Optical coherence tomography produces cross-sectional images of the retina, and separating out its layers is the first step before you can measure anything. Those measurements are drawing interest well beyond eye care, including as possible markers in neurodegenerative disease.

The trouble is that the images fight you. There is speckle noise, there are shadows cast by structures above, adjacent layers can sit at nearly the same brightness, and no two retinas are shaped quite alike. On top of all that sits domain shift: change the scanner, the acquisition protocol or the patient population, and a model that performed beautifully starts to slip.

That last problem is the awkward one, because it is not fixed by more accuracy on your own test set. A system trained on handwriting from one person can look flawless until it meets somebody else's. This work targets that generalisation gap directly through spatial normalisation, rather than treating segmentation quality on a single dataset as the finish line.