Diagnosing Diabetic Retinopathy With Artificial Intelligence
Diagnosing diabetic retinopathy is faster and more accurate than ever with the artificial intelligence used in Eyenuk’s EyeArt system, the first of its kind approved by the FDA.
By Todd Farley
Doctors say that diabetes is fast becoming a worldwide public health epidemic with more than 400 million total cases around the globe, including 30 million Americans. It is the seventh leading cause of death in the world and can lead to heart disease, stroke, and nerve and kidney damage. It can also lead to diabetic retinopathy (DR), an eye disease that is the leading cause of blindness in working-age Americans.
Diabetes Can Lead to DR
“Diabetes is a condition that ultimately affects blood vessels all over your body, including in your eye,” says Dr. SriniVas Sadda, president and chief scientific officer of Doheny Eye Institute and professor of ophthalmology at UCLA. “DR is the direct consequence of diabetes affecting the retina.”
DR occurs when the blood vessels in the retina are adversely affected by too much blood sugar (glucose). This can lead to excess blood and fluid leaking into the retina from those damaged blood vessels and a corresponding dearth of oxygen and nutrients being delivered to the retina. While DR initially results in swollen or blocked blood vessels, there are not necessarily any visual symptoms. But as the disease progresses, sight is often negatively affected.
In its most advanced form — what is called the “proliferative stage” — new blood vessels grow rapidly but are too fragile and prone to leakage, thus exacerbating the problem of too much blood and fluid already in the eye.
“If you think of the eye like a camera, the retina is the film of that camera,” Dr. Sadda says. “Diabetes damages the film of our eye camera, if you will, and that is where it can cause problems with vision that can lead to blindness.”
Symptoms of DR
According to the Mayo Clinic, symptoms of advanced DR can include spots or strings floating in one’s vision (known as floaters), blurriness, impaired color sight, dark or empty areas in vision, and eventually, blindness. In the US, new cases of blindness in people between the ages of 20 and 74 result primarily from diabetes and DR. Fortunately, DR can be both prevented and treated.
Prevention & Treatment
“It’s true that we have effective treatments for DR, but as with most diseases, the treatments are most effective if given early on,” Dr. Sadda says. “Ninety percent of diabetic blindness can be prevented if it’s detected early and then treated in a timely manner.”
For that reason, the American Academy of Ophthalmology (AAO) recommends that patients with diabetes get regular eye exams every year. The problem is that there are more diabetics today than there are ophthalmologists, meaning there are more eye test results to look at than there are eye specialists to review them.
“The challenge,” Dr. Sadda explains, “is that, because diabetes is growing at an epidemic rate, health systems aren’t necessarily able to keep up with screening everyone who needs to be screened for DR.”
Telehealth Diagnosis
One method to expedite the screening process is through telehealth, whereby patients can have their retinas checked remotely. Through telehealth, many working-age Americans suffering from DR would not have to leave work to travel all the way to their ophthalmologist’s office to get a yearly screening. In addition, many others residing in isolated locations would not have to travel great distances to be seen.
“Typically, a camera is set up, perhaps at a primary care doctor’s office or a diabetes center,” Dr. Sadda says. “The camera captures pictures of a patient’s retina, which are sent to another location to be reviewed by experts well-trained to evaluate the images.”
While such telehealth screenings are prevalent in Europe, including in Iceland and as part of the UK’s National Health Service (NHS), they are rarely an option in the US. Rather, most eye screenings in the US still happen the old-fashioned way.
“Even now, the bulk of DR screenings in the US happen in person at eye care providers’ offices,” Dr. Sadda says. “Telehealth screenings are not broadly deployed here because we don’t have a national health system, per se.”
Too Many Images, Not Enough Reviewers
Regardless of how those retinal images are captured — whether via telehealth screenings or on-site doctor’s visits — the real issue regarding DR scans is the large number of them. Because there are so many patients with diabetes getting so many yearly DR eye exams, countless images must be assessed.
“All of these images must be sent to well-trained humans for assessment, which means there is an expense involved,” Dr. Sadda says. “And there’s also a delay, so the patient doesn’t know right after the image is captured if they have an issue. There is no instant feedback.”
In trying to alleviate the glut of scans to be assessed, eyecare specialists are not looking to the past for help but to the future: artificial intelligence (AI). Computer algorithms can be trained based on the decades of knowledge from human assessments of eye scans. By feeding computers thousands of images that human assessors have reviewed to determine whether they show evidence of DR, AI algorithms learn how to do the same thing on their own.
“This is how AI could improve the screening paradigm, making screening timelier and more cost effective,” Dr. Sadda says. “Perhaps AI can also make screening deployable on a much larger scale.”
Cutting-Edge AI for DR Diagnosis
The first technology incorporating AI to screen for DR was the IDx-DR platform, which was approved by the FDA in 2018. This summer, Eyenuk’s EyeArt system became the first AI technology cleared by the FDA for use in the “autonomous detection of both more-than-mild and vision-threatening diabetic retinopathy.”
EyeArt has been tested on more than half a million patients and almost two million retinal images. Its success was originally seen in a study conducted by Dr. Sadda and the Doheny Eye Institute, which was featured at the 2019 AAO conference in San Francisco.1
In the study, EyeArt screened nearly 900 diabetic patients at 15 medical locations across the US, and its results were reviewed by certified human graders for clinical accuracy. Results showed that EyeArt’s sensitivity, meaning its ability to identify the presence of eye disease, was 95.5%. EyeArt’s specificity, meaning its ability to identify the absence of disease, was found to be 86%. For the small number of eyes that needed to be dilated in order to acquire clear images, EyeArt’s sensitivity remained the same while its specificity slightly improved.
“Diabetic patients already outnumber practicing ophthalmologists in the US, and that will just get worse,” Dr. Sadda says, “so accurate, real-time diagnostic tools like EyeArt hold great promise for patients with diabetes. Such a system is more accessible, and a prompt diagnosis by AI also means prompt identification of patients at risk of blindness, which helps get them to an ophthalmologist sooner.”
EyeArt is Fast & Accurate
The speed with which AI can make an initial diagnosis for a DR screening is vital. The ability to upload an image to the cloud and get results within a minute or two means patients will have an expedited doctor’s visit. Instead of having to wait days or weeks for the results of a screening, these more immediate AI responses mean patients would likely know within minutes whether their scans are clear or problematic. It also means that, if needed, a patient could schedule a follow-up appointment with an ophthalmologist before leaving the doctor’s office, making it less likely that they will forget the appointment or fail to follow up.
The Future is Now
Does using a computer algorithm to make an important medical decision seem troubling? It shouldn’t; it’s already the future. Consider the electronic devices many people tote about, like phones that check one’s heartbeat or blood oxygenation level. Much of that technology is AI-based, where a computer algorithm may suggest that a sign or symptom is sufficiently abnormal that it needs to be checked.
“It just shows the diagnostic power of AI,” Dr. Sadda says of EyeArt and other AI-based technologies. “Maybe these tools can’t make a final diagnosis because you’d need a physician to do that and to administer treatment, but many of these technologies can prove beneficial in identifying or screening issues.”
Dr. Sadda continues, “This is all just part of the trend in healthcare, and it’s probably not a bad solution either. If we, as consumers of healthcare, can take command of certain aspects of our own care, we might improve the efficiency of the healthcare system and drive down its overall costs.”
References
1. Campbell P. EyeArt Screening System Detects Diabetic Retinopathy. https://www.hcplive.com/view/eyeart-screening-diabetic-retinopathy. Published 2019. Updated October 14, 2019. Accessed January, 2021.
