Artificial Intelligence May Help Doctors Screen Diabetics for Eye Disorders
Doctors may soon begin using artificial intelligence to quickly diagnose eye disorders.
Diabetic retinopathy, often abbreviated DR, is a preventable blinding disease that affects about one-third of all people with diabetes mellitus worldwide. To screen patients for the disease, eye doctors take pictures of patients’ retinas and score abnormal changes. But this requires expensive retinal cameras and highly trained eye specialists -- not to mention the waiting list for an appointment with one.
Two recent studies, from Australia and India, want to rewrite this practice. Both aim to evaluate whether an artificial intelligence system could speed up diagnosis by allowing nonspecialists to screen patients for referral to the ophthalmologists.
“To try and bridge this gap we developed and validated an AI-based retinal diagnosis system for DR and other common blinding eye diseases, namely, glaucoma, age-related macular degeneration, and cataract,” said Stuart Keel, lead author of the Australian study published in March in the journal Scientific Reports.
Keel and his collaborators recruited 96 patients and shot photos of their retinas using an automated retinal camera. They then separated the patients into two groups. One set’s images were uploaded to an online AI platform created by a Chinese firm, which can analyze the images to grade abnormalities found in the images of the blood vessels of the eye and generate a report with referral recommendations.
Seven minutes after the photoshoot, these patients received a diagnosis with 93.7 percent accuracy and 92.3 percent sensitivity, roughly equivalent to a specialist's performance. Patients who needed further attention were referred to ophthalmologists for additional tests. The other set waited two weeks for results, until their images were manually analyzed.
In a similar small trial, researchers from Dr. Mohan's Diabetes Specialities Centre and Madras Diabetes Research Foundation, Chennai, India, took a different approach. In place of an expensive retinal camera, they chose to use a smartphone outfitted with a patented optics system that can capture retinal images, at about one-tenth the cost of the camera Keel's team used, according to Rajalakshmi Ramachandran, lead author of the study published in the journal Eye.
In the Indian trial, technicians imaged the eyes of 301 patients using the customized smartphone camera. They then separated the images into two sets, sending one to the ophthalmologists and the other to another cloud-based AI platform created by an American firm for analysis. The software diagnosed diabetic retinopathy with 95.8 percent sensitivity and 80.2 percent accuracy. As was the case with the Australian system, the system performed about as well as the manual diagnostic approach.
If approved by the country’s regulatory bodies, the Indian study's approach could be used to screen patients in remote villages with less to no supervision by eye doctors. The new system, said Ramachandran, could be very useful to effectively screen large numbers of people for diabetic retinopathy.
However, according to Ramachandran, scaling up the effort might require an offline-mode of AI-based diagnosis in countries like India where internet connections are often unreliable. This requires the development of offline databases for the AI to access data that are currently available on the cloud.
Unlike the Indian study, the Australian study scored the satisfaction of patients for an AI-based DR screening. While 80% of the patients reported their preference for AI over human doctors, the rest felt they trusted the human doctors more.
“A key challenge to the clinical adoption of this AI-based technology relates to a mindset shift in how patients and clinicians entrust clinical care to machines,” said Keel. Both groups are expanding their studies further.
“This is definitely a novel area that will innovate how screening is done and make it more affordable,” said Sobha Sivaprasad of University College London and a consultant at the U.K.'s Royal College of Ophthalmologists. However, she said, any AI-based approaches would need algorithms that are able to handle the array of potentially complicating issues that patients may present to the software.