Skip to content

Artificial intelligence technology provides potentially promising tool for diabetic retinopathy screening.

Researchers based at the Singapore National Eye Center have reported their findings on a novel “deep learning system” (DLS) – a machine learning technology with potential for screening diabetic retinopathy (DR) and related eye diseases. The research, published in the Journal of the American Medical Association (2017; 318 (22): 2211-2223) indicates that the DLS technology possesses a high sensitivity and specificity for identifying DR and other eye disease following an evaluation of a large set of retinal images from multi-ethnic cohorts of patients with diabetes. The researchers believe that the technology may provide a significant tool in efficiently screening retinopathy in an increasing global diabetic population.


According to the researchers, approximately 600 million people will be diagnosed with diabetes by 2040, with an estimated one third of diabetics suffering DR. Early diagnosis and treatment of DR is acknowledged as a key determinant in reducing visual impairment and blindness however, there is a significant numbers problem. The numbers of diabetic patients at risk of DR dwarf the number of qualified retinal image graders and ophthalmologists available to assess risk of DR. Consequently, technology enabled methods will be required in order to manage the challenge. One such method, DLS, uses artificial intelligence and machine learning methods to iterate computer algorithms trained to detect those retinal images that require referral to an ophthalmic clinician. The performance of the current system in detecting referable DR, vision-threatening DR, possible glaucoma, and age-related macular degeneration (AMD) was assessed after using a large dataset of retinal images for training.


The researchers outlined that close to half a million images from multi-ethnic cohorts were used to demonstrate that the DLS process had both high sensitivity and high specificity for recognizing referable DR, vision-threatening DR, and other ocular diseases, including potential glaucoma and AMD. The researchers stated that the “performance of the DLS was comparable and clinically acceptable to the current model based on assessment of retinal images by trained professional graders and showed consistency in 10 external validation datasets of multiple ethnicities and settings, using diverse reference standards in assessment of diabetic retinopathy by professional graders, optometrists, or retinal specialists.” A key objective in these and other DLS technologies is to ensure there is no degradation in health outcomes, which essentially translates to ensuring that the false-negative rate is no worse than human assessment by trained professional graders. In this respect, the researchers suggest that the DLS could be used with a semi-automated model for first-line screening followed by human grading for patients who test positive. Such an approach may radically increase through-put times for retinal image assessment in parallel with considerably reducing costs and facilitating earlier diagnosis and earlier treatment.