Synthesizing Post-Acetazolamide Cerebral Blood Flow Maps from Baseline MRI in Moyamoya Using 3D Generative AI

Julia Huang; Camila Gonzalez; Rydham Goyal; Aja Zou; Sasha Alexander; Michael Moseley; Moss Y. Zhao; Gary K. Steinberg

Source-linked AI summary

Synthesizing Post-Acetazolamide Cerebral Blood Flow Maps from Baseline MRI in Moyamoya Using 3D Generative AI

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

arXiv:2608.14758v1Aug 14, 2026eess.IVcs.AIcs.CV

TL;DR

In Moyamoya disease, missing post-acetazolamide CBF maps can limit CVR assessment when the challenge scan is unavailable. CAE3D synthesizes these maps from baseline ASL and achieved the lowest held-out MAE among the evaluated in-house models, supporting retrospective feasibility.

Problem
Paired pre- and post-acetazolamide ASL maps assess CVR in Moyamoya, but the post-challenge map is unavailable when acetazolamide is contraindicated or avoided.
Method
CAE3D is a deterministic 3D conditional autoencoder that synthesizes post-acetazolamide CBF maps from pre-acetazolamide ASL input.
Results
CAE3D achieved the lowest held-out MAE among nine in-house models, with near-zero Bland–Altman bias and lower territory-level ΔCBF error than evaluated diffusion baselines.
Takeaways & Limitations
The results demonstrate retrospective feasibility of post-acetazolamide CBF synthesis in Moyamoya patients who completed the standard two-scan protocol.
Takeaways & Limitations
This internal validation may not generalize across scanners, ASL sequences, or sites because external cohorts and out-of-distribution conditions were not evaluated.

Abstract

from arXiv · show

For patients with Moyamoya disease, impaired cerebrovascular reserve (CVR) is an important hemodynamic criterion for recommending extracranial-to-intracranial bypass surgery. Standard CVR assessment in this cohort uses paired arterial spin labeling (ASL) perfusion MRI acquired before and after acetazolamide (ACZ). When ACZ is contraindicated or avoided, the post-ACZ cerebral blood flow (CBF) map needed for hemodynamic assessment is unavailable. We propose CAE3D, a deterministic 3D conditional autoencoder that synthesizes post-ACZ CBF maps directly from pre-ACZ ASL input. We evaluated CAE3D against ten comparators, including deterministic and diffusion-style 3D baselines, a 2D contextual baseline, and frozen-encoder foundation-model adapters. CAE3D achieved the lowest held-out MAE (0.066), with SSIM 0.80 and PSNR 24.0 dB, and near-zero full-brain mean bias. Its MAE advantage was statistically significant over seven of eight trained-from-scratch baselines, excluding the 2D CAE_2D comparator; its SSIM and PSNR advantages were significant over all eight. Regional delta-CBF predictions compressed the dynamic range in high-response territories. These results establish the retrospective feasibility of post-ACZ CBF synthesis in patients who completed the standard two-scan protocol. Extension to ACZ-contraindicated patients, who were not represented in this cohort, requires external and prospective validation.

1. Introduction

Moyamoya disease involves progressive arterial narrowing, and impaired cerebrovascular reserve is an important hemodynamic criterion for disease severity and bypass-surgery consideration. This work investigates synthesizing post-acetazolamide cerebral blood flow maps from pre-acetazolamide ASL using 3D conditional image-synthesis models, while emphasizing that clinical replacement remains unvalidated.

