The US Food and Drug Administration’s (FDA) approval of an algorithm for diabetic retinopathy (DR) in 2018 was the first approval of an autonomous artificial intelligence (AI) technology, and that’s not just in ophthalmology, or even just in medicine.1,2 It was the first, true, fully automated technology in a field that includes self-driving cars, manufacturing robots, and social media monitoring, says Michael D. Abràmoff, MD, PhD. 

Since then, a wealth of ophthalmic research has shown the rapid development of these AI technologies to monitor diseases, identify their development preclinically, and even, in some cases, extend the role of the ophthalmologist into the realms of neurological and systemic disease.2

“AI opens a world of opportunities for clinicians,” says Dr Abràmoff, an educator at the University of Iowa Hospital and Clinics in Iowa City, IA. He is also founder of Digital Diagnostics, the company behind that first FDA-cleared AI technology, IDx-DR. Dr Abràmoff sees AI enhancing clinician workflow and productivity, allowing existing staff to treat more patients, improving patient outcomes, and overcoming health disparities.

Continue Reading

Possibilities for AI implementation include ophthalmic screening, diagnostics, treatment, and delivering precision medicine for DR, diabetic macular edema (DME), age-related macular degeneration (AMD), glaucoma, retinopathy of prematurity, cataract, refractive surgery, retinal vein occlusion, and myopia, often using fundus images. Some AI innovations are already available. Others are still in early clinical research and testing.

But the potential is real, even in some of the more science fiction-esque ideas — such as research published last year in Scientific Reports that found deep learning models could reliably predict age and sex using only fundus images.3 Those same images could also provide “unique information” related to blood pressure, hemoglobin A1c, and relative fat mass, researchers report.3

This raises  the question: What else can be learned when physicians apply AI to images of the eye? How else could it benefit patients and ophthalmologists, both now and in the future?

Just What Is Deep Learning?

AI has been in the making for decades, Dr Abràmoff explains. He credits clinicians including Matthew Davis, Rick Ferris, Morton Goldberg, and Stuart Fine as the “giants who came before us.” But the difference in technology now, as opposed to, say, 20 years ago, is something called deep learning, according to Aaron Lee, MD, MSc, an assistant professor at University of Washington Medicine Department of Ophthalmology in Seattle, WA.

“Deep learning really was a revolutionary technology that came out of computer science,” Dr Lee explains. “We live in this completely electronic world now, where all the images that we take of our patients are stored as electronic files and all the data about the patients in our clinical encounters are electronically codified in our health records. That became the recipe for the revolution of artificial intelligence and ophthalmology.”

Ophthalmology is uniquely suited to deep learning because of how much photographic data ophthalmologists collect digitally, especially in retina. Those images provide AI algorithms exposure to a great deal of information that it can use to learn, and evolve. 

“This kind of technology learns from many, many data points, usually thousands, if not tens of thousands of images,” says T.Y. Alvin Liu, MD, assistant professor of ophthalmology at The Johns Hopkins Wilmer Eye Institute, Baltimore, MD. 

Dr. Liu is director of the newly established Wilmer Precision Ophthalmology Center of Excellence, which is part of the university-wide inHealth Initiative that brings together expertise from the university’s School of Medicine, School of Engineering, and Applied Physics Laboratory. Those involved currently have multiple on-going ophthalmology deep learning projects in progress, including research into an algorithm that can predict uveal melanoma with 90% accuracy just by looking at a picture of the tumor’s cancer cells.

“This is superhuman,” Dr. Liu says. “No pathologist can just look at some cells and predict the genetic makeup of a tumor.”

Other fields of medicine, such as radiology, cardiology, and internal medicine/general practice (which have the most AI technologies developed thus far, according to a 2020 study) employ AI as well, but ophthalmology has been at the forefront of research into the technology thanks to deep learning, Dr. Lee points out.4 

Picture “a big Venn diagram,” Dr Lee explains. “You have a big circle that says ‘artificial intelligence.’ And inside of that, there’s a smaller circle that says, ‘machine learning.’ And inside of that, there’s a smaller circle that says, ‘artificial neural networks.’ And inside of that is, ‘deep learning.’ So, deep learning is a special form of artificial neural networks.”

It’s called “deep” because of how many layers of computation each image is passed through, as opposed to only three layers of imaging in past artificial neural networks, he says.

