Abstract
Multiple factors - from normally varying characteristics including age and sex to various disease processes - contribute to individual differences in how people see. In turn, these factors may be associated with subtle differences in the anatomy of the neural structures underpinning vision, from the eye to visual cortex. We - the OCTAHEDRON project team - examine whether AI models can learn to predict individual characteristics and diagnose neurodegenerative diseases from such structural variations embedded in retinal OCT images. The AI models are built from large datasets of annotated and unannotated OCT images, from northeast England NHS Hospital trusts, the UK Biobank and elsewhere. One model exploits a CNN-based retinal layer segmentation algorithm (NDD-SEG), designed to be robust across individuals, diseases and imaging instruments, to generate thickness maps feeding a further classification model which differentiates between individuals with and without multiple sclerosis, achieving 97% balanced accuracy. Other results I will describe compare different techniques – CNN, transformer and traditional machine learning regression methods – to predict sex and age, and ultimately to differentiate between generally healthy and unhealthy ageing trajectories.