Clinical motivation: Moyamoya disease progressively narrows and occludes major intracranial arteries, increasing long-term ischemic and hemorrhagic stroke risk.Cerebrovascular reserve reflects the brain’s capacity to augment cerebral blood flow after vasodilatory stimulation.
Clinical motivation: Impaired cerebrovascular reserve is an important hemodynamic criterion for assessing Moyamoya disease severity and recommending extracranial-to-intracranial bypass surgery.Standard evaluation uses paired pre- and post-acetazolamide ASL scans bracketing intravenous acetazolamide administration.
Study rationale: Synthesizing post-acetazolamide CBF maps from baseline pre-acetazolamide ASL could provide CVR-relevant information without drug administration, pending validation.The proposed direction builds on encoder-decoder networks, diffusion models, and pretrained 3D encoders for neuroimaging synthesis.
Study contributions: CAE3D achieved the lowest held-out MAE among the nine in-house models, with near-zero Bland-Altman bias and lower territory-level ∆CBF error than evaluated diffusion baselines.The work demonstrates feasibility of synthesizing post-acetazolamide CBF maps from pre-acetazolamide ASL input in patients with Moyamoya disease.
Study limitations: The synthesized post-acetazolamide map represents a potential future workflow rather than a clinically validated replacement for acetazolamide imaging.The task predicts an independently normalized post-acetazolamide map from the pre-acetazolamide input.
Generalizable implications: Extending baseline-to-challenge image synthesis to other clinical workflows requires task-specific validation using agreement, regional, and subgroup-level metrics.Global image-similarity scores alone may be insufficient for evaluating such extensions.

2. Related Work

This section situates post-ACZ CBF synthesis within ASL-based CVR assessment in Moyamoya and distinguishes the study’s full-volume 3D, unified evaluation from prior slice-based work. It also motivates baseline-driven prediction through Moyamoya physiology and related perfusion-synthesis approaches.

ASL and CVR: ASL non-invasively quantifies voxelwise CBF, while pre- versus post-ACZ changes encode regional vasodilatory responses used for CVR assessment and bypass planning.The synthesis task predicts the post-ACZ CBF map from the pre-ACZ map in independently normalized voxel space rather than physical CBF units.
Prior synthesis work: Goyal et al. established slice-based conditional and diffusion baselines, whereas this study evaluates full-volume 3D prediction with side-by-side 2D and 3D baselines.The present evaluation additionally includes agreement and territory-level analyses.
Physiological basis: Chronic large-vessel occlusion, collateral recruitment, and territorial steal can make resting pre-ACZ perfusion reflect partially exhausted or preserved autoregulatory reserve.This physiological basis explains why baseline perfusion may carry information about response to a later vasodilatory challenge.
Broader synthesis methods: Related work includes learning-based synthesis of perfusion-related targets from baseline imaging in MRI–PET settings and frozen-encoder comparators based on Med3DVLM and SAM-Med3D.The two publicly released 3D encoders are adapted as frozen-encoder comparators in this study.
3D modeling and evaluation: Medical image synthesis has used diffusion and cold-diffusion models, while this study unifies 3D comparisons with global image-quality metrics and atlas-based regional summaries.The implementation builds on standard MONAI encoder-decoder and diffusion components.

3. Methods

The study framed post-ACZ ASL perfusion synthesis as deterministic supervised image-to-image regression from pre-ACZ maps, using a shared evaluation protocol across 11 models and multiple 3D modeling approaches. CAE3D was implemented as a task-adapted volumetric encoder-decoder with combined L1+SSIM training and full-brain metric evaluation.

Model comparisons: 11 models were evaluated: CAE3D, eight in-house baselines, and two separately evaluated foundation-encoder adapters.The primary held-out comparison included CAE3D plus the eight in-house baselines; CAE3D had ten total comparators.
CAE3D design: CAE3D maps a pre-ACZ perfusion map to a predicted post-ACZ map in one forward pass using deterministic conditional regression.The design prioritizes stable full-volume reconstruction and spatial fidelity rather than multiple possible outputs.
CAE3D design: CAE3D uses a task-adapted MONAI 3D UNet encoder-decoder with skip connections and a combined L1+SSIM objective.The adaptation specifies spatial dimensions, channel widths, strides, residual units, inputs and targets, padding, pre→post mapping, and training protocol.
Experimental protocol: The shared task synthesized post-ACZ ASL perfusion maps from pre-ACZ maps and compared deterministic encoder-decoders, diffusion-style models, FNO 3D, and adapted pretrained 3D encoders.Models used a fixed subject-level train/validation/test split, with a separate rotating-test K-fold analysis over the full cohort.
Foundation-model adapters: Med3DVLM and SAM-Med3D used frozen pretrained 3D encoders paired with small trainable regression decoders for the same pre→post ASL task.Med3DVLM used 8 seeds and SAM-Med3D used 3 seeds; frozen means the pretrained encoder weights remained fixed.
Evaluation: MAE, SSIM, and PSNR were computed over in-mask voxels on [0, 1] data, using full volumes for 3D models and the middle axial slice for CAE 2D.The held-out test set contained N = 32 subjects; in-house metrics were reported across R=3 independently trained seeds.

