Federated Learning-Driven System for Improved Diabetic Retinopathy Diagnosis

Overview

Diabetic retinopathy is the leading cause of vision loss in working-age adults worldwide, yet under-resourced regions lack sufficient ophthalmologists for accurate early detection and prevention. Current state-of-the-art deep-learning systems for diabetic retinopathy diagnosis struggle in real-world application at under-resourced institutions due to limited generalizability and long inference times. Training a global model on centralized data across different medical institutions can address the challenges of data inadequacy and diversity, but faces restrictions of patient-privacy and data-ownership due to direct data transfer. This paper explores using a novel federated learning system for diabetic retinopathy diagnosis to address these challenges while enabling a multi-institutional collaborative approach. Our system trains models using updates from local data rather than directly shared medical data to preserve patient privacy and address confidentiality concerns. We simulated a federated learning diagnostic system with two well-resourced institutions and one under-resourced institution (lower image quality) using a separate dataset to represent each hospital and a local EfficientNetB0 architecture training at each hospital’s local dataset. The federated model was updated by all three of the local models at a central server. The federated model outperformed the three local models to improve diabetic retinopathy diagnosis accuracy with a 93.21% five-category-classification accuracy on an independent test-set of 6500 high-quality fundus images. The federated model also outperformed all local models on each institution’s test-sets. This included the under-resourced institution’s test-sets with lower-quality images in which the federated model achieved 91.05% accuracy. The study demonstrates the ability of a federated system to improve diabetic retinopathy diagnostic accuracy and generalizability by leveraging knowledge learned from different institutions. It also demonstrates the effectiveness of a federated system in generalizing to lower-quality fundus images at under-resourced institutions. As a proof-of-concept, the federated model was deployed to a free iOS mobile application and web application, allowing users to upload fundus images for accurate five-category-classification of diabetic retinopathy. After further testing and development, the federated learning-driven system has potential to revolutionize diabetic retinopathy diagnosis for under-resourced communities around the world with high generalizability while preserving patient privacy.

Competition history

• AJAS 2025 Category not listed

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Browse more like this

• All Biomedical and Health Sciences projects
• AJAS Biomedical and Health Sciences projects
• Every AJAS 2025 project