Deep Learning Prediction of Personalized Peripapillary Retinal Nerve Fiber Layer Thickness Norms from Fundus Images in Glaucoma
Authors
Affiliations (1)
Affiliations (1)
- Harvard Ophthalmology AI Lab, Schepens Eye Research Institute of Massachusetts Eye and Ear, Harvard Medical School
Abstract
PurposeTo predict retinal nerve fiber layer thickness (RNFLT) norms from fundus images. MethodsWe selected 18,000 OCT scans and visual fields (VF) from the Massachusetts Eye and Ear Glaucoma Service. A U-Net-based deep learning model was developed to predict RNFLT norms from OCT en face fundus images. A total of 10,000 OCT scans with normal VFs (mean deviation [MD] [≥] -1 dB, glaucoma hemifield test within normal limits, and pattern standard deviation probability > 5%) tested within 30 days were used for training, while the remaining 8,000 OCT scans (mean VF MD: -3.3 {+/-} 4.9 dB), including 2,419 scans with normal VFs, were used for evaluation. Structure-function correlations between RNFLT maps and VFs were assessed using linear regression and VGG-16 across original RNFLT maps, deviation maps, and their combination. Performance was evaluated using correlation coefficients, mean absolute error (MAE), and R2. ResultsPredicted RNFLT norm maps showed agreement with baseline RNFLT maps in eyes with normal VFs (R2 = 0.81 {+/-} 0.13). RNFLT deviation maps correlated more strongly with VF MD than original RNFLT maps (R = 0.42 vs. 0.19, p < 0.01). In deep learning-based VF prediction, combining original and deviation maps achieved the best performance (MAE = 3.31 dB, R2 = 0.39), outperforming the model (p < 0.05) using original RNFLT maps alone (MAE = 3.36 dB, R2 = 0.35). ConclusionsDeep learning can estimate individualized RNFLT norms and improve structure-function assessment in glaucoma. Translational RelevancePersonalized RNFLT norm prediction may improve detection of glaucomatous damage.