4. Results

CAE3D achieved the strongest held-out reconstruction performance among trained-from-scratch models, with statistically significant SSIM and PSNR advantages over all eight comparators. It also showed near-neutral bias and preserved territorial ordering, although it compressed high-response regional ΔCBF.

Primary test-set reconstruction: 0.066 MAE was the lowest among trained-from-scratch models on the held-out test set; CAE3D tied for highest SSIM (0.80) and PSNR (24.0 dB).The SSIM tie was with ResNet 3D, and the PSNR tie was with FNO 3D.
Statistical comparisons: CAE3D’s MAE advantage was significant over seven comparators but not CAE 2D, while its SSIM and PSNR advantages were significant over all eight comparators.The MAE comparison used Holm-adjusted paired Wilcoxon tests; CAE 2D had pHolm = 0.080.
Subject-level R2: 0.35 was CAE3D’s per-subject R2, compared with 0.47 for FNO 3D at comparable MAE values of 0.072 and 0.066, respectively.R2 reflects between-subject variance in mean post-ACZ signal, whereas MAE and SSIM reflect within-subject voxelwise fidelity; the authors do not interpret R2 = 0.35 as individualized CVR recovery.
Agreement and Bland-Altman: −0.003 was CAE3D’s smallest absolute bias, with limits of agreement of ±0.11, comparable to FNO 3D and CAE 2D.Agreement broadly followed reconstruction quality, with stable deterministic models showing near-neutral bias and poorly performing diffusion models showing larger bias and wider limits.
Territory-level ΔCBF: CAE3D preserved territorial ΔCBF ordering but compressed dynamic range in high-response territories, most strongly in the cerebellum and pons.It captured a substantial fraction of the true absolute change, while ResNet 3D tended to overestimate vascular responses.
Five-fold cross-validation: CAE3D remained the top-performing model across five cross-validation partitions, supporting the stability of the fixed held-out ranking.The cross-validation evaluated nine trained-from-scratch 3D models under the same protocol.

5. Discussion

CAE3D’s post-ACZ perfusion synthesis supports further clinical validation, while its territory-level ΔCBF compression suggests more reliable population-level mapping than subject-specific vasodilatory magnitude. The study remains internal validation, requiring external, prospective, and clinically audited evaluation before deployment.

Main findings and clinical interpretation: CAE3D synthesized post-ACZ perfusion maps from pre-ACZ input with reconstruction error motivating further clinical validation.Deterministic conditional autoencoders outperformed diffusion-style variants and frozen-encoder adapters, while errors appeared near mask boundaries and high-flow or low-signal regions.
Main findings and clinical interpretation: Territory-level ΔCBF compression indicates that CAE3D captures population-level pre→post structure more reliably than subject-specific vasodilatory magnitude.
Trustworthiness: uncertainty, agreement, and failure modes: Reported seed variability, bootstrap CIs, and Bland-Altman bias/LoA address uncertainty and agreement beyond correlation alone.Anticipated failure modes include registration or atlas mismatch, unstable ΔCBF at low baseline signal, mask-boundary artifacts, and scanner or cohort shift.
Transparency, explainability, and auditing: Explainability analyses used guided backpropagation for one held-out subject and representational similarity analysis of bottleneck features across 32 test subjects.Clinical-subgroup alignment was not tested because laterality and severity metadata were unavailable.
Limitations and future directions: The study provides internal validation rather than deployment-ready evidence and may not generalize across scanners, ASL sequences, or sites.Future steps include multicenter external validation, prospective blinded clinician-in-the-loop evaluation, physical-unit CBF reporting, range-aware training, and foundation-model fine-tuning.

Appendix A. Reproducibility

The reproducibility appendix documents the software stack, fixed data split, model-selection procedure, and CAE3D training and inference workflow. It also specifies evaluation outputs and supplementary analyses, with regeneration scripts and file-level provenance available in the code repository.

