AI in Ophthalmology: Clinical Practice Integration

Published Date : 2/10/2025 

The integration of artificial intelligence (AI) in ophthalmology is transforming the way eye care professionals diagnose and treat patients. Despite challenges, AI is poised to revolutionize the field, improving accuracy and efficiency in clinical settings. 

With the rise of artificial intelligence, the healthcare field is among many transforming industries that are incorporating AI into their practices. Whether it be through AI scribes taking notes during doctor appointments or AI assessing radiology images for cancer, the impact is significant. For many ophthalmologists, the boom in AI has raised a critical question: How should ophthalmologists bring AI into their practices?

It’s an evolving, multifaceted, and complex issue that ophthalmologists and researchers discussed at the 29th Annual Sue Anschutz-Rodgers Eye Center Symposium, a conference that brings together eye care professionals and trainees from across the country to the University of Colorado Anschutz.

“What frontier AI technologies such as those developed by OpenAI or Google can do is amazing, but very few of them have made their way into the clinic,” said Jayashree Kalpathy-Cramer, PhD, professor of ophthalmology at the CU Anschutz School of Medicine and chief of the Division of Artificial Medical Intelligence in Ophthalmology. “What we actually see routinely used in the clinic is light years away from the technology that is already in existence and usable, because there are many challenges to safely deploying AI in the clinic.”

The Data Challenges of AI

There are several reasons why AI is not currently being used as much in healthcare as it is in other fields, explained Niranjan Manoharan, MD, an associate professor of ophthalmology. “Data is really the fuel for AI, and this is where AI mostly gets stuck,” he said. “We need high-quality data that is large, diverse, accurately labeled, and reflects the real-world population. This is where a lot of AI stalls, especially in healthcare where our data quality is lacking.”

Training AI based on clinical records and images is difficult, he explained, because diagnoses that are pulled from the electronic health record often do not match what is actually visible in a routine clinical image. For example, there are times when a photo taken in a medical office does not fully capture the issue — such as a retinal tear in the eye — which can lead to the AI not properly learning how to scan for these issues and giving false positives.

“Unlike bank transactions, health care data is messy and inconsistent, which undermines large-scale training,” he said. Due to these challenges, there is still a reliance on humans to manually annotate the data to properly train AI systems — a process that Manoharan described as accurate but slow and difficult to scale.

Capturing enough data that is reflective of different ethnicities is also challenging, he explained. For instance, if an AI model that reviews images taken by ophthalmic fundus photography cameras is trained with mostly images of darker fundus pigmentation, this may cause the AI model to struggle to properly identify images with lighter pigmentation.

Manoharan further notes that the high volume of patient aggregate data needed for cross-institutional AI modeling is hindered by privacy and data-sharing constraints. However, despite these challenges, he is hopeful for the future.

“Our ophthalmic AI team, along with other institutions, are working on addressing these data challenges one by one,” he said. “I believe we’ll soon see AI and automation everywhere in healthcare and ophthalmology. The challenge will be to sift through what makes it better for our patients.”

Developing AI Tools

There is a difference between a health information technology tool and software as a medical device, Manoharan explained. “A health IT tool is something that documents, summarizes, and organizes information for the clinician,” he said, adding that these tools do not have to be regulated by the U.S. Food and Drug Administration. “Software as a medical device does have to go through the FDA. This is software that diagnoses, screens, triages, and drives clinical action.”

In ophthalmology, he noted, there are three different autonomous screening softwares — IDx-DR, LumineticsCore, and EyeArt — that are AI systems built to detect diabetic retinopathy. These were all approved by the FDA. “Less than 50% of diabetic patients in the United States get appropriately screened every year. We have three tools that really work,” he said. “But many people still prefer human graders.”

In addition to diabetic retinopathy, AI is also being developed and applied to help with other eye conditions. For example, retina specialist Emily Cole, MD, MPH, assistant professor of ophthalmology, discussed how AI could potentially be used in neonatal intensive care units to help screen infants for retinopathy of prematurity (ROP).

Talisa Forest (de Carlo), MD, medical director of imaging and assistant professor of ophthalmology, spoke about the use of AI in the Sue Anschutz-Rodgers Eye Center’s age-related macular degeneration (AMD) registry. This registry holds repositories of electronic data, imaging data, and blood samples for biomarker studies that allow researchers to further investigate AMD. The ophthalmology department’s AI team has developed tools to help visualize and characterize the growth of geographic atrophy, an advanced form of AMD.

Kalpathy-Cramer explained how using AI for image registration — a process of capturing and aligning multiple images together — can be helpful for ophthalmologists treating conditions like uveitis, allowing the ophthalmologists to see small changes in the eyes between various images. “If you can do this across diseases, across modalities, and across time, I think it really can help clinicians see things better,” Kalpathy-Cramer said.

Building Strong Datasets

The 27th Annual Phillip Ellis Lecture was given by Aaron Y. Lee, MD, MSCI, professor and chair of ophthalmology at the Washington University School of Medicine. He described how AI, deep learning, and large-scale model developments can impact healthcare. Lee has conducted various research projects centered around the use of AI, ultimately looking for ways that the field of ophthalmology can apply these tools to advance care for patients.

One of his major research projects, called the Artificial Intelligence Ready and Exploratory Atlas for Diabetes Insights (AI-READI), is an initiative supported by the National Institutes of Health to develop a large, ethically sourced dataset for type 2 diabetes. Lee and his team are working to collect health information and measurements from about 4,000 people with different stages of diabetes.

“The really cool part of the project is we actually sent people home with three different devices. One is a continuous glucose monitoring device, the second is a fitness tracker, and then a third is a custom-built environmental sensor,” he said. “Last November, we released our year two dataset that had over 1,000 participants in it, and it has millions and millions of measurements.”

The dataset can be downloaded by other research teams, allowing for more investigators to access and use the information in their own studies. “This is really the proof in the pudding, from the NIH perspective, whether this endeavor is having an impact on the world,” Lee said. “I think AI is here to stay, and I think it really is a transformative technology.” 

Frequently Asked Questions (FAQS): 

Q: What is the main challenge in using AI in ophthalmology?

A: The main challenge in using AI in ophthalmology is the quality and diversity of data. High-quality, diverse, and accurately labeled data is essential for training AI models effectively.

Q: What are some AI tools used in ophthalmology?

A: Some AI tools used in ophthalmology include IDx-DR, LumineticsCore, and EyeArt, which are approved by the FDA to detect diabetic retinopathy. AI is also being developed to help screen infants for retinopathy of prematurity (ROP) and to manage age-related macular degeneration (AMD).

Q: How can AI improve image registration in ophthalmology?

A: AI can improve image registration by aligning multiple images together, allowing ophthalmologists to see small changes in the eyes over time, which is particularly useful for conditions like uveitis.

Q: What is the AI-READI project?

A: The AI-READI project is an initiative supported by the National Institutes of Health to develop a large, ethically sourced dataset for type 2 diabetes. It collects health information and measurements from about 4,000 people with different stages of diabetes.

Q: What is the role of human graders in AI-assisted ophthalmology?

A: Human graders play a crucial role in manually annotating data to properly train AI systems, ensuring the accuracy and reliability of AI tools in clinical settings. 

More Related Topics :