PaperlayerView exact revision on arXiv

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:2608.14758v1eess.IVcs.AIcs.CV

TL;DR

Paired pre- and post-acetazolamide ASL is used to assess cerebrovascular reserve in Moyamoya, motivating alternatives when drug administration is constrained. This paper uses a 3D conditional autoencoder to synthesize post-acetazolamide CBF maps from baseline ASL, achieving the lowest held-out MAE among trained-from-scratch models and tying for top SSIM and PSNR.

  • Problem

    Paired pre- and post-acetazolamide ASL is an established approach to assess cerebrovascular reserve in Moyamoya, while practical constraints motivate computational alternatives.

    Source: §1 · p. 2

  • Method

    CAE3D is a deterministic 3D conditional autoencoder that maps a pre-acetazolamide perfusion map to a predicted post-acetazolamide map.

    Source: §3.1 · p. 6

  • Results

    MAE 0.066 was lowest among trained-from-scratch models, while SSIM 0.80 and PSNR 24.0 dB tied for the highest reported values.

    Source: §4 · p. 11

  • Takeaways & Limitations

    The study demonstrates the feasibility of synthesizing post-acetazolamide CBF maps from pre-acetazolamide ASL in patients with Moyamoya.

    Source: §1 · p. 2

    This was internal validation, not deployment-ready evidence, and results may not generalize across scanners, ASL sequences, or sites without external and prospective validation.

    Source: §5 · p. 16

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

Paired pre- and post-acetazolamide ASL maps support CVR assessment in Moyamoya, but the post-challenge scan may be unavailable. This study proposes full-volume 3D synthesis of post-ACZ CBF from baseline ASL, while emphasizing that the retrospective task is not a validated replacement for ACZ imaging.

Clinical motivation

  • Paired pre- and post-ACZ ASL is an established CVR assessment approach, but contraindications or practical barriers can leave clinicians without a key hemodynamic input for bypass planning.The cohort does not establish how prevalent the listed contraindication categories are among patients undergoing Moyamoya evaluation.Source: §1 · p. 2, Figure 1 · p. 3

Study approach

  • The study proposes CAE3D to synthesize post-ACZ CBF maps from pre-ACZ ASL, extending Goyal et al. (2026) from slice-based synthesis to full-volume 3D.The contribution is task-specific adaptation and comparative evaluation, not a novel backbone architecture; eleven models were evaluated in total, including CAE3D.Source: §1 · p. 2

Contributions

  • CAE3D achieved the lowest held-out MAE among nine in-house models, near-zero Bland–Altman bias, and lower territory-level ∆CBF error than evaluated diffusion baselines.The reported CAE3D-versus-CAE 2D comparison favored CAE3D on reported metrics, but the methods operate on different spatial domains.Source: §1 · p. 2

Study approach

  • The illustrated synthesis task predicts an independently normalized post-ACZ map from pre-ACZ input, but represents a future workflow motivation rather than a clinically validated replacement for ACZ imaging.The evaluation concerns retrospective synthesis in patients who completed the two-scan protocol.Source: §1 · p. 2, Figure 2 · p. 3

Generalizable insights

  • In this single-center evaluation, deterministic 3D autoencoders outperformed tested diffusion implementations, multimodal evaluation exposed limits beyond global metrics, and frozen-encoder adaptations did not outperform task-specific training.The authors present these as observations from the evaluated models and protocols, not as findings that generalize beyond them.Source: §1 · p. 3

2. Related Work

The related work frames baseline-only post-ACZ CBF synthesis as a normalized image-to-image prediction task, motivated by ASL-based CVR assessment and prior synthesis research. It also clarifies that the learned mapping reflects cohort-level associations, not recovery of an individual patient’s CVR.