Implementation: Python, PyTorch, and MONAI implemented preprocessing, training, and evaluation, using MONAI UNet / DiffusionModelUNet configurations for in-house models.Architecture details and task-specific recipes are referenced in §3.1–§3.2.
Model selection: Seed 42 fixed the train/validation/test split for all models and was predesignated before training, while checkpoint selection used the highest validation PSNR without test-set involvement.Exact checkpoint filenames and configurations are documented with the code release.
CAE3D procedure: 50 epochs and Adam (10−3) were used in the CAE3D training procedure, retaining the checkpoint when validation PSNR improved.Inference loads the best checkpoint and computes the predicted post-ACZ volume from the pre-ACZ input.
Evaluation: Full-brain MAE/SSIM/PSNR, Bland-Altman agreement, and regional ∆CBF summaries were specified as evaluation outputs.The algorithm preprocesses paired volumes using affine registration to MNI, brain masking, normalization to [0, 1], and pad/crop to a valid grid.
Supplementary reproducibility: Supplementary materials include pairwise statistical tables, residual-diffusion and interpretability analyses, regional evaluations, ablations, failure cases, and foundation-baseline aggregates.Regeneration scripts and file-level provenance are documented in the code repository.

Appendix B. Supplementary figures

Appendix B provides supplementary visualizations of atlas mask unions, cohort-mean territorial ∆CBF, and per-subject qualitative examples. The figures complement corresponding main-text figures without reproducing Figure 4.

Atlas mask unions: Figure S1 shows atlas mask unions for the left hemisphere, right hemisphere, and pooled vascular mask on one middle axial slice.It uses the same slice convention as Figure 5 in the main text.
Territorial ∆CBF: Figure S2 plots cohort-mean ground-truth territorial ∆CBF (%) with one bar per summary region.The figure supports comparison with Table 3 in the main text.
Per-subject examples: Figures S3 and S4 present per-subject qualitative examples, while the paired pre-ACZ, post-ACZ, and voxelwise post-minus-pre maps from Figure 4 are not reproduced.The supplementary figures shift from territory-level summaries to subject-level examples.
Qualitative layout: Figure S4 uses a five-panel layout containing pre, post, prediction, mask, and error images.It complements the best, median, and worst PSNR panels in Figure 6 of the main text.

Appendix C. K-fold sources and foundation baselines

The appendix documents evaluation aggregation, Bland–Altman agreement calculations, foundation-adapter reporting, and fold-level replication procedures. It also records a verified Residual 3D collapse caused by numerical instability in shared DDIM/cosine-schedule sampling code.

Foundation adapter aggregates: Frozen-encoder foundation adapters report mean±std MAE/SSIM/PSNR across eight Med3DVLM seeds and three SAM-Med3D seeds.Both use pre→post ASL perfusion maps with the same full-brain evaluator as CAE3D.
Bland–Altman agreement: Table 6 reports Bland–Altman Bias, standard deviation of paired differences, and 95% limits of agreement, with LoA defined as Bias ± 1.96 × SD.Bias is computed as predicted − ground truth on the normalized [0, 1] scale.
K-fold replication: Five-fold rotation reports mean±std across each model’s five fold-level cohort means for every metric.Fold-level outputs are stored with the released training scripts, and the protocol is reproducible from the code release.
Residual 3D collapse: All three saved Residual 3D seed-42/123/456 checkpoints were distinct, yet converged to a degenerate near-identity solution with predicted residual ≈0.The collapse was excluded from best/second-best ranking and from the territory-level ∆CBF comparison.
Residual 3D collapse: A floating-point artifact made the cosine schedule’s cumulative-product term slightly negative near the tail, generating NaNs that were floored to zero during DDIM sampling.This shared sampling-code issue drove predicted residuals toward zero regardless of the trained weights, and TIPS stabilization measures did not prevent the failure.

LLM Use Disclosure

The manuscript was drafted and revised with assistance from Claude, including prose, implementation and evaluation code, and figure generation; the authors reviewed and validated all content.

LLM Use Disclosure: Claude assisted with drafting and editing prose, model implementation and training/evaluation scripts, and figure generation.The authors state that this use was consistent with the MLHC LLM Use Policy.
LLM Use Disclosure: The authors reviewed and validated all experimental results, figures, and scientific claims and take full responsibility for the manuscript’s accuracy.