AI in ophthalmology: From code to clinic

Applications of AI

April 06, 2023

10 min read


Healio Interviews

Habash reports having disclosures for Allergan/AbbVie, Avellino, Bausch + Lomb, Johnson & Johnson and Zeiss. Keane reports consulting for DeepMind and Google. Lad reports serving on the steering committee for Apellis and having a provisional patent 63162741: “A system and method to predict progression of age-related macular degeneration.” Lim reports consulting for Eyenuk.

We were unable to process your request. Please try again later. If you continue to have this issue please contact

“With the advent of advanced machine learning algorithms and powerful computing capabilities, artificial intelligence has the potential to revolutionize the diagnosis and treatment of various eye disorders.”

That introduction was not written by a human but by the AI program known as ChatGPT, a system developed by OpenAI. It is one of the newest advances in AI and has become the topic of international interest and debate. All we had to do was ask the program to write an introduction to a story about AI in ophthalmology, and it did the rest.

Eleanora Lad
There is a significant need for integration of artificial intelligence across ophthalmology, particularly for age-related macular degeneration, according to Eleonora M. Lad, MD, PhD.

Source: Duke Eye Center

Few things sound more futuristic than artificial intelligence, but AI has been around for a long time.

“It’s a term that’s been around for more than 50 years,” Pearse Keane, MD, FRCOphth, said. “But what essentially happened in the last 10 to 15 years is that a subtype of AI called deep learning got very, very powerful, and tech companies, like Amazon and Apple, became AI-first companies.”

Research by these companies has brought forth a lot of technology that people take for granted every day, from language translation and image recognition to digital personal assistants such as Siri and Alexa.

Pearse Keane, MD, FRCOphth
Pearse Keane

“People should understand that we use AI every day,” Ranya Habash, MD, said. “When you’re shopping on Amazon for a red sweater, you don’t have to search through every website on the internet to find all the red sweaters. The algorithm aggregates them for you, but you make the decision about what you want to buy. Ultimately, it’s the same thing in medicine.”

Habash said AI systems are already in use in ophthalmology, helping with data aggregation, compiling information from multiple sources and presenting the data to the physician to inform decision-making.

“We have algorithms for intraocular lens calculations and autonomous systems for diabetic retinopathy screening that can look at a picture of the eye and cross-reference all of its characteristics — the size of the blood vessels, the hemorrhages — and give us insight about the findings,” she said. “We still make the ultimate decision about how to treat the patient, but AI can guide our decision-making. I use the term ‘augmented intelligence’ rather than artificial intelligence.”

While AI is already having an impact in the eye care space, recent advances could drastically affect the world of medicine in ways that are not even currently understood.

Foundation models

Keane said that since the deep learning breakthrough, AI technology has been accelerating at an exponential rate. Outside of health care, the advent of new models, known as foundation models, has ramped up that progress even more, including with applications such as ChatGPT.

“They are these large language models trained on every bit of information available on the open internet,” Keane said. “You get huge amounts of data and train the AI system in an unsupervised fashion. It can have billions or potentially trillions of parameters. It turns out that when you have loads of data and really powerful models, unexpected things start to happen.”

Keane said this has immense applications in health care and ophthalmology. Just like ChatGPT is trained in language and information, he said a similar program can be trained on millions of ophthalmic images and used for any number of conditions or diseases. However, the foundation is the entirety of the data, not just one disease.

“You might want to tweak the model to be able to diagnose glaucoma or diabetic retinopathy,” he said. “Since it’s been trained on so much data, it does much better than conventional models that have been trained on diabetic retinopathy alone.”

Ranya Habash, MD
Ranya Habash

Habash said having an advanced chat program itself could be a useful tool for an ophthalmologist.

“If we’re dictating notes during a visit, ChatGPT can use natural language processing to extract relevant information from the EHR and structure things into a SOAP note or referral letter, for example,” she said. “Some doctors are even using it now to refute denials in insurance cases or for prior authorization appeals. ChatGPT will formulate the letter and even add citations based on its foundation of medical literature.”

