Personalizing Circumpapillary Retinal Nerve Fiber Layer Thickness Norms with Individual Retinal Anatomy in Glaucoma

Abstract

Purpose : To generate individualized RNFL norms based on optic nerve anatomy from OCT-derived fundus images using deep learning and subsequently generate RNFL deviation maps to predict VF MD.

Methods : Reliable pairs of Cirrus OCT scans and 24-2 pattern visual fields (VFs) tested within 30 days were included in this study. First, we trained a deep learning model to use the OCT en-face fundus image to predict corresponding RNFLTs (Figure 1a) in eyes with normal VFs (mean deviation (MD) ≥ -1 dB and normal glaucoma hemifield test [GHT] and pattern standard deviation [PSD] results). The R-squared (R²) was used to measure the accuracy of RNFLT map prediction. Second, we used our deep learning model to predict corresponding RNFLT maps from OCT fundus images of eyes with normal and abnormal VFs. The RNFLT deviation map was computed by subtracting the predicted normal RNFLT maps from the actual RNFLT maps. To validate the utility of our personalized RNFLT norm prediction, we compared if the average RNFLT deviation is better correlated with MD compared with average actual RNFLTs.

Results : We included 18,000 reliable pairs of OCT scans and VFs from 17,835 eyes of 13,821 patients with an age of 57.2 ± 16.1 years. The average MD of VFs was -1.4 ± 3.7 dB. 10,000 OCT scans with normal VFs (MD: 0.06 ± 0.72 dB) were used to train the deep learning model, and the remaining 8,000 OCT scans with a mean MD of -3.3 ± 4.9 dB (including 2,419 OCT scans with normal VFs, MD: 0.04 ± 0.73 dB) were used for evaluating the prediction performance. The average R² of RNFLT map prediction on 2,419 OCT scans of individuals with normal VFs was 0.81 ± 0.13 (Figure 1b). The correction coefficient (R) of average RNFLT deviations with VF MDs was 0.42, almost doubling the correlation score (R = 0.19) the average actual RNFLT maps achieved. The two examples (Figure 2) show how the deep learning model reasonably predicts the RNFLT map from the corresponding OCT image, even though the actual RNFLT maps showed substantial RNFLT thinning.

Conclusions : The OCT fundus image encoding retinal anatomical information can be used to predict personalized RNFLT norms, which significantly improved the structure-function correlation with VFs. The deep learning predicted norms may improve the diagnostic accuracy of OCT for glaucoma patients.

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

 

Figure 1

 

Figure 2