Retinal imaging is critical in the field of ophthalmology and provides a trove of information that enables physicians to visualize a disease and its pathology, severity, and progress. For years, researchers have been exploring how to take the plethora of images available to gain more insight into eye health and to turn that insight into action with potential to improve patient health. With the integration of artificial intelligence (AI), these efforts are accelerating and generating compelling advancements in the field, from research to diagnosis to treatment.
AI builds on earlier computer vision efforts
Earlier efforts of applying computer vision techniques to analyze images involved creating a hard-coded algorithm containing rules that machines should follow in their search. While this method has worked well in many cases, where it fails is in cases that are outliers or are presented in a manner inconsistent with what models have been told to seek. To capture those types of instances, it would be necessary to revise the algorithm by writing new sets of rules as scenarios arise — an arduous and time-consuming process.
With AI-based approaches, researchers can feed thousands upon thousands of images to a system that will review and sort all the data it comes across, allowing it to detect patterns and to generalize. It is not foolproof, as encountering a not-yet-seen image will require some intervention to confirm whether that image should be added to a database, but this adjustment is relatively simple and allows AI to refine its learning quickly and serve as a more valuable, timely resource with compelling potential for patient care.
Medical AI algorithm offers clarity
Already, one of the first FDA-cleared medical AI algorithms is making a life-changing impact for patients around the world. A collaboration between Doheny Eye Institute and Eyenuk, leveraging NIH-sponsored grants, has resulted in a new screening system to detect diabetic retinopathy, a condition that can result in vision loss and blindness in people with diabetes. Screening is vitally important because 90% of diabetic blindness is preventable with timely detection and treatment.
For the AI screening tool, experts in image reading participated in algorithm development and training by furnishing annotated images of cases ranging from mild, non-proliferative retinopathy to the most severe stage of proliferative retinopathy. With the data input, the algorithm was trained to identify whether patients exhibited retinopathy that warranted a visit to an ophthalmologist to avoid future vision loss.
In the case of diabetic retinopathy, early stages of the condition may go unnoticed, or if the symptoms appear manageable, patients may postpone a visit to their physician. An AI screening tool offers the possibility of garnering greater compliance with recommendations for regular assessment. Capturing an image of the eye at a primary care physician’s office or even on a smartphone at home and sharing it with AI in the cloud for an immediate diagnosis could promote early detection and, hopefully, prevent blindness.
Similar work is being conducted in the area of age-related macular degeneration (AMD). Algorithms are being developed to detect features in patients with AMD to help physicians identify those who may be at high risk of progression to more advanced stages. The information can assist doctors in closer monitoring of at-risk individuals and in developing a treatment course.
Adding another tool to the toolbox
The advent of AI has ushered in countless conversations about the technology’s impact — good and bad — and the medical field has certainly not been immune to the debates. For one, AI is not perfect, and it can and will make mistakes. It’s necessary to consider the output of large language models that have generated hearsay or misinformation, and human experts must approach any output with a healthy dose of skepticism and be unafraid to challenge results.
Additionally, the application of any new technology, like AI, may result in unintended consequences. Should medical AI models become widely available, there is a risk that those without proper credentials may try to use them to make decisions, or to question decisions, which are best left to licensed medical professionals. AI models can provide actionable knowledge, but they will lack the broader context of a patient’s circumstances, values and conditions that their physicians should understand and consider in a holistic treatment plan.
If harnessed with care, AI has the ability to provide medical professionals with another tool to elevate patient care. Each case is unique and deserves a personalized approach, and AI may help doctors optimize their selection of treatment options. Physicians could leverage information furnished by AI, such as from clinical trials, and compare it alongside an individual’s characteristics to arrive at a care approach with the best chance for success.
AI is poised to be instrumental in the decision-making process for physicians, but ultimately, it cannot replace the human element. Medicine is much more than determining a diagnosis, and tough discussions and difficult decisions require empathy that only another human being can provide. With AI as an assistive tool, it opens the door for physicians to access more information and context to aid in making treatment recommendations and ultimately help improve their patients’ outcomes.