Manual assessment of optic disc parameters can be time intensive and vulnerable to intergrader variability, leading to possible phenotypic inaccuracy and misrepresentation of the genomic architecture.
In this study, the authors apply artificial intelligence (AI) to assess optic disc morphology and facilitate gene discovery in glaucoma. Using a transfer learning approach, the authors train a convolutional neural network model to estimate imaging gradeability, vertical cup-to-disc ratio (VCDR), and vertical disc diameter (VDD) using fundus photos from the UK Biobank (UKB) and an independent dataset, the Canadian Longitudinal Study on Aging (CLSA). They also compare differences in glaucoma-related traits across ancestries and perform genome-wide association study (GWAS) for VCDR and VDD as determined by AI using three cohorts: the International Glaucoma Genetics Consortium, UKB and CLSA.
Using the AI-phenotypes, the authors identified more significant single nucleotide polymorphisms (SNPs) in their GWAS of VDD-adjusted VCDR (UKB and CLSA cohorts) than previous GWAS using manual optic disc gradings (164 vs 76, respectively) and SNP heritability increased from 0.22 to 0.35. They also identified high concordance between AI-based and clinician-based gradings. After adjusting for differences in VCDR and IOP, the glaucoma risk in European and non-European ancestral groups were quite similar, suggesting differences in disease incidence between ancestry groups is largely affected by glaucoma- related parameters. In the meta-analysis GWAS, the authors identified 111 and 107 significant novel VDD-adjusted VCDR and VDD loci, respectively.
AI may efficiently and more accurately facilitate genomic discovery in glaucoma
These findings strongly suggest that AI may efficiently and more accurately facilitate genomic discovery in glaucoma. Yet, this study does have limitations, as noted by the authors. Specifically, cross validation was done with one dataset and with training completed on images taken from one imaging platform. Furthermore, validation using populations of mixed ancestries is needed. Nevertheless, the authors should be commended on this novel application of AI for glaucoma research.