Precision medicine

Jennifer I. Lim, MD, said AI could be a powerful tool in advancing precision medicine. Its ability to rapidly quantitate changes in values between ophthalmic imaging results and other test values could help physicians compare information between one patient visit and another, allowing them to make well-informed treatment choices.

“Quantitative assessments that would otherwise be tedious for somebody to perform in a clinic setting could be done more precisely and automatically. Calculations that otherwise would not get done or would only get done in the context of a clinical trial could become routine,” she said. “AI could give us that precision for an individual patient that we’re missing.”

Jennifer I. Lim, MD
Jennifer I. Lim

In glaucoma, Habash said there is a great opportunity for using predictive analytics to make more accurate assessments on progression and treatment decisions. AI could provide a new analysis to make personalized improvements to care.

“We diagnose a patient with glaucoma and use our normal protocols — one drop, two drops, maybe a laser,” she said. “But glaucoma is just so complex and multifactorial. We could use AI to aggregate all of the different data sets, apply analytics to our thought process and get a lot more precise about how that patient might do on certain treatments.”

Programs can provide an individualized risk score or give insights on a particular patient’s chance of progressing.

“There are very objective, tangible things that we can derive with the help of data analytics through artificial intelligence that we wouldn’t have been able to objectify before,” Habash said. “We use factors like age, race and family history to try to calculate risk, and we’ve had a best guess as to how we should treat our patients or where their disease might end up. However, we don’t know anything definitively. With AI, we can compare a patient’s characteristics against what is in the literature from real-world data. From there, based on all of the factors that are common to all of these patients, an algorithm can generate a score for the risk of progression.”

The level beyond this, Habash said, is known as digital twinning, in which AI creates a digital copy of a patient based on clinical factors with lifestyle factors as well.

“The digital twin has all of the same characteristics that make you who you are, so we can compare it against real-world data and run simulations through AI and data analytics,” she said. “Say we’re following a patient for glaucoma, and now we’re able to fast forward the clock through AI simulations and see what the glaucoma will look like in 10 to 15 years. We can also run simulations using different kinds of treatments. How would it look under different conditions? It’s kind of like looking into a crystal ball, seeing your patient’s future and then modifying it in the present. This is precision medicine, where we can gain a deeper understanding of our patients’ conditions, predict potential outcomes and customize treatment plans accordingly.”


The area in which AI has already given eye care a massive boost is in screening. In 2018, the IDx-DR (Digital Diagnostics) became the first AI device granted FDA clearance for diabetic retinopathy screening. This was followed by the EyeArt system (Eyenuk), the first FDA-cleared system to screen for vision-threatening diabetic retinopathy in addition to more than mild diabetic retinopathy in 2020.

Lim said the IDx-DR and the EyeArt are perfect examples of how AI can bring screening to more people more efficiently.

“It’s so impactful because these AI autonomous systems can help us screen patients who wouldn’t otherwise get screened,” she said. “There are patients who don’t get an annual diabetic examination, but they do frequently visit a primary care office where devices like this can be strategically placed.”

Devices such as the IDx-DR and EyeArt allow even non-ophthalmic photographers to capture images and upload them to the cloud to be read in a matter of minutes.

“The patient is screened with high enough sensitivity and specificity to make it useful to detect the presence or absence of more than mild diabetic retinopathy,” Lim said. “That means more diabetic retinopathy patients are getting diagnosed.”

While diabetic retinopathy has paved the way for AI in eye care, Eleonora M. Lad, MD, PhD, said there is a significant need for AI help in other areas of ophthalmology, particularly in age-related macular degeneration.

“We’re really interested in its potential to make a positive impact on diagnosis and management of patients with AMD,” she said. “However, the development of effective AI devices for clinical care faces numerous considerations and challenges, evidenced by the complete current absence of FDA-approved devices for AMD, unlike in the diabetic retinopathy space.”

However, Lad believes the recent FDA approval of Syfovre (pegcetacoplan injection, Apellis Pharmaceuticals) for geographic atrophy (GA) could lead to a new “therapeutic era,” further heightening the need for efficient and effective screening methods.

“Our challenge is that a lot of patients with AMD live in the comprehensive and optometric communities, and they don’t present for retina care,” she said. “Unless they have wet AMD, there has previously been no reason to refer them for retina care given the absence of treatments for the more common dry form of AMD.”

