A federal learning-driven artificial intelligence framework for fundus image myopia diagnosis

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Myopia is a significant global public health concern.
• Accurate classification of fundus images is crucial for diagnosing myopia and related pathologies.
• Existing methods face challenges with data privacy, heterogeneity, and class imbalance in multi-institutional settings.

Purpose of the Study:

• To develop a privacy-preserving federated learning (FL) framework for the triple classification of fundus images (normal, myopia, pathological myopia).
• To ensure the framework generalizes across institutions despite data heterogeneity and class imbalance.
• To maintain data privacy during collaborative model training.

Main Methods:

• Proposed a novel FL framework incorporating a genetic algorithm-inspired dynamic aggregator (FedProx_GA), a distance-aware attention module (OptiFocus), and a class-frequency dynamic loss.
• Trained and evaluated the framework on 1,279 fundus images from three heterogeneous medical centers.
• Compared performance against standard FL baselines using metrics like AUC, accuracy, sensitivity, and specificity.

Main Results:

• Achieved a high Area Under the Curve (AUC) of 0.9889, nearing the performance of centralized data processing.
• Significantly outperformed conventional FL methods in classifying fundus images.
• Demonstrated robust cross-center generalization with high sensitivity (0.9346) and specificity (0.9673), effectively handling data heterogeneity and class imbalance without compromising privacy.

Conclusions:

• The developed FL framework offers an effective, privacy-preserving solution for collaborative ophthalmic AI.
• The framework shows strong potential for multi-institutional clinical deployment in diagnosing myopia.
• Future research should involve prospective validation with larger, diverse cohorts.