Reconstruction fidelity is not clinical utility

If a model could generate the scan most clinics cannot do, from the one they can, would you still need the machine?

Optical coherence tomography (OCT) is the standard retinal scan. OCT angiography (OCTA) adds blood flow, down to the capillaries, and far fewer clinics have a machine that can do it. So recent work predicts the angiogram from a plain OCT scan the patient already had, reporting an impressive peak signal-to-noise ratio above 31 dB and structural similarity above 0.9 against the real thing.

We tested it. The generated images look right and score like it. The clinical measurement does not follow.

And it fails where it matters most. The outer retina should carry no blood vessels at all, and new vessels growing there are what puts a patient on treatment once they leak. On our own hospital data, neither method reproduced that lesion in the eyes we checked.

Reconstruction fidelity is not clinical utility. We call the difference the reconstruction-utility gap, and we hope closing it becomes the target: evaluation that reflects what the scan is actually for.

"In Defense of OCTA: The Reconstruction-Utility Gap in OCT-to-OCTA Synthesis," at OMIA 2026, the Ophthalmic Medical Image Analysis workshop at MICCAI 2026, with co-authors at Rabin Medical Center and Tel Aviv University. The paper is on arXiv: arxiv.org/abs/2608.15626.