An AI model that fakes contrast-enhanced CT scans can also flag its own worst mistakes.
Researchers tested whether "uncertainty" - how unsure a diffusion model is about its own output - can substitute for ground truth when checking synthetic medical images. They used AortaDiff, a diffusion framework that generates contrast-enhanced CT (CECT) images from plain, non-contrast CT scans while also producing a segmentation of the blood vessel it just drew. That segmentation gives a concrete, checkable stand-in for anatomical accuracy. The team compared six uncertainty methods - Ensemble, HyperDiff, BayesDiff, MCDropout, RDS, and TTA - across pixel, region, and whole-image levels, and tested detection of out-of-distribution cases on an external multi-centre dataset. MCDropout came out on top: strong at every scale, consistent on outside data, and free to turn on in any model already trained with dropout.
This matters because synthetic CECT exists to cut out contrast dye, which carries its own costs and risks for patients. But a realistic-looking fake vessel is not the same as a correct one, and the usual pixel-similarity scores used to judge these models don't actually check anatomy. A cheap, bolt-on uncertainty signal gives hospitals a way to triage generated scans without extra training runs or added compute.
It's not a cure-all. The uncertainty signal is good at catching obvious failures but, per the researchers, "discriminates poorly among already high-quality images" - meaning it will catch the scan that's clearly wrong, not the one that's subtly wrong. For a tool meant to replace a dye injection, subtly wrong is the failure mode that should worry clinicians most.