Diabetic Eye Disease and AI

Say you’re not a retina specialist, but have patients with diabetes who might be at risk for DR. DR is the most common reason for vision loss in the more than 30 million Americans who have diabetes — and the leading cause of vision impairment and blindness among working-age adults.5 So it’s likely you’ve encountered patients at risk.  AI could provide you with a way to detect and monitor patients for DR without having to refer them to a subspecialist. 

All you would have to do is obtain images of a patient’s retina with a retinal camera, upload those images to a cloud server, and AI technology would take it from there. Ideally, these technologies can tell if your patient has DR or DME and you can make treatment or referral decisions from there. That’s precisely how the IDx-DR works, a device that uses an AI algorithm to analyze images with a high rate of accuracy, and it’s now on the market with another approved autonomous AI system EyeArt (Eyenuk), which received approval in 2020. The same year, the Centers for Medicare and Medicaid Services (CMS) announced that Medicare would reimburse for autonomous AI with a new code specifically created for the technology.

In DR, autonomous AI provides what Dr Lee calls a “leveler of clinical care.” He and colleagues published a study in Diabetes Care earlier this year that compared seven automated AI DR screening systems, finding wide variety in performances, but excellent outcomes in several.5

“I think in clinical practice in the United States, there’s a great variety and spectrum of clinical care,” he says. “Some people are experts at a cataract surgery, because they do a ton of cataract surgery, but maybe they’re not an expert at taking care of macular degeneration, because they just don’t see that often. They also may be in a situation where they don’t have access to that kind of expertise. I think what AI models can do is provide a form of that expertise in a way to elevate the standard of care across the country.”

It will also hopefully increase screening rates, Dr. Liu says, which might “translate to earlier treatments, and lower rates of visual impairments and blindness from diabetes on a population level.”

Cataract Surgery Workup and AI

Right now, approximately 70% of patients receive their desired refractive outcome after cataract removal and intraocular lens (IOL) implantation, says Aazim A. Siddiqui, MD, an ophthalmology resident at Albert Einstein College of Medicine in New York City. That translates to approximately 30 out of 100 patients who are not happy and might need further interventions, such as glasses or additional surgery, he points out.

“This is where artificial intelligence can really make a meaningful difference in closing the gap. I think the area of artificial intelligence and IOL calculations is an extremely promising landscape,” Dr. Siddiqui says. Last year, he and colleagues published a paper looking at AI’s use in cornea, refractive, and cataract surgery in Current Opinions in Ophthalmology, including sample calculations from, a free AI formula he and his team created, the Ladas Super Formula, after John Ladas, MD, PhD, who initially developed it.6,7 

Dr. Siddiqui advises ophthalmologists to see AI as a guide. It’s a way to achieve even better outcomes and higher rates of patient satisfaction — whether due to earlier screenings that save patients’ vision to the satisfactory refractive outcomes that prevent more intervention down the road.

“I think you can never replace the art of medicine, the surgical decision making, and the intuition we all develop as physicians as we go through our day-to-day, and learn from our own experiences,” he says.

Glaucoma and AI

Turns out, there’s more than one type of AI itself — and one of those could be particularly helpful in glaucoma, Dr Abràmoff says. Called assistive AI might offer opportunities to  glaucoma specialists, who may sometimes find themselves overwhelmed with data.The assistive AI could sift through all that information, he says, reaching a better-quality diagnosis from all imaging available and helping a clinician not miss anything vital.

Two examples from the literature demonstrate how deep learning can assist in glaucoma diagnosis: Researchers in a 2018 study published in Ophthalmology found that a deep learning algorithm showed a high sensitivity (95.6%) and specificity (92%) to detect referable glaucomatous optic neuropathy using fundus photos.8 Another study, published in 2019 in the American Journal of Ophthalmology, used a deep learning algorithm with pretraining to diagnose glaucoma based on macular optical coherence tomography (OCT) of the retinal nerve fiber layer and ganglion cell complex.9

Cornea and AI

For corneal conditions, AI technology can be especially helpful in the detection and monitoring of keratoconus, Dr. Siddiqui says. 

These factors could benefit therapeutic outcomes, first by catching those who might not otherwise be screened, and second by determining a patient’s progression risk once diagnosed. 

One study, published last year in Eye and Vision, looked at a machine learning-based automated classification system using a Scheimpflug camera data and ultra-high-resolution-OCT imaging data.10 Researchers found that the system was effective at distinguishing subclinical keratoconus eyes from normal eyes.10

“It’s a little bit more challenging, certainly, because we have far more variables and the disease process itself is variable,” Dr. Siddiqui says of AI in keratoconus. “But I think AI can certainly play a huge role. There are research studies that have tried to integrate AI and keratoconus, and have shown that AI may be beneficial in the screening, diagnosis, and management of keratoconus.”

