ARVO 2025: Using graph neural networks to differentiate between fast and slow glaucoma progressors

At ARVO 2025, in Salt Lake City, Utah, Patipol Tiyajamorn, talked about his poster on using graph neural networks to identify fast glaucoma progressors solely using optic nerve head morphology at a single time point

Video Transcript:

Editor's note: The below transcript has been lightly edited for clarity.

Patipol Tiyajamorn:

I'm Patipol Tiyajamorn. I'm from Emory University, and my work is about using graph neural networks to try to differentiate between fast glaucoma progressors and slow glaucoma progressors. What we do is that, we only use a single time point to try to predict whether they will develop fast progression or slow glaucoma progression. And the main motivation behind this is that when patients come to the ophthalmologist periodically, sometimes they can develop a fast glaucoma progression then can worsen before they come at a second visit.

So, let's get into my work. So the way we do this is that we use an OCT scans of the patient before they develop any glaucoma or progression, and then we use a segmentation model to identify the main layers, main tissue layers, and then we can construct a 3D graph network based on that and feed into a network called a graph attention network to try to do classification. After the model has learned the main features and the main landmarks, when they do the prediction, we can also display it to see where are the critical features, critical tissue, that they pulled in the data to do the prediction. And you can see that mainly the laminar corbosa and the inter no limiting membrane are the important features that it looks [for] when it's doing the prediction. We can plot this in 3DSL, and you can see that mainly the key notable features that make it do the prediction is around the nasal inferior part of the IOM. Apart from that, we can also do patient-specific critical landmarks. So, if we perform the model into a single patient, we can tell where MDA patients opting [inaudible] looked wrong and might be the cause of the fast progression.