According to study results published in Diabetes Care, an artificial intelligence-based deep learning algorithm can accurately provide automated detection of vision-threatening referable diabetic retinopathy in retinal images.
Researchers developed an algorithm in an effort to better detect forms of vision-threatening referable diabetic retinopathy (preproliferative, diabetic macular edema, or both). They tested the deep learning algorithm on 71,043 retinal images and a panel of 21 ophthalmologists graded each image for diabetic retinopathy severity.
Ophthalmologists graded all images in the local data set 3 to 8 times, with an average agreement of 87.3% and a range of 78.6% to 97.2% for severity grading. Each ophthalmologist graded between 137 and 21,024 photographs (median 3501). Of all images, 6.1% were deemed poor quality, leaving 66,790 images available for conclusive severity grading; 8000 of these photographs were held for internal validation while the rest were graded as part of the training data set.
Of the images included in the training data set, 18.5% had vision-threatening referable diabetic retinopathy and 27.5% showed evidence of diabetic macular edema. In the internal validation set, the researchers reported an area under the curve of .989, a sensitivity rating of 97%, and 91.4% specificity of the deep learning algorithm for vision-threatening referable diabetic retinopathy.
The ophthalmologists also graded retinal images from 13,657 eyes in an external data set from 3 population-based studies. Area under the curve for this external validation data was .955, sensitivity was 92.5%, and specificity was 98.5%.
Of false-positive cases, 93.2% were attributed to a misclassification of mild or moderate diabetic retinopathy. In addition, 77.3% of false-negative cases were due to undetected intraretinal microvascular abnormalities.
This study had certain limitations, including that all previously reported deep learning algorithms were based on a definition of severity that warranted notably earlier referral for diabetic retinopathy. However, no specific or effective eye care management is currently available for patients with mild diabetic retinopathy, and adoption of this clinical criterion may lead to over-referral and a strain on eye care resources.
Based on their findings, the researchers said, “…this [artificial intelligence]-based [deep learning algorithm] shows robust performance in the detection of vision-threatening referable [diabetic retinopathy]… This technology offers [the] potential to increase the efficiency and accessibility of [diabetic retinopathy] screening programs, particularly in developing countries… and in minority and underserved populations.”
Li Z, Keel S, Liu C, et al. An automated grading system for detection of vision-threatening referable diabetic retinopathy on the basis of color fundus photographs [published online October 1, 2018]. Diabetes Care. doi:10.2337/dc18-0147