How is AI used in ophthalmology?
Broadly, AI has made great strides in computer vision and natural language processing tasks. Eye care is a specialty that relies heavily on imaging, which makes ophthalmology ripe for the implementation of AI. In fact, one of the very first large-scale studies demonstrating the potential for AI in image analysis was in ophthalmology in 2016. Since then, we have seen a continued push to applying AI to many clinical problems.
What are some examples of applications?
Diagnosing eye diseases is a common example of the kind of AI applications being developed in ophthalmology. Many eye diseases are diagnosed by examining different types of images of the eye, and computers are getting quite good at such tasks. As a specific example, in the last few years, the Food and Drug Administration approved two AI-powered screening tools for detecting diabetic retinopathy. These applications could enable generalists using AI to take retinal images and refer patients requiring care to an ophthalmologist for further screening. This as an important step in standardizing treatment for diabetics and increasing access to screening.
Machines have also been trained to review retinal images to screen for retinopathy of prematurity (ROP), a potentially blinding disease affecting premature infants. Working closely with ophthalmologists at Oregon Health & Science University and other institutions, we have developed a diagnostic tool, a measure of disease severity, and a risk model that could improve current protocols. Current ROP screening requires weekly or biweekly in-person examinations of infants born before 31 weeks of gestation, creating a burden for their families. If we could identify all babies at high risk for treatment requiring (TR) disease, we might be able to reduce overtreatment and improve the quality and consistency of care in infants at risk of ROP.
AI might be also used "behind the scenes" for automated measurement of disease features in ophthalmic images. Quantification of attributes that historically were more qualitative might be useful for research and potentially to uncover biomarkers of disease in the future. AI might also aid in diagnosis by providing clinicians with objective measures of disease severity and improve diagnostic consistency between clinicians.
What are some challenges facing the implementation of AI in ophthalmology?
The applications are only as good as the data they are drawn from, and there’s still work to be done to collect data from diverse populations and build algorithms. A 2020 study showed that that geographical distribution of cohorts used in studies to develop AI algorithms in the U.S. was quite limited and predominantly came from just three states – New York, Massachusetts, and California. There may also be racial bias in the collection of patient information, with white patients making up a disproportionate percentage of data.
What are some opportunities for AI implementation in ophthalmology?
In addition to using AI for ophthalmic diagnoses including ROP and diabetic retinopathy, we see potential opportunity in diagnosis and monitoring the effectiveness of treatment in many diseases including age-related macular degeneration (AMD) and glaucoma. By accurately ruling out patients who do not need treatments, AI could help clinicians more effectively target patients who do need help.
We see opportunities for screening or triaging, especially in underserved communities. If we can identify those patients who are more likely to require care, it really does relieve a lot of the burden on the health care system.
The use of AI in workflow and operations in a clinic setting represents another potential area for AI to improve access to care. AI tools may be used to optimize efficiency in the clinic by tailoring appointment times to patient needs. They may help with identifying patient no-shows and suggest mitigating strategies such as help needed to get to the appointment.
What progress has already been made? What are the remaining obstacles?
There are many algorithms that have been published in the literature but the full potential of AI to transform health care has yet to be realized. The “final mile” of implementing AI in the clinic is still a challenge for most institutions. Even if the infrastructure for deployment might not be fully developed yet, AI is already moving forward. The open-source nature of the computer science community and the advances in computer technology has allowed the field of medical AI to build on successes from other domains. In addition, large amounts of data being collected from electronic health records and the expansion of databases like those supported by the National Institutes of Health are paving the way for more AI-powered tools. The widespread adoption of devices like smartwatches that collect health information from individuals is another opportunity to explore. Overcoming resistance to change may be the first hurdle to clear. Ensuring that our AI algorithms are safe and fair and do not cause harm to vulnerable populations is key to broad adoption.
What’s next for ophthalmic AI at the Sue Anschutz-Rodgers Eye Center?
We are working closely with our clinician colleagues to explore building AI algorithms for the diagnoses of many diseases including AMD and glaucoma while continuing to expand on our work in ROP. In addition to diagnosis for specific diseases, we are looking to signals in the eye as a measure of overall patient health.
Our true measure of success will be determined by the impact our work has on the health of the patients we serve at the Rocky Mountain Lions Eye Institute. Ultimately, instead of thinking of AI as replacing doctors, we need to think about what humans are good at versus what computers are good at to ultimately improve the "art and science" of medicine.
.png)
