Stanford's CAE3D AI synthesizes post-acetazolamide CBF maps from baseline MRI

A Stanford team led by Julia Huang developed CAE3D, an AI model that generates post-acetazolamide (ACZ) cerebral blood flow (CBF) maps from baseline MRI scans in Moyamoya disease. Moyamoya patients often need cerebrovascular reserve (CVR) assessment to decide on bypass surgery, which normally requires two arterial spin labeling (ASL) scans: one at rest and one after taking acetazolamide. But ACZ is contraindicated for some patients, making the second scan unavailable. CAE3D, a deterministic 3D conditional autoencoder, learns to predict the post-ACZ CBF map directly from the pre-ACZ ASL input, bypassing the need for the drug entirely.

The model was evaluated on a retrospective cohort against ten baselines, including 3D deterministic and diffusion-style models, a 2D contextual baseline, and foundation-model adapters. CAE3D achieved the best performance: lowest MAE of 0.066, SSIM of 0.80, PSNR of 24.0 dB, and near-zero whole-brain mean bias. Its MAE advantage was statistically significant over seven of eight trained-from-scratch baselines, while SSIM and PSNR gains were significant over all eight. The authors note regional delta-CBF predictions compressed the dynamic range in high-response areas. The work is accepted at MLHC 2026, but extension to ACZ-contraindicated patients requires prospective validation.