Volumetric reference data of the orbit: a deep learning MRI analysis in the German national cohort
TL;DR
This deep learning tool effectively resolved the bottleneck of manual segmentation, providing comprehensive orbital reference data for the German population and enabling future high-throughput epidemiological research into the associations between orbital anatomy, systemic health, and disease.
Abstract
Manual segmentation of orbital magnetic resonance imaging (MRI) is labor-intensive, hindering large-scale morphometric studies. To overcome this, we developed a fully automated deep learning pipeline to segment orbital MRIs and establish age- and sex-stratified normative reference data. We analyzed T1-weighted brain MRIs from 30,868 participants in the population-based German National Cohort (NAKO). After quality control, 28,779 participants (mean age 48.1 years; 44.1% female) were included. The model, validated against expert manual segmentations, accurately extracted 34 volumetric and geometric parameters across 15 orbital structures (Dice Similarity Coefficients: vitreous 0.97, lens 0.89, optic nerve 0.85). Mean [SD] axial length was 23.5 [1.2] mm. Mean [SD] volumes were 34.4 [3.9] cm³ for total orbital contents, 6.3 [0.8] cm³ for the vitreous, and 0.17 [0.03] cm³ for the lens. Males exhibited significantly larger dimensions across all parameters (p < 0.001). Age-stratified percentile curves revealed continuous age-dependent lens growth (Spearman’s ρ = 0.57 in men, 0.51 in women) alongside modest volume increases in the orbit, optic nerve, and extraocular muscles. This deep learning tool effectively resolved the bottleneck of manual segmentation, providing comprehensive orbital reference data for the German population. This foundation enables future high-throughput epidemiological research into the associations between orbital anatomy, systemic health, and disease.