AI/ ai · medical-imaging · ophthalmology · research

A Retinal AI Model That Learns From Scans It Never Sees

EyeMVP was trained on paired eye scans from 112,642 patients to make standard fundus photos smarter — without needing the expensive scan at diagnosis.

A new AI model for retinal screening extracts depth information from cheap eye photos — without actually taking a depth scan.

Researchers trained EyeMVP on nearly 675,000 paired image sets, each pairing a standard color fundus photo with an optical coherence tomography scan from the same eye on the same day, drawn from 112,642 patients across eight hospitals. The idea: during training, the model learns what an OCT scan would reveal, so that at diagnosis time it can work from the fundus photo alone. OCT provides the kind of cross-sectional, depth-resolved detail that fundus photography cannot, but OCT equipment is expensive and far less common in mass screening programs. EyeMVP sidesteps that access gap by baking OCT knowledge into a model that only needs the cheaper image at inference.

The clinical stakes are real: the model hit AUROCs of 0.923 for macular edema and 0.867 for myopic macular schisis, two conditions that are notoriously hard to catch in standard fundus photos. In a reader study, EyeMVP outperformed junior and intermediate ophthalmologists on macular edema — though not senior specialists — and surpassed all human groups on myopic macular schisis. That gap matters for healthcare systems where senior ophthalmologists are the scarce resource.

The honest caveat is that this is a preprint, not a peer-reviewed clinical deployment, and reader studies are a controlled setting — not a busy screening clinic. Still, the approach of offloading expensive modality knowledge into a cheaper-to-run model at inference is a legitimate architectural direction, one that has parallels in radiology AI where MRI-informed CT models have drawn similar interest.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →