Screening of Diabetes and Hypertension Based on Retinal Fundus Photographs Using Deep Learning

ABSTRACT Purpose The aim of this study was to use deep learning to screen for hypertension and diabetes based on retinal fundus images. Methods We collected 1160 retinal photographs which included 580 from patients with a diagnosis of hypertension or diabetes and 580 from normotensive and non-diabetic control. We divided this image dataset into (i) a development dataset to develop model and (ii) test dataset which were not present during the training process to assess model’s performance. A binary classification model was trained by fine-tuning the classifier and the last convolution layer of deep residual network. Precision, recall, the area under the ROC (AUC), and the area under the Precision-Recall curve (AUPR) were used to evaluate the performance of the learned model. Results When we used 3-channel color retinal photographs to train and test model, its prediction precision for diabetes or hypertension was 65.3%, the recall was 82.5%, the AUC was 0.745, and the AUPR was 0.742. When we used grayscale retinal photographs to train and test model, its prediction precision was 70.0%, the recall was 87.5%, the AUC was 0.803, and the AUPR was 0.779. Conclusions Our study shows that trained deep learning model based on the retinal fundus photographs alone can be used to screen for diabetes and hypertension, although its current performance was not ideal.