Collaborative, Privacy-Preserving Federated Learning Framework for the Detection of Diabetic Eye Diseases

Diabetes Mellitus is one of the chronic metabolic disorders that is easily identified by elevation in blood sugar levels. The high blood sugar levels over time can lead to substantial damage to the blood vessels, heart, eyes, kidneys, and nerves. Diabetic Eye Diseases (DED) such as Diabetic Macular Edema (DME), Diabetic Retinopathy (DR), and glaucoma, are serious complications of diabetes that can result in irreversible vision loss. Thus, affecting the health and quality of life of the population worldwide. The primary objective of the proposed work is to evaluate the efficacy of a privacy-preserved Federated Learning (FL) framework for detecting DEDs. To detect DEDs a collaborative and privacy-preserving Federated Deep Learning (FDL) framework using a lightweight MobileNetV2 architecture is proposed. The aggregation scheme selected for the FDL framework is FedProx and to achieve a high accuracy score the proximal term, µ is tuned with values ranging from µ= (0.01, 0.02, 0.03, 0.04 and 0.05). The impact of the variance in the proximal term on model performance is analyzed, and the results show that higher µ (0.03,0.4 and 0.5) values provide more stable and consistently high precision, recall, and F1-scores. The findings suggest that FedProx aggregation successfully stabilizes the performance of the framework with the least standard deviation of 0.035 in accuracy at µ = 0.05. The balance between accuracy and stability is a key factor in the successful application of federated learning in disease prediction.