Federated Transfer Learning For Diabetic Retinopathy Detection Using CNN Architectures
Mohammad Nasajpour, Mahmut Karakaya, Seyedamin Pouriyeh, Reza M. Parizi
Abstract
Diabetic Retinopathy is a complication of diabetes that could cause vision loss or even blindness if not detected in the early stages. As a result, having a regular eye exam is critical for maintaining a healthy retina and preventing damage to the eyes. Due to the lack of ophthalmologists in developing countries, there should be a faster approach to analyze the condition of these fundus images collected by different opticians. In this case, developing deep learning approaches comes in handy to enhance the effectiveness of Diabetic Retinopathy (DR) diagnosis accurately. Our main motivation is to develop a system that could manage various medical institutions. The critical part is to preserve privacy while training such deep learning models, which Federated Learning enables a decentralized training method by only sharing the parameters not the actual data. This study investigates three main models using standard transfer learning, Federated Averaging, and Federated Proximal frameworks. We demonstrate that our three models, including standard, FedAVG, and FedProx, were able to detect DR or non-DR images with the accuracy of 92.19%, 90.07%, and 85.81% respectively.