Background: ASL perfusion imaging

  • ASL provides voxelwise CBF maps, and the pre-to-post-ACZ CBF change encodes the regional vasodilatory response used in CVR assessment.CAE3D predicts post-ACZ maps from pre-ACZ maps in independently normalized [0, 1] voxel space rather than physical CBF units, which affects how normalized-space ∆CBF relates to physiological CVR.Source: §2 · pp. 3–4

Post-ACZ perfusion synthesis from baseline ASL perfusion maps

  • Goyal et al. (2026) studied 2D middle-slice post-ACZ prediction, whereas this study evaluates full-volume 3D prediction with 2D and 3D baselines plus agreement and territory-level analyses.Both studies predict synthetic post-ACZ maps from pre-ACZ maps alone.Source: §2 · p. 5

Physiological basis and scope of the learned mapping

  • In Moyamoya, chronic occlusion, collateral recruitment, and territorial steal can make resting perfusion statistically associated with reserve, but CAE3D learns cohort-level correlations rather than subject-specific CVR recovery.The authors do not claim that visually identical baseline maps imply identical CVR; such cases are especially difficult for baseline-only prediction.Source: §2 · p. 5

Broader perfusion and multimodal synthesis

  • Perfusion-target synthesis has been explored in MRI–PET settings, and this study includes frozen-encoder comparators Med3DVLM (Xin et al., 2025) and SAM-Med3D (Wang et al., 2023).Adaptation details are provided in §3.3.Source: §2 · p. 5

3D modeling, evaluation rigor, and clinical interpretability

  • Diffusion and cold-diffusion models have been applied to medical image synthesis, while this study uses a unified 3D comparison with atlas-based regional summaries alongside global image-quality metrics.The cited diffusion work includes Ho et al. (2020) and Bansal et al. (2023).Source: §2 · p. 5

3. Methods

The study evaluates supervised synthesis of post-ACZ ASL perfusion maps from pre-ACZ maps using a shared held-out protocol and a deterministic 3D conditional autoencoder alongside comparison models. The cohort comprises patients with paired pre- and post-ACZ scans, so it does not represent patients unable to receive ACZ.

3.2. Baselines and comparison models

  • The primary comparison evaluates nine in-house models and two separately evaluated foundation-encoder adapters; CAE 2D is a middle-slice comparator, unlike the full-volume 3D models.The in-house baselines include deterministic, diffusion-style, and FNO 3D models; CAE3D-ES isolates the effect of early stopping from the full-schedule CAE3D.Source: §3 · pp. 5–6, §3.2 · pp. 8–9

3.1. Proposed method: 3D conditional autoencoder and diffusion variants

  • CAE3D deterministically maps a pre-ACZ volume to a post-ACZ volume with a 3D MONAI encoder-decoder and a combined L1+SSIM objective.It uses skip connections, three downsampling levels, and a full 50-epoch schedule selected by validation PSNR; primary results use an unweighted full-brain loss.Source: §3.1 · p. 6
  • The study used 251 paired-scan patients, split into 188 training, 31 validation, and 32 held-out test subjects, with the test set reserved for final reporting.The fixed subject-level split used random seed 42; in-house models were trained with three seeds, and a separate five-fold analysis was performed.Source: §3.1 · pp. 6–8
  • Volumes were affinely registered to MNI152 2 mm, resampled to 91 × 109 × 91, masked, and independently min–max normalized to [0, 1] per scan.One failed registration was excluded; the independent per-volume scaling partly absorbs absolute between-scan intensity differences.Source: §3.1 · p. 8

3.3. Foundation model adapters: Med3DVLM and SAM-Med3D

  • Med3DVLM and SAM-Med3D adapt frozen pretrained 3D encoders to the same pre-to-post ASL task by training lightweight regression decoders.The adapters use separate seed counts and an aggregation protocol distinct from the in-house models.Source: §3.3 · pp. 9–10

3.4. Metrics

  • Evaluation used in-mask MAE, 3D SSIM, PSNR, and R2, with regional delta-CBF, Bland–Altman agreement, and paired Wilcoxon tests with Holm adjustment as additional analyses.Regional delta-CBF was computed from territory-mean intensities across predefined vascular territories; the authors note it is not a formal gas-reactivity CVR measure.Source: §3.4 · p. 10, Table 1 · p. 11

4. Results

CAE3D achieved the strongest overall held-out reconstruction performance among trained-from-scratch models, but subject-level and territory-level findings show limitations in capturing individualized vasodilatory response.

