Federated Learning for Diabetic Retinopathy Screening Across Regional Hospitals

Abstract

Training accurate retinopathy screening models requires large image collections, but hospitals are often unable to share patient data. This study trains a convolutional classifier with federated averaging across four regional hospitals, so that only model updates leave each site. Each hospital contributed between 2,100 and 5,600 graded fundus images. The federated model reached an area under the ROC curve of 0.92 for referable retinopathy, close to the 0.93 achieved by a model trained on pooled data and well above the 0.84 average of models trained at a single site. Sites with fewer images benefited the most. The results show that federated learning can deliver near-centralised accuracy while keeping retinal images inside each hospital.