Abstract
Cupping of the optic nerve head, a highly heritable trait, is a hallmark of glaucomatous optic neuropathy. Two key parameters are vertical cup-to-disc ratio (VCDR) and vertical disc diameter (VDD). However, manual assessment often suffers from poor accuracy and is time-intensive. Here, we show convolutional neural network models can accurately estimate VCDR and VDD for 282,100 images from both UK Biobank and an independent study (Canadian Longitudinal Study on Aging), enabling cross-ancestry epidemiological studies and new genetic discovery for these optic nerve head parameters. Using the AI approach we perform a systematic comparison of the distribution of VCDR and VDD, and compare these with intraocular pressure and glaucoma diagnoses across various genetically determined ancestries, which provides an explanation for the high rates of normal tension glaucoma in East Asia. We then used the large number of AI gradings to conduct a more powerful genome-wide association study (GWAS) of optic nerve head parameters. Using the AI based gradings increased estimates of heritability by ~50% for VCDR and VDD. Our GWAS identified more than 200 loci for both VCDR and VDD (double the number of loci from previous studies), uncovers dozens of novel biological pathways, with many of the novel loci also conferring risk for glaucoma.
Introduction
Glaucoma is the leading cause of irreversible blindness worldwide. To catch it early, clinicians look at the optic nerve head (ONH), specifically measuring the vertical cup-to-disc ratio (VCDR)—the size of the "cup" in the center of the nerve relative to the nerve itself. A larger cup often signals the loss of retinal ganglion cells.
The challenge is scale. Manual grading of these images is slow, expensive, and suffers from human inconsistency. Previous genetic studies (GWAS) were limited by the number of human-graded images available, leaving much of the "heritability" (the genetic component) of these traits undiscovered. Furthermore, we haven't fully understood why glaucoma prevalence varies so much across different ancestries.
Methods and Study Design
The authors employed a transfer learning approach using a ResNet-34 convolutional neural network (CNN). The AI was designed for three specific tasks: determining if an image is "gradable" (quality control), estimating VCDR, and estimating vertical disc diameter (VDD). Once trained, the AI could process ~300,000 images in about an hour—a task that would take clinicians years of collective effort.
The study leveraged two massive population cohorts: the UK Biobank (UKB) and the Canadian Longitudinal Study on Aging (CLSA). These AI-generated "phenotypes" were then used as the basis for a massive Genome-Wide Association Study (GWAS) to find links between specific genetic variants and eye morphology.
Key Results and Insights
AI Consistency and Genetic Discovery
Surprisingly, using AI labels increased the estimated SNP-based heritability of these traits by ~50% (from 0.22 to 0.35 for VCDR). This suggests that the AI is more consistent than human graders, reducing the "noise" in the data and making it easier for genetic signals to stand out. The study identified over 200 genetic loci for both VCDR and VDD—effectively doubling the number of known loci compared to previous studies using manual gradings.
Solving the East Asian Paradox
East Asians often suffer from "normal-tension glaucoma," where the eye is damaged despite having low intraocular pressure (IOP). The AI analysis revealed that East Asians naturally have higher VCDRs (larger cups). This genetic predisposition to a larger cup counteracts their lower IOP, explaining their high glaucoma risk.
Biological Pathways
Pathway analysis of the new genetic loci revealed unexpected but robust links between eye development and kidney development (involving genes like PAX2, BMP2, and BMP4). This points toward shared biological pathways in the development of different tube-like or filtering structures in the body.
Conclusion
While the AI performed exceptionally well, a small percentage of images remained "ungradable" due to poor quality. Future work could use adversarial architectures to further improve robustness. Additionally, while this study significantly expanded data for non-European ancestries, these groups are still underrepresented in major biobanks. Nonetheless, this research demonstrates how AI-driven phenotyping can unlock massive datasets to provide new insights into disease risk and human genetics.
Summary updated