Collaborative Differentially Private Federated Learning Framework for the Prediction of Diabetic Retinopathy

Abstract

The diagnosis of diabetic retinopathy may be streamlined and expedited with the help of deep learning, which is an efficient way to help an eye specialist examine the enormous amount of retinal images. For these strategies to be effective, big datasets must be consolidated and used for training. Medical data privacy laws frequently make it impossible to gather and share patient data on a single system. In this paper, we introduce a collaborative differentially private federated learning system that enables deep learning image analysis without transferring patient data between healthcare organizations. We investigated four different machine learning algorithms—AlexNet, ResNet50, SqueezeNet1.1, and VGG16—for varying amounts of noise using a dataset of 35120 retina images divided into five classes—No Diabetic Retinopathy, Mild, Moderate, Severe, and Proliferative Diabetic Retinopathy (PDR). Our ResNet50 model outperformed the state-of-the-art diabetic retinopathy prediction models with an accuracy 83.05 % when we added no noise, and with an accuracy 79.35% with a noise multiplier of 8.0. By including our checkpoint techniques, we have reduced the total communication overhead by 49 % when compared to federated learning without checkpoints.

Ähnliche Arbeiten