How deep learning may play a role in predicting retinitis pigmentosa visual impairment

T. Y. Alvin Liu, MD, sat down with Ophthalmology Times to discuss how combined imaging and deep learning may predict visual impairment in individuals with retinitis pigmentosa (RP). This ability may help to speed the screening process for clinical trials of therapies for RP. The results were featured in the British Journal of Ophthalmology.

Editor's note - This transcript has been edited for clarity.

Hello, everyone. Thank you for having me. My name is Alvin Liu. I'm a retina specialist at the Wilmer Eye Institute at Johns Hopkins University. I'm also the founding director of the Wilmer Precision Ophthalmology Center of Excellence. It's a pleasure to be here today. My project was done in collaboration with a close friend and colleague of mine, Dr. Mandeep Singh. Our overall hypothesis or question that we want to answer is whether we can predict function from objective imaging in patients with retinitis pigmentosa using a technology called deep learning.

Deep learning is a subtype of artificial intelligence that has really garnered a lot of attention in recent years, particularly because deep learning is very good at pattern recognition and image analysis. Just broadly speaking, in the field of inherited retinal disease, there are two end point outcomes that we care about. Number one is whether there are changes in structural findings, such as images, and second, the changes in a patient's function.

We thought it would be nice to do a study to see if it's even possible to use deep learning to directly deduce a patient's function from an image alone.

What were the findings?

We start off with a question of whether we can predict visual impairment in an eye with retinitis pigmentosa using deep learning, and a single OCT image slice. By visual impairments, we defined that as a visual acuity of worse than Snellen 20/40 and we demonstrated that we were able to predict with a high accuracy whether a particular eye has better or worse than 20/40 vision, based only on a single foveal scan of the OCT, and fortunately, our work was published in the British Journal of Ophthalmology.¹

What are the next steps?

Honestly, when we first started on this journey, we were not sure whether deep learning was able to predict anything. This really was a hypothesis that deep learning is powerful enough in pattern recognition to be able to deduce that kind of function from an OCT image. Now that we have demonstrated that it's possible and we can deduce functional metrics from OCT images and deep learning, we plan to make the model more fine grained, meaning the current version of the model can only predict whether it's a finer prediction of whether it's better or worse than 20/40.

As a next step, we will train a model to be able to predict different levels of vision. In addition, in this particular study, we looked at combining both OCT images and infrared images for prediction. As the next step we'll also provide models with different kinds of admission, such as all fluorescent images, or genetic information, or other clinical variables that may be relevant to functional prediction.

What does this mean long range?

We think this is exciting for the following reason. Oftentimes, in clinical trials for inherited retinal dystrophy, such as retinitis pigmentosa, the functional end point is very important. However, we know that in the commonly used functional end points that we use in clinical trials, there are drawbacks for each one of them. There are their abilities in measuring visual acuity. I'm going to test this commonly used visual field, it's highly variable as well.

What's really attractive to us when it comes to imaging is that it really is quite objective. The variability, it's a lot lower than typical functional testing. And also the time for acquisition of these images is typically a lot shorter as well.

If we can consistently demonstrate that different functional outcomes can be predicted directly from an objective image that can be obtained very quickly, and using deep learning, that really could change how we do clinical trials for retinal pigmentosa and inherited retinal dystrophies. In addition, the same kind of approach can also be used to stratify patients in terms of the rate of progression and treatment response.

Reference

1. Liu TYA, Ling C, Hahn L, Jones CK, Boon CJ, Singh MS. Prediction of visual impairment in retinitis pigmentosa using deep learning and multimodal fundus images [published online ahead of print, 2022 Jul 27]. Br J Ophthalmol. 2022;bjophthalmol-2021-320897. doi:10.1136/bjo-2021-320897