Abstract
Purpose :
To improve glaucoma diagnosis, we estimated an individualized baseline of retinal nerve fiber layer (NFL) thickness based on vessel patterns and other confounding factors.
Methods :
In this study, normal participants were selected from the Hong Kong FAMILY Cohort, a large random sample of people from different districts in Hong Kong. The study involved measuring the NFL thickness (NFLT) and the area of the optic disc using spectral-domain OCT scans of the optic nerve head region. To create a customized normal reference for each eye based on its vascular pattern, a generative deep-learning model was trained. This model also accounted for other features such as age, gender, axial length, spherical equivalent error, signal strength index, and disc area. The generative deep learning model utilized two parallel conditional variational autoencoder (CVAE) models to reconstruct the NFLT and vascular pattern. The individualized baseline using the deep learning model was compared to the reference using population means without any adjustment, or adjusted by multiple linear regression (MLR). Cross-validation was used for training the models and evaluating the performance.
Results :
Total of 1152 healthy eyes were divided into four subgroups: high myopia (SE<-6 D), low myopia (SE -6 to -1 D), emmetropia (SE -1 D to 1 D), and hyperopia (SE> 1 D). Compared to the population mean reference, the individualized baseline significantly reduced the prediction error for overall and quadrant averages of NFLT (from 9.0~14.3 to 8.2~13.2 µm. The individualized baseline also decreased the false-positive rate of identifying abnormal NFL thinning in myopia groups (high/low) for overall average (from 13.3%/27.1% to 6.7%/6.3%) and for quadrant sectoral averages (p<0.0125) except the temporal sector, where all models had good false positive rates (<7.3 %). Analysis of the NFLT profile also showed that the individualized baseline significantly reduced prediction error around the arcuate bundles, compared to both population mean and MLR-adjusted reference.
Conclusions :
The generative AI approach can be used to construct an individualized NFLT baseline profile using the vascular pattern derived from the same OCT scan. The individualized baseline reduced the prediction error of NFLT in healthy eyes and prevented the false positive identification of abnormal NFL thinning in myopic eyes.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.