Screening for diabetic retinopathy with a deep learning network: Evaluation of retinal images from the Qatar Biobank

Diabetic retinopathy (DR) is the leading cause of impaired vision and blindness in working age adults. Regular eye screening is advised to prevent progression to blindness. However, screening programs are labor- and capital-intensive and suffer from limited access to trained professionals. Artificial intelligence has been employed to develop algorithms for automated classification of retinal images. We developed an ensemble model of two deep convolutional neural networks to score DR in fundus images. The model was trained using tens of thousands images from a public database and allows the determination of DR stage, as well as referable DR. The model was validated on 1748 fundus images from the Messidor-2 database. All 190 patients with referable DR were successfully detected by our deep learning (DL) model (sensitivity 100%, 95%CI 98.1%-100%). The model labelled 444 of the 684 diabetes patients without referable DR accordingly (specificity 65.0%, 95%CI 61.2%-68.5%). The DL model was further evaluated on retinal images from the Qatar Biobank. Images from 740 individuals enrolled in the Qatar Biobank were received. Image quality was poor in 72 individuals leaving 668 individuals with manually graded images which were independently analyzed using the DL model. The model scored the DR class with an accuracy of 0.88 and a precision of 0.95. Forty-two individuals (6.3%) have some form of DR and referable DR was identified in twenty-six individuals. The model for referable DR had an accuracy of 0.90 and a precision of 0.97. In conclusion, our DL model has been successfully tested on fundus images from the Qatar Biobank. The Automated Retinal Image Analysis System (ARIAS) is promising in relation to supporting medical professionals to undertake cost-effective screening, especially in the context of large population screening programs.