Are There Limitations With AI?

With so much promise surrounding this burgeoning technology, what drawbacks should we be aware of?

Dr. Abràmoff notes what he calls “glamour AI.” “Now that AI has become mainstream, ophthalmologists will need to understand whether or not their patients will benefit” from AI, he explains, instead of being captivated by the thrill of new tools.

He also highlights a second limitation: Potential for liability issues or unethical practices with the data collected. The key part of using AI correctly is having an ethical framework, including “choices such as sensitivity and specificity, how racial, ethnic, and other bias were addressed during design, validation, and implementation,” he says. That’s why it is so important to have peer-reviewed publications on preregistered and hypothesis-testing trials of autonomous AI compared with prognostic standards, he adds.

Dr Lee, warns of a third possible limitation. In his study with colleagues, they found that some AI systems worked better than others, and AI itself is only as good as the data it’s receiving. “All AI is not created equal,” he says.

“You can have AI models that perform really well and you can have AI models that perform really poorly,” he points out, which shows the importance of studies that have head-to-head comparisons of AI systems.

But overall, Dr. Lee and other researchers are actively pursuing study of AI in its various forms, examining how it can benefit patients — and ophthalmologists.

“Compared to other fields of medicine, I would say that ophthalmology is at the bleeding edge, if not the bleeding edge of how AI is transforming clinical care,” Dr Lee says. “We’re in this kind of freefall moment where there was a big innovation, a disruptive innovation that occurred, and we’re still learning how much new is possible, because of this confluence of technology and innovation. So it’s really an amazing time to be an ophthalmologist,” Dr Lee adds.


  1. U.S. Food and Drug Administration. FDA permits marketing of artificial intelligence-based device to detect certain diabetes-related eye problems. FDA. Published online April 11, 2018.
  2. Du XL, Li WB, Hu BJ. Application of artificial intelligence in ophthalmology. Int J Ophthalmol. 2018;11(9):1555-1561. doi:10.18240/ijo.2018.09.21
  3. Gerrits N, Elen B, Craenendonck TV, et al. Age and sex affect deep learning prediction of cardiometabolic risk factors from retinal images. Sci Rep. Published online June 10, 2020. doi:10.1038/s41598-020-65794-4
  4. Benjamens S, Dhunnoo P, Meskó B. The state of artificial intelligence-based FDA-approved medical devices and algorithms: an online database. npj Digit. Med. Published online September 11, 2020. doi:10.1038/s41746-020-00324-0
  5. Lee AY, Yanagihara RT, Lee CS, et al. Multicenter, head-to-head, real-world validation study of seven automated artificial intelligence diabetic retinopathy screening systems. diabetes care. Diabetes Care. Published online January 5, 2021. doi:10.2337/dc20-1877
  6. Siddiqui AA, Ladas JG, Lee JK. Artificial intelligence in cornea, refractive, and cataract surgery. Curr Opin Ophthalmol. 2020;31(4):253-260. doi:10.1097/ICU.0000000000000673 
  7. Introducing Ladas super formula AI. 2015.
  8. Li Z, He Y, Keel S, et al. Efficacy of a deep learning system for detecting glaucomatous optic neuropathy based on color fundus photographs. Ophthalmol. 2018;125(8):1199-1206. doi:10.1016/j.ophtha.2018.01.023
  9. Asaoka R, Murata H, Hirasawa K, et al. Using deep learning and transfer learning to accurately diagnose early-onset glaucoma from macular optical coherence tomography images. Am J Ophthalmol. 2019;198(2):136-145. doi:10.1016/j.ajo.2018.10.007
  10. Shi C, Wang M, Zhu T, et al. Machine learning helps improve diagnostic ability of subclinical keratoconus using Scheimpflug and OCT imaging modalities. Eye and Vis. Published online September 10. 2020. doi:10.1186/s40662-020-00213-3
  11. Abràmoff MD, Tobey D, Char DS. Lessons learned about autonomous ai: finding a safe, efficacious, and ethical path through the development process. Am J Ophthalmol. 2020;214(6):134-142. doi:10.1016/j.ajo.2020.02.022

This article originally appeared on Ophthalmology Advisor