Optic Disc Colour Ratio: An Image-Derived Proxy for Subgroup Fairness Analysis in Diabetic Retinopathy Screening

<title>Abstract</title> <p> Automated diabetic retinopathy (DR) screening achieves specialist-level accuracy on curated benchmarks, yet subgroup performance disparities are not well characterised. We introduce the Optic Disc Colour Ratio (ODCR), an image-derived measure computed from the peripapillary annulus in CIE L <italic>a</italic> b* space and motivated by choroidal pigmentation, though it is not validated against ground-truth ethnicity or pigmentation labels; we use it only to define image groups for a five-variant ablation on EyePACS (patient-level splits, locked test set n = 5,270). What is established: a specific training recipe — dark-group oversampling, dark-fundus augmentation, and pos_weight=2.0 combined (the Full variant) — produces the only statistically significant sensitivity gap between ODCR-dark and ODCR-light groups (Δ = +0.107, p = 0.015, h = 0.218); the same gap in the unmodified baseline model is not significant (Δ = +0.065, p = 0.100). What is not established: that a stable disparity exists in the unmodified model, or that pigmentation causes the gap through a verified contrast-loss pathway — only training recipe and ODCR group were manipulated, pigmentation and contrast were not. A separate, associated observation is a group-level SNR reduction between ODCR-light and ODCR-dark image groups (1.14×, Mann–Whitney p = 0.0005; three-group Kruskal–Wallis p = 3 × 10⁻¹⁰), alongside a shift toward higher predicted probabilities for ODCR-dark images on true-positive examples in four of five variants, and on true-negative examples in three of five; we call this the score-elevation signature, a descriptive label for the shift, not a claim about its cause. The specificity gap is small and non-significant in every variant (p &gt; 0.13), so the referral-burden estimate — about four excess referrals per 1,000 ODCR-dark images in the Full variant — is directional, not a significant effect, and is computed per image, not per patient. The gap is invisible to AUC (light–dark gaps ≤ 0.008, all non-significant). As an exploratory stress test, zero-shot transfer to APTOS 2019 (n = 3,662, Indian population) retains discriminative AUC but does not replicate the score-elevation pattern on its primary (dark-vs-medium) comparison, where the direction reverses in all five variants — a result we report as a failure to transfer, not a confirmation. Class balancing reduces the gap relative to the Full variant but not below the unmodified baseline, indicating that training recipe, not data imbalance alone, drives most of the significant gap. Each variant was trained once with a single seed, so this single significant result cannot be distinguished from run-to-run variation. Code is available at https://github.com/aaronjji/fundusnet. </p>