August 2020


Artificial Intelligence
Artificial intelligence in retina

by Ellen Stodola Editorial Co-director

AI-based identification and quantification of intra- and subretinal fluid on home OCT images of a patient with neovascular AMD (Notal Vision)
Source: Notal Vision


Several experts discussed applications of artificial intelligence (AI) in the retina subspecialty.
Anthony Joseph, MD, thinks that the diseases it’s initially going to be most useful for are the ones that make up the majority of the retina practice: macular degeneration, diabetic macular edema, diabetic retinopathy, and vein occlusion.
Dr. Joseph thinks AI could play a role in screening and monitoring diseases for those who aren’t to the point of requiring active treatment.
It may also help in patients being treated or followed more closely for macular degeneration. “We hope it might be something to help predict when they may convert and develop wet macular degeneration or something that can help measure treatment response or treatment needs,” he said.
Allen Ho, MD, thinks AI will be useful for retinal diseases that have rich datasets available to analyze. AI analyzes and discerns information that physicians might not typically see in the dataset; for example, there are many structure and function correlations in the OCT imaging datasets from a variety of common retinal diseases. Color fundus imaging is another rich dataset from which disease prognosis and treatment responsiveness may be refined.
Glenn Stoller, MD, thinks the two most obvious applications for AI in retina are for diabetes and macular degeneration. It could also be used for retinopathy of prematurity. AI algorithms have already been shown to be effective at detecting clinically significant macular edema, as well as advanced stages of diabetic retinopathy. It can track disease progression by comparing current images to those that were initially screened, helping to provide insight into the progression of diabetic retinopathy, Dr. Stoller said.
He added that there is already a system on the market (IDx-DR, IDx Technologies) that uses deep learning and is able to screen outside the ophthalmologist’s office, helping determine when clinical referral is indicated.
In using retinal OCT images, AI systems can be trained to perform a segmentation, Dr. Stoller said. The system has been shown to display a high degree of accuracy in segmenting different layers of the retina.
AI may also enable personalized healthcare. He mentioned a home-based OCT device from Notal Vision that uses a machine learning algorithm. This machine is intended to help monitor patients from home and determine when someone who has wet macular degeneration will need an injection. The technology is expected to be commercially available in the first half of next year. An AI-based random forest classifier is already being applied to analyze visual field data of the ForeseeHome device (Notal Vision), used to determine when a patient converts from dry to wet macular degeneration.
Jennifer Lim, MD, said that AI could help decrease the screening burden of diabetic retinopathy. Many diabetic patients never undergo screening (up to 70%, depending on the population and location studied). “It would help to identify patients at risk of visual loss and hopefully result in referrals for appropriate treatment,” she said.

AI facilitating interaction

Dr. Lim sees the potential for providers across medical specialties to work together to reinforce the importance of diabetic glucose control, as well as control blood pressure, cholesterol, and lipid levels. All physicians could see the results of the patient in the outputs. The data would be presented to the patient by whomever is seeing the patient next, and all team members could be easily updated, she said.
In terms of interaction and collaboration with AI among primary care, endocrinologists, ophthalmologists, and retina specialists, Dr. Joseph said the primary care doctor or endocrinologist may be interacting with patients more frequently at the point-of-care level, especially for those with diabetes. This could be a point where data is gathered for screening, as opposed to patients going into ophthalmologists’ offices for screening exams.
Specifically relating to diabetic retinopathy, Dr. Ho said it’s important to improve communication among all members of the care team. In large-scale disease screening, the interaction would hopefully become more refined, providing risk assessments to the general care team from specialists.

Cost effectiveness

Dr. Ho said AI could create value by identifying patients who are more responsive to medication, those who may not respond to certain treatments, those at risk for vision loss, and patients more likely to be lost to follow-up. Currently AI algorithms are employed in home monitoring solutions for the early detection of neovascular AMD and in the near future with home diagnostic imaging that may refine treatment of macular diseases with increased value to the patient and the healthcare system in general. The cost-efficient interpretation of daily home OCT images for the identification and quantification of intra- and subretinal fluid will require AI assistance.
Dr. Stoller added that AI may be a cost-effective way to improve patients’ access to care because remote visits/teleophthalmology help reach patients who may not be located in areas with easy access to care. “I think screening initiatives combined with AI technologies may help to diagnose eye disease that might otherwise get missed,” he said. It could also require fewer resources to operate, as opposed to some of the more time- and labor-intensive tests, Dr. Stoller said.
Dr. Joseph said that if AI reaches the point where it’s able to predict disease progression and more tailored therapy, physicians could hopefully use resources more efficiently, instead of relying on a cookie-cutter approach. “If we’re looking at earlier disease detection and more accurate detection, hopefully we’re treating patients before they develop a serious disease burden,” he said.

