Researchers have built an AI model that guesses what a Moyamoya patient's brain scan would show after a drug challenge - without giving the drug.
Moyamoya disease narrows arteries in the brain, and doctors judge how badly by measuring cerebrovascular reserve: how much blood flow increases when a patient is given acetazolamide (ACZ), a drug that dilates brain vessels. That before-and-after comparison, captured on MRI, is what surgeons use to decide whether a patient needs bypass surgery. But when ACZ is skipped or off the table for a patient, the "after" scan simply doesn't exist. A team's new model, called CAE3D, generates that missing post-ACZ scan directly from the baseline MRI. Tested against ten rival methods, including diffusion-style models and adapters built on pretrained foundation models, CAE3D came out on top on error and image-similarity scores, with a statistically significant edge over most of them.
The real value here isn't the accuracy numbers - it's what this could remove from a patient's day: a drug challenge, plus the discomfort and rare side effects that come with it. That fits a broader pattern in medical imaging AI, where models increasingly try to infer stress-test-style data from a single resting scan.
The catch is that this was tested only on patients who already completed both scans, and the model underestimated the biggest blood-flow jumps - exactly the territory doctors care about most. The people who'd actually benefit, those who can't take ACZ at all, weren't in the dataset. Promising proof of concept; not yet a substitute for the real test.