Trends of artificial intelligence and machine learning technology and retinal imaging

Nitish Mehta, MD discusses the latest trends in Artificial Intelligence and Machine Learning technology in retinal imaging and its uses for ophthalmologists with David Hutton, Managing Editor, Ophthalmology Times®.

Editor’s note: This transcript has been edited for clarity.

David Hutton:

Hello, I'm David Hutton of Ophthalmology times. I'm joined today by doctor Nitish Mehta to discuss some of the latest trends in Artificial Intelligence, Machine Learning technology in retinal imaging.

Thank you so much for joining us today to discuss some of these cutting-edge advancements. Artificial Intelligence is gaining favor in ophthalmology. Tell us about some of the trends of AI/ML technology and retinal imaging.

Nitish Mehta, MD:

Great, thank you so much for having me for this interview. It's very exciting to talk about these advancements in our fields.

I think it'd be best to start by backing up a little bit and understanding what AI is. And basically, the conjecture, is that any aspect of learning or a feature of intelligence can be so precisely described that a machine can be made to simulate it. And in terms of medical AI, you're designing a system that's going to learn patterns from clinically annotated datasets, and this can be clinical data or imaging data, and we're going to use these insights to assist in clinical practice. And within that, there are a variety of flavors per se of AI. And the one that's most relevant for images and retinal imaging, or in retina in particular, is what we call Machine Learning, which is essentially a family of statistical methods that can do more than what we normally do with our simple statistical models, say for example, a regression.

And particularly deep learning, which is a method to look very closely at images, has shown a lot of promise within retina in particular. So now you have an understanding basically, on what are the particular tools that we use, which essentially are clusters or sort of groupings of different statistical methods to allow us to gain more insight and prediction.

The next question is essentially going to be what are you going to be able to do with these tools? And sort of the major categories are, can you have a machine categorize images similar to the way that a retinal physician would?

So you know, we look at a retinal image, and we decide if this patient has no disease, diabetic retinopathy, macular degeneration. Can these algorithms do this on their own? And that is the first major category of AI in retina, which is screening. And screening, you're going to find a way that machine can automatically detect images. So that's our first major category in trends within ophthalmology, particularly in retina.

And secondly, as an extension of the first is diagnostic grading, can you take an image and determine how advanced a particular condition is. Can you use an algorithm to aid in the creating of diagnosis? Or can you use AI to create a whole new paradigm of grading? So for example, now you have an image, and you can determine if they're mild, moderate, severe diabetic retinopathy.

Or the next step would be to say, here are 1000 patients with a variety of different flavors of diabetic retinopathy, how would you design the grading to split them out based on clinical outcomes.

So there's kind of 2 different ways to think about this, could we use AI to sort of teach us more in a different way than we're thinking currently. And sort of the last and maybe the Holy Grail of AI and retinal imaging is guidance of therapy.

For example, we've all internalized the teaching from the DRCR NET that patients with 20/50 vision or worse with DME may have better outcomes with treatment of Aflibercept versus Bevacizumab. So can an AI system identify the patient as having DME, having the vision worse than 20/50 and then recommend a particular treatment to the provider.

And this can be anything for you know disease activity or disease level with home OCT forthcoming, an AI algorithm that could basically alert a condition of disease activity without manual oversight could allow for seamless and timely treatment with the patient who has a conversion from dry to wet AMD. So that was sort of the major overarching goals here; Screening, grading and treatment guidance.

David Hutton:

You've kind of hinted at some of the applications of AI in retina - Are there any other applications or ways that this technology can help retinal surgeons in their practice?

Nitish Mehta:

First, in terms of actual current application, screening is the furthest along. So there are 2 FDA approved now devices, 1 from IDx-DR and the other one from EyeArt, which can in a primary care setting, identify referral-warranted diabetic retinopathy. This can be very useful for a busy retinol practice.

If your primary care physicians or endocrinologist or referral networks can essentially streamline and prioritize the patients that have referral-oriented disease. Originally in the primary care setting, you may find that we wouldn't be able to bring the patients that actually need treatment to the retina physician in a timely manner. And we know that that does improve outcomes. In terms of the second 2 things we talked about, grading and treatment guidance.

