Federated Learning for Early Detection of Diabetic Retinopathy in Distributed Healthcare Systems

Diabetic retinopathy (DR) is one of the leading
causes of preventable vision loss in the world, and early
detection is critical for an effective intervention. However, the
sensitive nature of patient data, together with regulatory
considerations, restricts the use of centralized model training of
healthcare data across institutions. This study presents a
federated learning (FL) framework, to train a deep
convolutional neural network(CNN) (EfficientNet-B0), to
provide remote, early detection of DR type using retinal fundus
images collected from multiple clinics that are distributed
geographically, without sharing un-identified raw health data.
The performance of the FL training-and-test scheme was
evaluated on a range of AUC, accuracy, sensitivity, and
specificity against a central, local only, ensemble, and pretrained
model. It is found that the FL model achieved an AUC
of 0.91, and accuracy of 88.3%. Federated learning is chosen
for this study to ensure patient privacy to handle distributed
non-IID data, and achieve near-centralized diagnostic
accuracy. This work also explores the future research
opportunities for federated learning, and suggests that
federated learning represents an advanced, scalable and
privacy-respecting avenue for the implementation of AIsupported
diagnostic imaging tools in the healthcare
distributed healthcare systems.