This project aims to develop an AI-driven framework that can infer genetic variants associated with glaucoma directly from retinal OCT (Optical Coherence Tomography) images. The central research question is: Can deep learning models trained on OCT scans predict genotype-level information relevant to glaucoma risk?
Our objectives are threefold:
– To train convolutional neural networks (CNNs) on OCT images to classify phenotypic indicators of glaucoma.
– To extend the model to infer gene-level and SNP-level associations from imaging data.
– To validate the model’s predictive performance using UK Biobank’s linked imaging and genetic datasets.
The scientific rationale is based on the hypothesis that subtle structural biomarkers-such as retinal nerve fiber layer thickness and optic disc morphology-encode genetic signals that can be decoded by AI. While traditional genotype-phenotype studies rely on statistical associations, our approach leverages high-dimensional image data and deep learning to uncover latent patterns that may not be apparent through conventional analysis.
UK Biobank’s large-scale, high-quality dataset of retinal OCT images, clinical diagnoses, and genotypes provides a unique opportunity to explore this hypothesis. By integrating these data sources, we aim to build a non-invasive, image-based screening tool for early glaucoma risk stratification. If successful, this approach could be extended to other complex diseases where imaging biomarkers are available, offering a scalable and cost-effective alternative to genotyping in clinical settings.