In-person verification of deep learning algorithm for diabetic retinopathy screening using different techniques across fundus image devices

Abstract

Purpose: To evaluate the clinical performance of an automated diabetic retinopathy (DR) screening model to detect referable cases at Siriraj Hospital, Bangkok, Thailand. Methods: A retrospective review of two sets of fundus photographs (Eidon and Nidek) was undertaken. The images were classified by DR staging prior to the development of a DR screening model. In a prospective cross-sectional enrollment of patients with diabetes, automated detection of referable DR was compared with the results of the gold standard, a dilated fundus examination. Results: The study analyzed 2533 Nidek fundus images and 1989 Eidon images. The sensitivities calculated for the Nidek and Eidon images were 0.93 and 0.88 and the specificities were 0.91 and 0.85, respectively. In a clinical verification phase using 982 Nidek and 674 Eidon photographs, the calculated sensitivities and specificities were 0.86 and 0.92 for Nidek along with 0.92 and 0.84 for Eidon, respectively. The 60°-field images from the Eidon yielded a more desirable performance in differentiating referable DR than did the corresponding images from the Nidek. Conclusions: A conventional fundus examination requires intense healthcare resources. It is time consuming and possibly leads to unavoidable human errors. The deep learning algorithm for the detection of referable DR exhibited a favorable performance and is a promising alternative for DR screening. However, variations in the color and pixels of photographs can cause differences in sensitivity and specificity. The image angle and poor quality of fundus photographs were the main limitations of the automated method. Translational Relevance: The deep learning algorithm, developed from basic research of image processing, was applied to detect referable DR in a real-word clinical care setting.