ASRS 2024: Improving OCTA performance using AI and differential artery-vein analysis

Jennifer I. Lim, MD, FARVO, FASRS, sat down with Sydney M. Crago, Assistant Managing Editor of Ophthalmology Times and Modern Retina, to discuss her ASRS 2024 presentation, "Differential Artery-Vein Analysis Improves OCTA Performance for Artificial Intelligence Classification of Diabetic Retinopathy."

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

Sydney M. Crago: Hi, I'm Sydney Crago with Modern Retina and I'm here today with Dr Jennifer Lim to talk a little bit about her upcoming presentation at ASRS 2024. Dr Lim, can you share a little bit about what you'll be presenting?

Jennifer I. Lim, MD, FARVO, FASRS: Absolutely. It's my pleasure to be here, Sydney. And I'll be presenting on the differential artery-vein analysis in eyes with diabetic retinopathy, and how this improves the OCTA classification using artificial intelligence.

SC: Can you talk a little bit about how the data for this study came about?

JL: Yes. So, in the past, we've done some work looking at particular quantitative measures on OCTAs of patients who have diabetes versus patients who don't have diabetes. And we've been very interested in looking specifically at the vascular changes, because we know that in certain diseases like diabetes, that the arteries have changes that are different from the changes that we see in veins. For example, when we look at the vascular caliber, in patients with diabetes, we see that the arteries narrow, whereas the veins get wider. So when we look at the ratio, the artery-to-vein ratio specifically for caliber, we found that this can differentiate patients with mild, moderate and severe diabetes from each other better than just looking at the overall vascular caliber. And it makes sense, because the polarities [are] in different directions. And so what we sought to do in this particular study is see whether using this differential analysis of our arteries and veins change over time, in different various severity states of diabetes, and whether this affects the classification system using AI.

SC: AI has been a very popular tool in diabetes diagnosis, with diabetic eye disease. How is this AI kind of overlapping those efforts?

JL: That's a great question, Sydney. And of course, artificial intelligence alone, looking at images, can differentiate mild from moderate from severe diabetes, in terms of retinopathy. But we've been very interested in seeing whether applying this differential analysis of arteries versus veins can even improve this classification. And so in our study, we looked at patients who had diabetes without retinopathy, compared them to control eyes of patients who had no diabetes, and also compared them to patients who had mild, moderate and severe diabetic retinopathy. And what we found when we did this differential analysis, using OCTA images and the quantitative analysis that I alluded to earlier, was that indeed, the artery-vein differentiation can improve the classification by anywhere from 6 up to 16% in these patients. That is, the accuracy can be improved in differentiating, say, control eyes from eyes that have diabetes, but no retinopathy, or that patients who have diabetic retinopathy mild, moderate or severe. And in this project, we also look specifically at regional analysis in addition to artery-vein analysis. And we found that when we look at, say, the parafoveal versus perifoveal zones, and this differential analysis, that this, too, increases the accuracy of the classification. In fact, this regional analysis can improve the classification accuracy by about 9%. It was also interesting that when we looked specifically at multi-class classification, that is differentiating mild, moderate and severe from each other, versus, say, a two-class classification, where we just differentiate moderate from severe—in each of these, again, artery-vein differentiation can also improve the classification within each of these regional analyses. So, I think overall, it's a very useful distinguishing analysis to differentiate arteries from veins when we're using AI to improve the accuracy. And hopefully in the future, these algorithms will be applied in AI systems so that we can improve the classification in terms of sensitivity, specificity and overall accuracy.