Automated triage of open globe injury from external-eye photographs for low-resource settings

Abstract Open globe injury is a vision-threatening emergency whose outcome depends on rapid escalation to surgical repair, yet in rural, military, and other low-resourced settings the escalate-or-not decision is usually made by non-ophthalmologists without access to a slit lamp. We present a smartphone-deployable triage tool that detects likely open globe injury from an ordinary external-eye photograph and runs fully offline, with no image or protected health information leaving the phone. From a heterogeneous, multi-source, real-world cohort (1,305 images, 58 open globe positives) with four-reviewer adjudicated labels, we trained and evaluated vision backbones under a de-duplicated, leakage-free split with a frozen held-out test (261 images, 12 positives). A fine-tuned ViT-B reached AUROC 0.934 on the frozen held-out test (0.930 mean across ten seeds; Supplementary Fig. S3); a frozen-feature probe and ensemble, with no backbone fine-tuning, nearly matched it (VisionFM probe 0.906, ensemble 0.908) while running on edge hardware, with a conservative cross-validation estimate of 0.88 to 0.89. We package the model into a working iOS and Android application and map its use across military, community, emergency department, and telemedicine pathways. To our knowledge, this is the first automated open globe injury triage signal from external-eye photographs, delivered as a deployable, offline tool for low-resourced settings.