Enabling Accessible Eye Screening: Privileged Bilateral Learning for Monocular Inference in Individuals with Disabilities

Machine learning systems for medical imaging are often developed under data-rich conditions but deployed when only incomplete information is available. This challenge is especially relevant for individuals with vision impairment or disabilities that make bilateral imaging difficult or impossible. In colour fundus photography (CFP), both eyes are commonly imaged because they provide complementary information for diagnosing ocular diseases such as age-related macular degeneration (AMD) and diabetic retinopathy (DR). However, bilateral acquisition may not be feasible for people with certain physical or cognitive disabilities, or for those with only one eye. We propose a privileged learning framework in which a teacher model has access to bilateral CFP images during training, while a student model learns from the teacher using only a single-eye image. At inference, the student requires only the available eye and does not need the contralateral image. Experimental results show consistent improvements in AUROC over conventional monocular training, with the largest gains for left-eye images for AMD (0.931 vs. 0.966) and DR (0.789 vs. 0.836). By enabling models to benefit from bilateral information during training while requiring only a single eye at inference, our approach may support more inclusive and accessible screening for individuals with disabilities and in resource-constrained settings.