Deep Learning Tools Can Accurately Identify Common Eye Disease

Researchers at New York Eye and Ear Infirmary of Mount Sinai (NYEE) have developed deep learning tools that can detect age-related macular degeneration (AMD), a leading cause of blindness in the US.

In patients with AMD, the central area of the retina called the macula deteriorates, causing blurry vision that can worsen significantly over time.

To help providers predict the risk of AMD progression and severity, NYEE researchers set out to build a deep learning screening and prediction model using data from the Age Related Eye Disease Study, a large study of AMD over 15 years.

Patients between 55 and 80 years old were grouped into categories for normal, early, intermediate, and advanced or late-stage AMD. For screening, researchers took 116,875 color fundus photos, which capture the interior surface of the eye, from 4,139 participants and trained the algorithm to classify them as “no,” “early,” “intermediate,” or “advanced” AMD along a 12-level severity scale to match the findings of human experts.

The algorithm achieved an overall accuracy of 98 percent when matching decisions of experts.

“We are excited to have built a deep learning form of AI that can be trained to match the performance of a human expert to accurately diagnose AMD grade and stage based on scanning retinal photographs, without using other information,” said lead researcher R. Theodore Smith, MD, PhD, Professor of Ophthalmology at the Icahn School of Medicine at Mount Sinai.

“This is an important step in identifying those at risk for late-stage AMD and may allow them to get quick referral to an eye specialist for timely, preventive treatment.”

Researchers then took the severity scores and combined them with patients’ sociodemographic clinical data, including age, gender, and medical history, as well as other imaging data in a second algorithm to predict AMD progression. Specifically, they developed the algorithm to identify risk for progression to late AMD within one to two years.

They trained and validated the model on 923 participants who had AMD progression within two years, 901 patients who had progression within one year, and 2,840 patients who did not progress within two years. The AI model further refined the risk of progression to late AMD so that researchers were able to predict the exact type of progression of late AMD – either dry or wet.

Dry AMD develops more slowly, with layers of the macula becoming progressively thin and losing function. Wet AMD is more rapid, and involves abnormal blood vessels forming behind the retina and leaking.

“The prediction program will produce a report that can help eye doctors counsel AMD patients on their risk for progression based on their retinal photographs and other lifestyle (diet and smoking) and demographic variables: age, gender, and medical history. The ophthalmologist can then recommend changes in modifiable factors in consultation with family and the primary care physician, and patients at high risk can be followed up with sooner,” said Smith.

These new algorithms could help providers manage eye disease patients and offer more preventive care, the researchers noted.

“The proposed noninvasive technology thus proceeds in two steps: we first screen high volumes of patients in the community to find the at-risk patients with intermediate and advanced AMD for referral to an ophthalmologist, and second, we help the eye doctor manage these patients by predicting if they will develop late AMD in one to two years,” said Smith.

“This can allow screening to take place more efficiently and cost-effectively in primary care clinics, with detection of a much smaller at-risk group for referral to specialty care.”

New York Eye and Ear Infirmary of Mount Sinai has been testing the algorithm for detection and staging of AMD in its eye clinics and has seen favorable results. Once these systems are brought on line for widespread use with automated, inexpensive cameras at primary care facilities, patients will have access to quick, non-invasive screening for blinding eye disease.

The technique could also serve as a useful tool in the current COVID-19 outbreak.

“This algorithm can easily be applied in the ophthalmology telemedicine landscape as the practice of medicine transforms under the impact of the COVID pandemic to embrace ‘medicine at a distance.’ For example, our large ambulatory facilities can strategically place teleophthalmology kiosks with inexpensive cameras that take these retinal images to screen underserved populations for AMD,” researchers said.

“The AI algorithm would instantly generate results, so patients get immediate diagnosis, and if they need additional care, they could have a same-day follow-up at a nearby ophthalmic center. This may become an important and cost-effective tool for high-risk or low-income groups who may not have direct or frequent access to eye screening, as early detection is critical to preventing AMD.”