Primary test-set reconstruction

  • MAE 0.066 was lowest among trained-from-scratch models; SSIM 0.80 tied ResNet 3D, and PSNR 24.0 dB tied FNO 3D; SSIM and PSNR advantages were significant over all eight comparators.The MAE advantage was significant over seven of eight comparators, but not the 2D CAE 2D baseline (pHolm = 0.080).Source: §4 · pp. 11–12

Seed-to-seed stability

  • Residual 3D collapsed to a near-zero-residual solution across three distinct seeds, while Patch 3D’s near-zero variance reflected a shared pretrained VAE rather than seed independence.Residual 3D is reported for transparency, not as evidence of stability; Table 7’s fold replication retrained the VAE and showed nonzero variance.Source: §4 · p. 11, Table 2 · p. 12

Interpreting subject-level R2

  • R2 0.35 for CAE3D was below FNO 3D’s 0.47 despite comparable MAE (0.066 vs. 0.072), reflecting weaker reproduction of between-subject mean-signal variance than voxelwise fidelity.The results do not establish recovery of individualized CVR; they indicate population-level structure is captured more reliably than subject-specific vasodilatory variation.Source: §4 · pp. 12–13

Agreement and Bland-Altman

  • Bias −0.003 was CAE3D’s smallest absolute bias, with limits of agreement ±0.11, comparable to FNO 3D and CAE 2D.A small mean difference does not guarantee narrow subject-level spread, so bias should be interpreted alongside limits of agreement and reconstruction error.Source: §4 · p. 13

Territory-level ∆CBF (exploratory)

  • Approximately 10-fold smaller predicted ∆CBF in the cerebellum and pons showed CAE3D compressed high-response dynamic range, although it retained a substantial fraction of true absolute change.Absolute ∆CBF magnitudes are not calibrated CVR values; conversion to physical units was not performed.Source: §4 · p. 13

Five-fold cross-validation

  • Across five-fold cross-validation partitions, CAE3D remained the top-performing trained-from-scratch 3D model, although the ordering of other models varied somewhat.The rotating protocol trained on K−2 folds, validated on one fold, and tested on one fold.Source: §4 · p. 14

5. Discussion

CAE3D showed retrospective feasibility for synthesizing post-ACZ perfusion maps, but its predictions capture population-level structure more reliably than subject-specific vasodilatory magnitude. Generalizability and clinical use remain unestablished, especially for patients who did not undergo ACZ challenge.

Main findings and clinical interpretation

  • CAE3D’s reconstruction error motivates further clinical validation, though it is not directly comparable to ASL scan-rescan repeatability; deterministic conditional autoencoders outperformed diffusion-style variants and frozen-encoder adapters.These findings support retrospective synthesis feasibility, not clinical readiness.Source: §5 · p. 14
  • Modest per-subject R2 (0.35) and territory-level ∆CBF compression indicate that CAE3D captures population-level structure more reliably than subject-specific vasodilatory magnitude.Compression may partly reflect the L1+SSIM objective, with the largest percent-scale compression in the cerebellum and pons/medulla.Source: §5 · pp. 14–16

Trustworthiness: uncertainty, agreement, and failure modes

  • The study reported seed variability, bootstrap CIs, and Bland-Altman bias/LoA; anticipated failure modes include registration or atlas mismatch, low-signal ∆CBF instability, boundary artifacts, and scanner or cohort shift.Future evaluation should pair outputs with predefined quality checks and human review.Source: §5 · p. 15

Transparency, explainability, and auditing

  • CAE3D’s explainability analyses included guided backpropagation for one held-out subject and representational similarity analysis across 32 test subjects, but clinical subgroup alignment was not tested.Synthesized maps are research outputs, not decision-support tools or substitutes for clinical imaging before prospective validation; data sharing requires de-identification and drift-monitoring governance.Source: Figure 7 · p. 15, §5 · p. 16

Limitations

  • Validation was internal, and performance in ACZ-contraindicated patients is unknown because none were represented in training or testing.Affine registration and atlas readouts may degrade outside template anatomy; [0, 1] normalization leaves MAE unitless, and physical-scale inversion was not performed. Voxel/cluster permutation tests, confounder-adjusted analyses, calibrated uncertainty maps, and prospective workflow validation were also absent.Source: §5 · p. 16

