Detecting diabetic retinopathy with deep learning and ultra-widefield scanning laser ophthalmoscope images

Article

During her talk at Angiogenesis, Dr. Loewenstein outlines how artificial intelligence could revolutionize diabetic retinopathy screening.

Dr. Anat Loewenstein discusses her presentation at the Angiogenesis, Exudation and Degeneration meeting. Her talk, “Detection of Diabetic Retinopathy from Ultra-Wide Field Scanning Laser Ophthalmoscope Images: A Multi-Center Deep-Learning Analysis,” delves into how artificial intelligence can aid in patient screening. In testing, the deep learning model scored well in sensitivity, specificity and accuracy.

Video transcript

David Hutton: I'm David Hutton of Ophthalmology Times. The Bascom Palmer Eye Institute at the University of Miami is hosting its Angiogenesis, Exudation and Degeneration 2022 virtual edition. Joining me today is Anat Lowenstein, MD, who is presenting “Detection of Diabetic Retinopathy from Ultra-Wide Field Scanning Laser Ophthalmoscope Images: A Multi-Center Deep-Learning Analysis.” Thank you for joining us today. Tell us about your presentation.

Dr. Loewenstein: Thank you, David. I'm very happy to tell you about my presentation on diabetic retinopathy detection from ultra-wide field scanning laser ophthalmoscope images.

So we all know that usually diabetic retinopathy screening uses a limited field of view, 45 to 50 degrees. And theoretically, ultra-wide field imaging can be an alternative to the existing model. But interpretation of these images does require expertise of human graders that are very knowledgeable in diabetic eyes.

We're suggesting models of deep learning to detect variability and assess the referral and vision-threatening diabetic retinopathy, which are important parameters that we want to use.

So in order to do that, we develop models to assess both ability to evaluate a referral diabetic retinopathy and vision-threatening one. And we use the transfer learning procedure for each such a model.

Basically, we looked at almost 10,000 images from almost 2,000 eyes with diabetes. And we had sets for training and validation, the primary set and external testing the external data sets. And when we looked at the sensitivity, specificity and accuracy, we could see that they're pretty high for the primary dataset, 86.5% 82.1% and 83.1%, respectively. So pretty high sensitivity, specificity and accuracy, which were shown by our model.

Also, we were able to show on the receiver operating curves that the performance of the deep learning is on referable diabetic retinopathy were pretty high. So pretty good results on that front as well.

And also for vision-threatening diabetic retinopathy. Again, the area under the receiver operating curve, it was able to show us the rate of accuracy pretty high with our model.

Basically, we were able to show that the diagnosis performance of the deep learning system for detecting referrable diabetic retinopathy and vision-threatening diabetic retinopathy in both primary and external validation sets were pretty high with the ratio of sensitivity, specificity and positive predictive value for both was around 90%. So pretty high and good performance.

And we will show some examples of AI with diabetic retinopathy with referral that the current neuropathy and so forth. So basically, we are able to show high performance in both variability and severity of diabetic retinopathy. And we suggested as a model to facilitate screening of patients with diabetes, for referrable and vision-threatening diabetic retinopathy.

We believe that such readability is crucial for the screening of this disease, it can identify hidden lesions and allows for objective identification of diabetic retinopathy severity.

Of course, our study has advantages and limitations. One of the advantages is that we have multiple ethnicities and geographical location.

But of course, there are also limitations such as the fact that some lesions cannot be captured by the system, maybe artifacts that do affect the prediction. And in the proliferative diabetic retinopathy primary data set, we had a limited number of images.

Of course, the fact that OCT imaging was not used to evaluate the ground truth of diabetic macular edema is a major limitation of our trial. But I think I can conclude that the deep learning system did demonstrate their high rate of identifiable referable diabetic retinopathy cases and that the model has the potential to be implemented in clinical settings, and therefore reduce the burden of both patients and clinicians. Thank you very much, David.

David Hutton: Thank you.

Note: This transcript has been lightly edited for clarity.

Recent Videos
WIO 2024: An educator's perspective on shattering glass ceilings in ophthalmology
Hannah Chiu, MD, FRCSC, highlights some of the early benefits of an AI-operated telephone call system for postoperative patient care at WIO 2024
Retinal Inner Layer Disorganization and OCT in Uveitic Macular Edema: Insights from Dr. Amitha Domalpally
ARVO 2024: Study Reveals Faricimab's Potential for Extended Dosing in nAMD
© 2024 MJH Life Sciences

All rights reserved.