Federated Learning for Multi-Disease Ophthalmic Diagnostics Using OCT Angiography.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinal disease classification using OCT angiography (OCTA) is crucial for diagnosis and treatment.
• Federated learning (FL) offers a privacy-preserving approach for multi-institutional data analysis.
• Challenges exist in optimizing FL for heterogeneous medical data.

Purpose of the Study:

• To systematically evaluate federated learning (FL) strategies for multi-disease retinal classification using OCT angiography (OCTA).
• To establish foundational feasibility and optimize FL performance under realistic heterogeneous conditions.
• To ensure privacy preservation in multi-institutional OCTA data analysis.

Main Methods:

• A retrospective multi-center FL study with a 2-part experimental design (homogeneous and heterogeneous conditions).
• Evaluation of five FL aggregation strategies (FedAvg, FedProx, FedMRI, FedAdagrad, FedYogi) across various architectures, transfer learning, and local epoch configurations.
• Integration and assessment of security mechanisms like differential privacy and secure aggregation.

Main Results:

• FL achieved superior performance over centralized training in simplified classifications (72.09% accuracy).
• DenseNet121 architecture with specific freezing strategies optimized performance (79.55% accuracy, 89.68% ROC-AUC).
• Federated proximal demonstrated resilience to heterogeneity; secure aggregation balanced privacy and utility.

Conclusions:

• Federated learning (FL) provides a comprehensive, privacy-preserving solution for multi-institutional OCTA-based disease classification.
• Optimized FL strategies can match or exceed centralized approaches.
• Careful selection of architecture, optimization, and security mechanisms is key for clinical utility and regulatory compliance.