It was very exciting to learn about all of the sophisticated new imaging techniques and diagnostic tools on "Retina Subspeciality Day" during the AAO 2022 meeting. These technologies can provide us with a vast array of data that will help shape our understanding of retinal disease. One of the daunting challenges that we, as retina specialists, will face is determining how we can take advantage of this imaging biomarker data in the context of a busy retinal clinical practice. This is where AI is going to play a critical role. From attention maps to automation, AI should boost efficiency.

A key question that many of our patients have right now is: How fast will my GA grow? Thus, we would like to leverage machine learning by training the system on longitudinal data sets from patients with GA so that we can have a better answer to that question. Predicting the expected progression will be vital in treatment decision making, and it may also help motivate patients and clinicians to persist with the newer agents, which have been shown to delay GA progression by slowing the rate of lesion growth. At AAO 2022, Ursula M. Schmidt-Erfurth, MD, a leading expert in AI, outlined some of the possibilities with AI-guided optical coherence tomography (OCT) imaging in GA. Several groups, including mine and hers, have shown that AI can be used with OCT, fundus autofluorescence, and other imaging modalities to predict GA progression.

The presence of GA means that some visual function loss has already occurred, and, currently, that loss of vision is irreversible. To prevent a patient from developing GA, we would need to intervene earlier in the disease course. As such, there is a need to identify biomarkers that are associated with those earlier phases, particularly biomarkers that can predict whether an eye will progress from early or intermediate dry AMD to GA. AI and machine learning can help us accomplish this goal. The beauty of AI is that it does not presuppose any knowledge of biomarkers. You train it with information that a group of patients progressed with their disease, and it takes that information and deciphers encoded signals that could potentially predict that progression. 

During my talk at AAO 2022, I highlighted a number of OCT-based biomarkers that are linked to an increased risk for GA, including intraretinal hyperreflective foci, hyperreflective cores within drusen, subretinal drusenoid deposits, and a high central drusen volume. Further, we can add to this list features that I did not focus on during my presentation, such as thick basal laminar deposits and photoreceptor loss. All of these are predictors that an eye is going to develop GA.

There are several challenges in applying machine learning and AI to medicine, some of which were discussed during the conference. In the AI world, we use a process called transfer learning in which you essentially give the model a head start by having it learn from something unrelated (eg, training on a diabetic retinopathy data set for a GA application). A challenge in transfer learning relates to the need for labor-intensive annotation, as well as the use of 2-D images to account for 3-D structures. 

During my presentation, I highlighted a technology that was developed at the University of California, Los Angeles called SLIVER-net that allows us to use transfer learning with 2-D images while still preserving the 3-D information in OCT, producing a prediction that is more accurate. We published our findings in npj Digital Medicine, concluding that SLIVER-net significantly outperformed standard deep learning techniques that are used for medical volumes. Our hope is that we will continue to refine this model, adding more biomarkers, and that it will ultimately be something that could be made available to retina specialists. This is all still in the research phase, but it is a promising tool for the future.

Challenges in AI due to data availability and real-world constraints were further magnified in the talk by Aaron Y. Lee, MD, at AAO 2022. There is a problem in AI called overfitting, which means that your model is perfectly tuned to the data set on which it has been trained, but it is very rigid otherwise. As long as everything in the world looks exactly like this data set, it operates perfectly, but it fails spectacularly when it sees things that are outside of those parameters. Dr Lee discussed an innovation in training AI models that allowed them to be a bit more flexible, referring to it as “solving the last mile problem.” He illustrated this by showing how algorithms that were designed to operate on data from one OCT device could be adapted to images from other devices.

Even though the focus in ophthalmology is on imaging biomarkers, we need to remain mindful that the eye does not operate entirely independently. Just like the eye is potentially a window to systemic diseases, systemic biomarkers may be relevant to our understanding of AMD. Joan W. Miller, MD, gave a nice presentation on developing systemic biomarkers for AMD, during which she highlighted work that is being conducted at Massachusetts Eye and Ear in Boston. This group has examined metabolomic and genomic markers that could signal an increased risk for the development and/or progression of AMD. Genomic markers are currently widely used in oncology, and, just as oncologists use circulating tumor DNA to query genomic alterations in tumors, we might, in the future, be able to analyze alterations that are specific to retinal diseases such as AMD and GA.