An automated grading system for detection of vision-threatening referable diabetic retinopathy on the basis of color fundus photographs

OBJECTIVE The goal of this study was to describe the development and validation of an artificial intelligence-based, deep learning algorithm (DLA) for the detection of referable diabetic retinopathy (DR). RESEARCH DESIGN AND METHODS A DLA using a convolutional neural network was developed for automated detection of vision-threatening referable DR (preproliferative DR or worse, diabetic macular edema, or both). The DLA was tested by using a set of 106,244 nonstereoscopic retinal images. A panel of ophthalmologists graded DR severity in retinal photographs included in the development and internal validation data sets (n = 71,043); a reference standard grading was assigned once three graders achieved consistent grading outcomes. For external validation, we tested our DLA using 35,201 images of 14,520 eyes (904 eyes with any DR; 401 eyes with vision-threatening referable DR) from population-based cohorts of Malays, Caucasian Australians, and Indigenous Australians. RESULTS Among the 71,043 retinal images in the training and validation data sets, 12,329 showed vision-threatening referable DR. In the internal validation data set, the area under the curve (AUC), sensitivity, and specificity of the DLA for vision-threatening referable DR were 0.989, 97.0%, and 91.4%, respectively. Testing against the independent, multiethnic data set achieved an AUC, sensitivity, and specificity of 0.955, 92.5%, and 98.5%, respectively. Among false-positive cases, 85.6% were due to a misclassification of mild or moderate DR. Undetected intraretinal microvascular abnormalities accounted for 77.3% of all false-negative cases. CONCLUSIONS This artificial intelligence-based DLA can be used with high accuracy in the detection of vision-threatening referable DR in retinal images. This technology offers potential to increase the efficiency and accessibility of DR screening programs.