These are still in the research era, and we're not seeing them yet applied in the clinical practice. But we're starting to see some of the lessons be taught back to us to help us guide our own practice using AI/ML algorithms to derive prognostication from retinal imaging, for example.

David Hutton:

Is this a technology that's really cost prohibitive? Or is it something that could be adopted by practices nationwide?

Nitish Mehta:

So in 2021, the AMA did release a new CPT code which would allow clinicians to bill government and private insurers for the use of these services. So CPT code 92229, will refer to the imaging of the retina to detect disease with automated analysis and report at the point of care.

So we're hoping payers will continue to incorporate this code, and you could see this actually, in practice being used as an AI tool within retinal imaging practices. In terms of the final stages we talked about in terms of guidance or grading, that we'd have to see if these algorithms are then incorporated, for example, into an EMR or into an OCT platform to basically in real time provide guidance. And we have yet to see how that is going to be implemented.

David Hutton:

And ultimately, how can AI help ophthalmologists provide better outcomes for their patients?

Nitish Mehta:

So this is a great question that can be answered by an example.

So imagine you have a patient with retinal vein occlusion, he's been receiving, or she has been receiving monthly anti-VEGF injections for about 6 months. And now you'd like to consider treating at an as needed basis. So clinical trial data would suggest that these patients do well. But a study that we actually performed here at NYU took clinical data from the Copernicus and Galileo trials and input all of the imaging and clinical data into an ML algorithm.

And then we asked are there clinical or imaging biomarkers present during the first 6 months that could present, that could predict the outcome during the as needed treatment arm. And what was fascinating about the outcome of the Machine Learning model was that patient that had higher central subfield thickness, or thicker retinas at baseline, or at week 4, were more than likely to receive 2 as needed treatments during the fixed dosing window, which was the following year.

So here you have actionable clinical information. So you have a patient that has thick retinas at baseline, now you can counsel the patients that they're more likely to need more injections in the first year than after the second year. So here's an example where you've utilized an advanced statistical method, a Machine Learning statistical method, to provide you actionable clinical guidance and information. So as this research continues to progress, you'll be able to see these lessons that we derive from AI being put, in terms of treatment guidance into clinical practice.

David Hutton:

And what are some of the limitations to this technology?

Nitish Mehta:

So this is an incredibly important question, because these models are only as good as the data that you provide, and the questions that you ask on the data. And this has to do with the actual external validation of your model.

So if you train, for example, on a certain subset of patients, and try to apply that in a different subset of patients, you may not get the greatest results. This was most famously demonstrated recently by a Google group that had a model for primary care referrals, and applied it abroad and had very poor outcomes in terms of prediction ability. So who you train on is very important on what you can actually use that for.

Secondly, there's a lot of concerns about explainability of the models. So how are they deriving the results, and you have blackbox models that provide you outcomes, it's very difficult sometimes to understand where that's coming from and what you're learning about it. So there has to be transparency in the model development, transparency in the ground truths.

There are regulatory considerations. These can have impact on patient outcomes that we'd have to track. Of course, we have to talk about cost effectiveness. How would we logistically incorporate this into clinical workflow? And then of course, ethical considerations. As in a very important field in AI right and now is making sure that we do not incorporate biases into the data sets that could then impact patients in the real world.

David Hutton:

And lastly, looking to the future, where do you see AI and retina imaging going in, say the next 3 to 5 years?

Nitish Mehta:

So I see AI helping us reveal features within imaging that we haven't had the foresight or the ability yet to answer. We have a large set of biomarkers in our retinal imaging. Perhaps there's things that we haven't looked at yet. I could also see AI and retinal imaging, allowing us to collaborate effectively with other specialties.

For example, a patient in a retina clinic gets scanned with an OCT, we may be able to reveal neurodegenerative disorders, cardiovascular risk that we can help then get the patient to the right place and get the care that they need. Challenging clinical questions regarding drug choice or surgery choice, for example, may be guided by AI based clinical guidance.

And most importantly, I do think that the route for patients to present to our retinal clinics will definitely be guided by AI based screening platforms.