Spatial Normalization Improves Cross-Domain Retinal Layer Segmentation in OCT

arXiv·medium signal

Iker Moran-Cavero, Monica Hernandez and Elvira Mayordomo (arXiv 2607.16065, cs.CV/cs.AI) address the domain-shift problem that breaks OCT retinal layer segmentation when models trained on one scanner are applied to another, using spatial normalization as the bridge. Retinal layer thickness is the input to a range of quantitative biomarkers, so segmentation that survives a scanner change is the gating requirement for multi-site clinical deployment. A useful reminder that cross-device generalization, not headline accuracy, is what blocks medical imaging models.