Abstract
Purpose :
To investigate whether a deep neural network algorithm could predict future rates of retinal nerve fiber layer (RNFL) loss in glaucoma from analysis of enhanced depth imaging (EDI) spectral-domain optical coherence tomography (SDOCT) scans acquired at baseline.
Methods :
The study included 847 eyes of 446 patients. Data were extracted from the Duke Glaucoma Registry (DGR), a database of electronic medical records and research data. All eyes had EDI scans of the optic nerve head obtained with the Spectralis SDOCT. Eyes were followed with SDOCT RNFL scans over time. Rates of global peripapillary RNFL thickness were obtained by linear mixed models. A deep learning (DL) convolutional neural network (ResNet) was trained to predict the slopes of RNFL change from analysis of the baseline raw EDI B-scans. Predictions were compared to the observed slopes by evaluating mean absolute error (MAE), correlation and Bland-Altman plots.
Results :
Eyes were followed for an average of 5.3 ± 2.3 years, and had an average of 9.4 RNFL OCT scans over time. The mean slope of RNFL loss over time was -0.11mm/year, whereas the mean predicted slope was -0.14 mm/year. There was a statistically significant, but relatively weak, correlation between predicted and observed slopes (r = 0.30; R2 = 9%; P<0.001). The MAE was 0.54mm/year.
Conclusions :
Baseline deep learning analysis of deep optic nerve structures was significantly associated with future rates of RNFL loss. However, the association was relatively weak, explaining less than 10% of the variance in slopes. These findings suggest that image analysis of the lamina cribrosa and deep optic nerve structures may have a limited role by itself in predicting which eyes may be at most risk for fast glaucoma progression.
This is a 2020 ARVO Annual Meeting abstract.