Detecting Glaucoma Using Iris Photos with Deep Learning

Abstract

Purpose : To assess deep learning’s performance using iris photos to detect glaucoma.

Methods : We included 24,253 patients from Massachusetts Ear and Eye tested between 2010 and 2022 with 112,243 iris photos. Iris photos were captured during the optical coherence tomography test. Eyes with a visual field (VF) mean deviation (MD) < -3 dB and abnormal glaucoma hemifield test (GHT) and pattern standard deviation (PSD) results were categorized as glaucoma. Eyes with VF MD greater >= -1 dB and normal GHT and PSD results were categorized as non-glaucoma. Our data was split into training (80%), validation (10%), and testing (10%) subsets. We used the four most widely used deep learning models (VGG-19, MobileNet, EfficientNet, and Vision Transformer [ViT]) in our study. The area under the receiver operating characteristic curve (AUC) was used to assess model performance, which was then compared by bootstrapping with t-test. We also created Grad-CAM maps to understand the important regions the deep learning models focused on.

Results : The average age of our patients was 62 ± 16 years. A summary of our results is presented in Figure 1. VGG-19 and ViT performed the best (AUC: 0.736 for VGG-19 and 0.729 for ViT), which were significantly better (p < 0.001) than MobileNet (0.702) and EfficientNet (0.698). There was a notable difference in model performance between males and females, with males achieving a better performance (male AUC: 0.745, female AUC: 0.730, p < 0.001). Compared to Asian and White demographic groups, we found that our model performs better on Black patient groups (Asian AUC: 0.739, White AUC: 0.736, Black AUC: 0.748, p < 0.001). We also found that the model performed slightly better on Hispanic groups than non-Hispanic groups (Hispanic AUC: 0.737, non-Hispanic AUC: 0.735, p = 0.003). In addition, analyzing the generated regional importance maps with VGG-19, it seems our model primarily focuses on areas of the peripheral region of the iris (Figure 2).

Conclusions : Models trained on iris photos have the potential to perform glaucoma detection tasks. In addition, there are substantial performance disparities in different gender and racial groups. Deep learning models also tend to focus on the periphery of the iris when assessing for glaucoma.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

 

Generated regional importance map, where red regions are more important. The top two eyes are glaucomatous, while bottom two are normal.