Researchers at the University of Michigan Kellogg Eye Center have demonstrated the effectiveness of a smartphone technology which combines high-quality retinal imaging with artificial intelligence (AI) to enable the early detection of diabetic retinopathy (DR).
Diabetic retinopathy (DR) is a complication of diabetes caused by high levels of blood sugar damaging the back of the eye. While it can take several years for the condition to reach a stage where it threatens a patient’s sight, when is does it can lead to permanent vision loss.
Lead author of the study and vitreoretinal surgeon Yannis Paulus said: “The key to preventing DR-related vision loss is early detection through regular screening. We think the key to that is bringing portable, easy-to-administer, reliable retinal screening to primary care doctors’ offices and health clinics.”
The team at the Kellogg Eye Center worked to develop a device which turns a smartphone into a retinal camera, known as the RetinaScope. While traditional retinal cameras are expensive, large, immobile devices which require specialist training to operate, RetinaScope is cheap and easy to use with no specialist training necessary.
RetinaScope was used to collect data from 69 adult patients with diabetes seen in the Kellogg Eye Center Retina Clinic. The images of the patients’ retinas taken after induced pupillary dilation were analysed with an AI eye screening system known as EyeArt, which graded them as referral-warranted diabetic retinopathy (RWDR) or non-referral-warranted DR.
The same images were also independently evaluated by two expert readers trained to recognise signs of DR.
These results were assessed alongside previously-recorded results of dilated slit-lamp fundus examinations by their treating clinicians, which had confirmed RWDR in 53 of the subjects. A silt-lamp exam is a standard diagnostic procedure, also known as a biomicroscopy, where a doctor shines a bright light into a patient’s medically dilated eye and examines it through a microscope to look for abnormalities.
The study measured two factors: whether the AI screening method was sensitive enough to detect disease, and if it was specific enough to confirm its absence.
The AI interpretation had a sensitivity of 86.8%, well above the 80% recommended for an ophthalmic screening device, and specificity of 73.3%. While one of the human image graders did achieve a significantly higher level of sensitivity at 96.2%, both had far lower specificity, at 40% and 46.7%.
Encouraged by these findings, Paulus and his team are now looking into how to improve the technology, notably seeking a way to make it work without inducing pupillary dilation in patients. They will also be seeking US Food and Drug Administration (FDA) clearance to roll out the test commercially.
“We’re focused on overcoming patients’ reluctance to seek DR screening by bringing it to them, making it easy, immediate, and available in a familiar clinical environment,” Paulus said.