Risks and challenges

Dr. Ho said that one possible risk is that you can get misled by large datasets. Just because there’s an association doesn’t mean there’s causation, he said. AI may give you hypotheses but not definite causation or optimal treatment. Expert, experienced human intelligence will still be an essential component of the interpretation and implementation of AI analyses. Another risk is trying to obtain more information than what the dataset can yield, he added.
“I think this is something that has been on the radar for awhile, but it takes time,” Dr. Joseph said, adding that he expects a number of regulatory and ethical concerns for patient privacy and data protection. AI requires large amounts of patient data, he added, so there is the question of who owns the data. There are also technical limitations in terms of computational power that’s required.
“We’re producing or obtaining a large amount of clinical imaging, and that’s what we want to use, but I think processing it takes manpower now and computational power later,” Dr. Joseph said. “I think if you’re looking further down the road, once we develop algorithms or tools, what kind of human oversights are we going to have?” He thinks some degree of human oversight will be needed. Adding this new tool won’t happen overnight, he said. “I think people think that this is not something that will be a replacement for clinical judgment, but it will certainly be an important tool in how we manage disease,” Dr. Joseph said.
Dr. Stoller said that one major challenge for AI will be acceptance from patients and doctors. “Many patients are willing to embrace high-tech devices, but I think they still might not trust computer diagnosis and would rather have their diagnosis made by an in-person visit,” he said. With regard to physician acceptance, it’s not always obvious how the computer algorithm reaches its conclusion, so the physicians are forced to trust the AI system without being able to evaluate the value of the metrics and data used by the computer program.
Another challenge is managing the technology itself. It’s important to assess the quality and validation of the datasets to ensure the results are applicable across diverse populations. “The dynamic nature of an AI system is that machine learning makes it difficult to predict at what point the algorithm should be reviewed,” he said. There is also a risk for false negatives.

Impact on clinical trials

“I think we’re going through a learning phase about how AI can be used in clinical trials,” Dr. Stoller said. He thinks AI could help identify patients most likely to develop disease or progress and could impact dosing regimens used in clinical trials. “I think there’s a great deal of promise,” he said. “It’s not something that’s been widely adopted yet in the ophthalmic community, but I think we’re in the early phases of seeing that paradigm shift and embracing the technology.”
AI could pinpoint groups at higher risk of progression, identifying subgroups that would be best suited for preventive therapies, Dr. Lim said. In addition, AI could analyze the images for treatment response instead of relying on human graders. Lastly, AI may find patterns of disease that respond better to a treatment, she added.
Dr. Joseph also thinks that AI could have an impact in clinical trials. “I think clinical trials are more rigorous in their controls and designs than regular clinical practice,” he said. “I think it’s a good opportunity for gathering and curating data and clinical images that would be useful in developing machine learning tools and accurate algorithms.” Dr. Joseph added that he thinks researchers will realize when designing clinical trials that they may want to capture data in a fashion that will lend itself well to developing AI tools, in addition to testing whatever medications or interventions that they’re trying to study.

At a glance

• Physicians said AI may be most applicable to diabetes and macular degeneration in the retina subspecialty.
• There may also be opportunity for enhanced collaboration among ophthalmologists, primary care providers, and other specialists when using AI and monitoring patients.
• There are some potential concerns, including acceptance by patients and doctors, the potential of being misled by large amounts of data, and the extent to which manpower will still be needed.

About the doctors

Allen Ho, MD
Director, Retina Research
Wills Eye Hospital
Philadelphia, Pennsylvania

Anthony Joseph, MD
Ophthalmic Consultants of Boston
Boston, Massachusetts

Jennifer Lim, MD
Director, Retina Service
University of Illinois at Chicago
Chicago, Illinois

Glenn Stoller, MD
Ophthalmic Consultants
of Long Island
Rockville Centre, New York

Relevant disclosures

: None
Joseph: None
Lim: None
Stoller: None



Artificial intelligence in retina Artificial intelligence in retina
Ophthalmology News - EyeWorld Magazine
283 110
220 265
True, 8