The Application of Deep Learning System to Screen for Diabetic Retinopathy in an Underprivileged African Population with Diabetes

Abstract

Purpose : Zambia is a low-income developing country overwhelmed with widespread rural poverty, ranked 191^th for gross domestic product per capita in 2018. Diabetes exerts an emerging burden in the region. Related complications such as diabetic retinopathy (DR) are expected to increase dramatically in prevalence. Challenged by shortage of ophthalmic services and poor accessibility to DR screening, the application of artificial intelligence (AI) using deep learning (DL) may be an alternative solution. This study aims to evaluate the real-world clinical effectiveness of a DL system in screening for DR and vision-threatening DR (VTDR) in the Zambian population with diabetes.

Methods : A total of 4513 images from 3101 eyes of 1578 Zambians with diabetes were prospectively recruited for this study. Two-field color 45-degree retinal fundus photographs were captured for each eye and graded according to International Classification of Diabetic Retinopathy Severity scale. Referable DR was defined as moderate non-proliferative DR (NPDR) or worse, diabetic macular edema and ungradable images; VTDR was designated as severe NPDR and proliferative DR. With reference to the retinal specialists’ grading, we calculated the area under the receiver operating curve (AUC), sensitivity and specificity for referable DR, and the detection rate of VTDR, using an Ensemble convolutional neural network.

Results : The prevalence of referable DR and VTDR in the Zambian population with diabetes were 23.1% and 6.4%, respectively. The AUC of the DL system for referable DR was 0.973 (95%CI, 0.968-0.977), with corresponding sensitivity of 91.35% (95%CI, 89.12-93.31) and specificity of 89.26% (95%CI, 88-90.48). VTDR detection rate was 93.91% (95%CI, 93.07-94.75).

Conclusions : The developed DL system shows clinically acceptable performance in detection of referable DR and VTDR for the Zambian population. This demonstrates the potential application to adopt such sophisticated cutting-edge AI technology for the underprivileged population in developing countries, to ultimately reduce the incidence of preventable blindness.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.