Generating Post-Acetazolamide Cerebral Blood Flow MRI for High-risk Stroke Patients

Rydham Goyal, Camila Gonzalez, Sasha Alexander, Aja Zou, Michael Moseley, Gary K Steinberg, Moss Zhao

Keywords: Generative models, Conditional Autoencoder, Pixel-space diffusion, Image-to-image translation, Arterial Spin Labeling, MRI, Cerebrovascular reserve

TL;DR: We show that a simple conditional autoencoder can synthesize post-acetazolamide CBF maps directly from baseline ASL, enabling drug-free estimation of cerebrovascular reserve in high-risk stroke patients.

Abstract: Cerebrovascular reserve (CVR) quantifies the brain’s ability to augment cerebral blood flow in response to a vasodilatory stimulus and is a key biomarker in Moyamoya disease and other steno-occlusive cerebrovascular disorders. Clinically, CVR is typically assessed by administering acetazolamide (ACZ) and acquiring post-ACZ perfusion maps, but this workflow is time-consuming, costly, and contraindicated in a subset of patients. In this work, we investigate whether deep learning models can predict post-ACZ perfusion directly from baseline arterial spin labeling (ASL) MRI, enabling pharmacological-free CVR estimation. We curate a single-center dataset of Moyamoya ASL perfusion imaging, comprising pre/post-ACZ scan pairs from 194 patients. We design a post-ACZ conditional Autoencoder (cAE) network to regress the middle axial post-ACZ slice from the corresponding pre-ACZ slice using a combined L1 and SSIM loss. We evaluate our method against two diffusion-based approaches: (1) a conditional Diffusion model implementing a 2D DDPM that learns to denoise post-ACZ slices conditioned on pre-ACZ inputs, and (2) Cold Diffusion model which replaces stochastic Gaussian noise with deterministic interpolation between pre- and post-ACZ images as the degradation operator. On a held-out test set of 49 patients, our proposed post-ACZ cAE encoder achieved the highest reconstruction fidelity (SSIM ∼ 0.78) with a principled generative formulation. Region-wise analysis of CBF percentage change in affected versus healthy MCA territories showed that model predictions generally followed ground truth patterns of cerebrovascular reserve. Our results demonstrate the feasibility of non-invasive CVR assessment using MRI for high-risk patients. These findings suggest that our data-driven approach could reduce reliance on ACZ challenges in routine clinical workflow and expand access to CVR testing to evaluate brain health.

Primary Subject Area: Generative Models

Secondary Subject Area: Application: Neuroimaging

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 368