Related Experiment Video
Updated: Sep 15, 2025
Author Spotlight: Ex Vivo OCT-Based Multimodal Imaging of Human Donor Eyes for Research into Age-Related Macular Degeneration
Published on: May 26, 2023
Genotype Prediction from Retinal Fundus Images Using Deep Learning in Eyes with Age-Related Macular Degeneration
Avishai Halev1, Denis Huang2, Shahbaz Rezaei3
1Department of Mathematics, University of California, Davis, Davis, California.
Purpose:
To employ deep learning models to predict high-risk genetic variants associated with age-related macular degeneration (AMD) from retinal fundus photographs of patients with this condition.
Design:
Deep learning algorithm development to classify single-nucleotide polymorphism in the complement factor H (CFH) and age-related maculopathy susceptibility 2 (ARMS2) genes using retinal fundus images.
Participants:
Thirty-one thousand two hundred seventy-one retinal color fundus photographs of 1754 participants from the Age-Related Eye Disease Study.
Methods:
We trained deep learning models including convolution neural networks and vision transformers (ViTs) to classify patients into high-risk (homozygous high-risk alleles) or low-risk (heterozygous or homozygous low-risk alleles) genotypes for CFH or ARMS2, then evaluated algorithm performance on an independent test set. The complexity of genotype predictions was compared with AMD severity or gender classification tasks using V-usable information. Attribution mapping was performed to identify fundus regions used to predict genotype from phenotype.
Main Outcome Measures:
Area under the receiver operating characteristic curve (AUROC), balanced accuracy, and average precision for predicting high-risk genotypes.
Results:
Our trained ViT models predicted high-risk genotypes in CFH and ARMS2 with an AUROC of 0.719 and 0.741 across all eyes, respectively. For genotype predictions for ARMS2, model performance is improved in eyes with advanced AMD (AUROC 0.867), choroidal neovascularization (AUROC 0.833), and geographic atrophy (AUROC 0.957). Genotype predictions from fundus images appear more difficult than AMD severity or gender classification tasks, although saliency mapping supports biological plausibility by demonstrating attention to the central macula for genotype predictions.
Conclusions:
Deep learning can predict high-risk genotypes in CFH and ARMS2 from retinal fundus images of patients with AMD. Our findings provide proof of principle for inferring genotype from noninvasive eye imaging and reveal insights into genotype-phenotype relationships in AMD.
Financial Disclosures:
Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.
