A system accurately segments eye layers on new devices it has never been trained on.

Eye doctors use special scans called optical coherence tomography, or OCT, to view the thin layers of the cornea. Measuring these layers helps track eye diseases and surgery outcomes. However, computer tools struggle to map these layers automatically because the boundaries are thin, grainy, and look different across scanner brands. To fix this, researchers created ARCOS. Instead of labeling entire regions, ARCOS zooms in on small, overlapping image patches to spot boundary heatmaps, which show where the interface edges likely sit. The network uses multi-scale feature fusion and a self-conditioned refinement step to sharpen spatial details using early boundary hints. Then, it stitches these patches together to build continuous, ordered boundaries across the full scan. ARCOS mapped matched-device boundaries with 95.1% off-by-one accuracy. When tested on unseen devices without extra training, it reached 84.3% accuracy, beat standard baselines, and yielded low error for layer thickness measurements.

Why it matters

Doctors use eye scans to diagnose and treat corneal conditions. This tool can measure layer thicknesses accurately even with different scanning machines, improving consistency in patient care.

Questions answered