Clinical Computer Vision for Retinopathy of Prematurity Screening Under Variable Fundus Camera Illumination: A Domain-Generalization Evaluation of Referral Threshold Safety Margins

Retinopathy of prematurity (ROP) screening programmes increasingly deploy convolutional networks on fundus photographs captured by cameras whose illumination profiles differ across devices, sites and operators. A referral threshold calibrated on one camera may therefore shift when the same network is applied to images from another, and the direction of that shift determines whether the programme misses treatable disease or overwhelms ophthalmology services with false referrals. the present analysis treats the problem as a domain-generalization question about the safety margin of a fixed referral threshold rather than as a question about average accuracy. It formalises illumination-induced covariate shift as a bounded perturbation of the image-formation operator, derives the conditions under which a threshold calibrated on a source domain retains a declared sensitivity floor on a target domain, and defines the safety margin as the gap between the calibrated operating point and the sensitivity floor. The analysis is a methodological and theoretical treatment: no clinical images were collected and no network was trained. Instead, the paper specifies a domain-generalization evaluation protocol that a screening programme can execute with its own cameras, and it characterises how the margin behaves as illumination shift grows. The central result is that threshold safety margins are asymmetric under illumination shift: sensitivity degrades faster than specificity improves, so a threshold that appears conservative on the source camera can become unsafe on a target camera without any change in average accuracy. The paper concludes that ROP referral thresholds should be reported with an explicit illumination-shift envelope and a margin, not as a single operating point.