Lad said there is hope that Syfovre and other therapies in the pipeline will help prevent dry AMD progression.

At Retina 2023, Lad presented an algorithm she developed with colleagues at Duke University that may help predict progression of intermediate AMD to GA. Lad said the model had excellent performance characteristics for detecting GA and predicting progression in the short term. The algorithm was also able to automatically identify the specific structural features on OCT that were most associated with or predictive of GA.

“It’s called the multiscan position-aware volumetric image classification model,” she said. “We trained the model on two data sets, and in the end, the performance was excellent. It was predicting GA in the current year with an AUC of 0.945. The following year, it was 0.937.”

Lad said the model made these predictions without human grading. When human grading was added in later, the advantage was only minimal, of 0.008 in AUC.

“We were surprised that all of the laborious human gradings on OCTs did not add that much,” she said. “The algorithm is very good independent of human expert contribution.”

Lad said the algorithm actually thinks a lot like human clinicians, paying close attention to the parafoveal area where GA starts.

“A lot of meaningful work remains to be done,” she said. “But there are so many avenues of investigation.”

From code to clinic

Keane said developing AI in health care comes in two phases. The first is “idea to algorithm.” In this phase, developers come up with an idea, gather data and train a model with the goal of publishing a paper.

“Everybody gets excited about it, but the other phase, which I think is longer and harder in some ways, is going from code to clinic,” Keane said. “Once you publish a paper, you need to show that it properly works. You need to do trials that will satisfy the FDA or other regulators that can often be expensive and require a lot of expertise. It’s a massive undertaking.”

Even after trials are done, a health care AI system still needs regulatory approval, a business model and other infrastructure before it can be integrated into clinical practice.

“At the dawn of the electrical age, Thomas Edison understood that not only did the light bulb need to work, it also had to have a network of innovations around it, like a proper generator, a distribution system and a way to allow for payment,” Keane said. “For us in the AI world, we’re still at the prototype light bulb stage and not the Edison-level integrated system stage.”

To better understand the road map for the integration of AI in ophthalmology, collaborators from the FDA, the National Eye Institute and other organizations formed the Collaborative Community on Ophthalmic Imaging (CCOI). Lim was part of the team that specifically looked at what was needed to bring an AI system to the clinic in AMD care.

“We tried to look at all the information and determine exactly what infrastructure we needed to advance the field,” she said.

While large data sets of color fundus photography and OCT already exist, the CCOI found that the image data often did not include all of the information necessary to create reliable AI models. Sharing data was also made more difficult due to privacy regulations.

“This concept of federated learning will allow us to share developments and analysis of the images within the algorithm without having to move images with HIPAA-protected information on them,” she said. “We realized that we needed an infrastructure that allows us to work across multiple institutions. If everyone is working in their silo creating all different algorithms that can’t talk to each other, then it’s not going to push the field ahead.”

According to the CCOI’s AMD road map, institutions could contribute to a secure cloud-based server accessed only through a virtual private cloud. The algorithm could transfer its analysis back without distributing any of its raw data. Another option is to have the AI model trained on local databases, with the model itself transferred but not the data.

More innovation is needed to reach a wider population and more diseases, but Lim said AI can already provide a massive benefit to patient care.

“In your day-to-day practice, having an AI assistive algorithm available would actually help you interpret findings faster and help inform your clinical decision-making,” she said. “It decreases the screening burden and gives you time to concentrate on taking care of patients who actually need you to do something for them. It’s going to shift what we’re doing from screening and diagnosing to treating and ultimately preventing vision loss.”

Keane said he expects massive advancements in health care AI in the next 5 years or so, with these new technologies becoming just another tool for physicians.

“If we’ve learned anything, it’s that these technologies seem magical at first, but then about a half hour later, you forget that they’re anything different,” he said. “If you think about our phones, we can look up any piece of information in the world within seconds, and it’s just natural to us. That would have been mind blowing 20 years ago.”

Click here to read the Point/Counter to this Cover Story.

Source link

Leave a Reply

Your email address will not be published. Required fields are marked *