Predicting Coronary Artery Calcium Score from Retinal Fundus Photographs Using Convolutional Neural Networks

Abstract

Coronary Artery Calcium Score (CACS) is a prognostic indicator for coronary atherosclerosis that can cause a stroke or heart attack. Cardiac computed tomography (CT) is widely used to calculate CACS. For asymptomatic patients, however, a CT-based screening tes is not recommended due to an unnecessary exposure to radiation and high cost. In this paper, we propose a deep learning approach to predict CACS from retinal fundus photographs. Our approach is non-invasive and can observe blood vessels without any side effects. Contrasted to other approaches, we can predict CACS directly using only retinal fundus photographs without the electronic health record (EHR) data. To overcome data deficiency, we train deep convolutional neural nets (CNNs) with retinal fundus images for predicting auxiliary EHR data related to CACS. In addition, we employ a task-specific augmentation method that resolves flare phenomenon typically occurred in a retinal fundus image. Our empirical results indicate that the use of auxiliary EHR data improves the CACS prediction performance by 4.2%, and flare augmentation by 2.4% on area under the ROC curve (AUC). Applying both methods results in an overall 6.2% improvement. In the light of feature extraction and inference uncertainty, our deep learning models can predict CACS using only retinal fundus images and identify individuals with a cardiovascular disease.