Abstract
Purpose :
Early detection of glaucoma is critical in preventing irreversible vision loss. This study evaluates the efficacy of combining deep learning (DL) analysis of color fundus photos (CFPs) with a polygenic risk score (PRS) in improving glaucoma detection.
Methods :
We developed DL models for glaucoma classification using over 4,000 glaucomatous and 100,000 normal CFPs, employing the ConvNeXt architecture pre-trained on ImageNet 1K. For the PRS, we utilized cross-ancestry meta-analysis GWAS summary statistics, excluding the UK Biobank (UKB) cohort, and applied the C+T method to derive a weighted PRS for primary open-angle glaucoma (POAG). Our DL models and PRS were evaluated on a subset of 270 POAG cases and 18,642 healthy controls from the UKB, all with available CFPs and genetic data. The performance was assessed using logistic regression and the area under the receiver operating characteristic curve (AUC).
Results :
Both the DL score from CFPs and the PRS showed significant associations with POAG, with p-values of 8.76x10-153 and 9.78x10-8, respectively, after adjusting for age, sex, and the top four principal components of genetic ancestry. The AUC was 0.963 (95% CI: 0.952 - 0.974). Utilizing the Youden index to determine the optimal cutpoint, we achieved a sensitivity of 0.856 and a specificity of 0.945.
Conclusions :
The study demonstrates that the DL score from CFPs, combined with PRS, significantly enhances the detection of glaucoma. This integrated approach of DL and genetic risk scoring presents a promising tool for glaucoma screening, potentially facilitating earlier intervention and better patient outcomes.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.