [FedProx], federated magnetic resonance imaging [FedMRI], federated Adagrad, and federated Yogi) were systematically evaluated across multiple optimization dimensions: 7 architecture configurations spanning vision transformers, established convolutional neural networks, and hybrid models; 5 transfer learning freezing strategies; 3 local epoch configurations (2, 5, and 10); and scalability analysis across 2, 3, and 5-client federations. Security mechanisms including differential privacy (ε = 1.0-8.0) and secure aggregation were integrated and evaluated. Performance was assessed across 3 classification scenarios: 7-class, 4-class modified, and 4-class streamlined.

Classification accuracy, receiver-operating-characteristic area under the curve (ROC-AUC), and macro-averaged F1-score with comprehensive privacy-utility analysis and computational efficiency metrics.

Under controlled conditions, FL achieved superior performance in simplified classifications, with FedAvg, FedProx, and FedMRI reaching 72.09% accuracy versus 69.77% centralized training. Comprehensive optimization identified DenseNet121 as optimal architecture (79.55% accuracy, 89.68% ROC-AUC), with "most" freezing strategy (75% frozen layers) providing 60% training time reduction while maintaining superior performance. Federated proximal demonstrated exceptional resilience to heterogeneity (-11.7% degradation). Bonawitz secure aggregation achieved optimal privacy-utility balance (63.64% accuracy with cryptographic guarantees), whereas differential privacy maintained clinical utility under moderate constraints (ε ≈ 4-6).

This systematic evaluation establishes FL as a comprehensive solution for privacy-preserving multi-institutional OCTA-based disease classification, with careful architectural selection, optimization strategies, and security mechanisms enabling performance that matches or exceeds centralized approaches while maintaining regulatory compliance and clinical utility.

The authors have no proprietary or commercial interest in any materials discussed in this article.