Artificial intelligence-based deep learning algorithms that scan retinal images can be used to detect referable diabetic retinopathy (DR) with high accuracy, according to a recent study presented earlier this week at the ARVO annual meeting.
A group of researchers developed a cloud-based deep learning system for automated detection of DR. It was then tested on 71,043 non-stereoscopic retinal images. Retinal photographs were graded for DR severity by a panel of ophthalmologists with a gold standard grading for each image assigned when three or more consistent grade outcomes were achieved. For external evaluation, the algorithm was tested using 13,406 retinal images from three-population based cohorts of Malay, Caucasian and Indigenous Australians. Sensitivity and specificity for the algorithm was 97% and 91.4%, respectively.
“Early detection of significant diabetic retinopathy is crucial,” says Joseph P. Shovlin, OD, of Scranton, PA. “Retinal images apparently can be evaluated using convolutional neuron network for detection of referable diabetic retinopathy (> pre-proliferative DR and diabetic changes with macular edema) with reasonably good sensitivity and specificity in validation of the dataset. The technology seems to offer significant potential for telemedicine to increase efficiency and accessibility for detection of diabetic retinal changes.”
|Keel S, Li Z, He Y, et al. The development and validation of a deep learning algorithm for referable diabetic retinopathy. ARVO 2018. Abstract 738.|