3D Generative AI Synthesizes Post-Acetazolamide CBF Maps in Moyamoya
A recent study has unveiled CAE3D, a deterministic 3D conditional autoencoder designed to create post-acetazolamide (ACZ) cerebral blood flow (CBF) maps from baseline MRI scans in individuals with Moyamoya disease. This research, available on arXiv (2608.14758), tackles the difficulty of evaluating cerebrovascular reserve (CVR) when ACZ cannot be used. Typically, CVR assessment relies on paired arterial spin labeling (ASL) perfusion MRI conducted before and after ACZ; however, if ACZ is contraindicated, post-ACZ CBF maps are inaccessible. CAE3D produces these maps directly from pre-ACZ ASL data, achieving a minimal held-out mean absolute error (MAE) of 0.066, SSIM of 0.80, PSNR of 24.0 dB, and nearly zero full-brain mean bias. The model was tested against ten alternatives, showing a statistically significant MAE advantage over seven out of eight trained-from-scratch baselines. This innovation may facilitate hemodynamic assessments without ACZ, broadening the evaluation process for surgical candidacy in Moyamoya patients.
Key facts
- CAE3D is a deterministic 3D conditional autoencoder for synthesizing post-ACZ CBF maps.
- It uses pre-ACZ ASL input to generate post-ACZ CBF maps.
- CAE3D achieved the lowest held-out MAE of 0.066.
- SSIM was 0.80 and PSNR was 24.0 dB.
- Near-zero full-brain mean bias was reported.
- Evaluated against ten comparators including deterministic and diffusion-style 3D baselines.
- MAE advantage was statistically significant over seven of eight trained-from-scratch baselines.
- Study published on arXiv with identifier 2608.14758.
Entities
Institutions
- arXiv