Future directions

  • Next steps include multicenter external validation and prospective blinded clinician-in-the-loop evaluation, alongside physical-unit error reporting, range-aware training, and foundation-model fine-tuning.Territory-aware models may improve regional calibration at the cost of less data per territory and added inference complexity; flow matching is another alternative, and synthesis needs separate evaluation in intracranial atherosclerosis and sickle cell disease.Source: §5 · p. 16

Appendix A. Reproducibility

The appendix documents the fixed-split, validation-based checkpoint selection used for in-house models and provides the CAE3D training and inference procedure. It also describes the implementation framework, evaluation measures, and supporting reproducibility materials.

Appendix A. Reproducibility

  • Seed 42 was predesignated as each in-house model’s primary instance, and its highest-validation-PSNR checkpoint was retained without test-set involvement.The split was fixed using random seed 42; exact checkpoint filenames and configurations are documented in the code release.Source: Appendix A · p. 20, Table 5 · p. 20
  • Preprocessing, training, and evaluation used Python with PyTorch and MONAI, including MONAI UNet and DiffusionModelUNet implementations for in-house models.The code repository documents the training configurations and file-level provenance.Source: Appendix A · pp. 20–21
  • CAE3D preprocesses paired volumes with MNI affine registration, brain masking, [0, 1] normalization, and pad/crop, then trains for 50 epochs using Adam at 10^-3.Inference loads the retained checkpoint to predict post-ACZ volume from pre-ACZ input; evaluation includes full-brain MAE, SSIM, PSNR, Bland–Altman agreement, and regional ΔCBF summaries.Source: Appendix A · p. 21
  • Supplementary materials provide adjusted pairwise tests, ablation and regional analyses, failure-case panels, and per-territory and foundation-baseline results.The appendix also reports bootstrap resampling with B=2000.Source: Appendix A · p. 21

Appendix B. Supplementary figures

Appendix B presents supplementary atlas-mask and regional ∆CBF figures, followed by per-subject qualitative examples that complement main-text figures.

Appendix B. Supplementary figures

  • 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 main-text Figure 5.Source: Appendix B · p. 21
  • Figure S2 plots cohort-mean ground-truth territorial ∆CBF (%) with one bar per summary region, for comparison with main-text Table 3.Source: Appendix B · p. 22
  • Figures S3 and S4 shift to per-subject qualitative examples; paired pre-ACZ, post-ACZ, and voxelwise post-minus-pre maps for an example appear in main-text Figure 4, not reproduced here.Source: Appendix B · p. 22
  • Figure S4 presents a five-panel qualitative layout of pre-ACZ, post-ACZ, prediction, mask, and error, complementing the best-, median-, and worst-PSNR panels in main-text Figure 6.Source: Appendix B · p. 22

Appendix C. K-fold sources and foundation baselines

The appendix documents agreement, K-fold, foundation-adapter, and territory-level reporting, while clarifying collapse and variance caveats for diffusion-style baselines. It also specifies how the reported aggregates were computed and which checks were performed.

Foundation adapter aggregates

  • Foundation-adapter MAE/SSIM/PSNR aggregates are mean±std over eight Med3DVLM seeds and three SAM-Med3D seeds, using pre→post ASL maps and the full-brain evaluator.Their R2 and Bland–Altman bias/LoA are seed-42 checkpoint point estimates, and bootstrap bias CIs were not computed for these rows. Table 8 reports the adapter aggregates alongside the foundation rows in Table 1.Source: Appendix C · pp. 22–24, Table 8 · p. 24

Bland–Altman detail

  • Table 6 reports the standard deviation of paired differences underlying each model’s limits of agreement, where LoA = Bias ± 1.96 × SD.It reproduces per-model Bias and LoA already reported in main-text Table 1.Source: Appendix C · p. 23, Table 6 · p. 23

K-fold replication

  • Table 7 summarizes five-fold cohort means as mean±std across five test-fold values per metric; the collapsed Residual 3D tips row is excluded from best/second-best ranking.Fold-level outputs and scripts support reproduction, with optional per-fold table checks before exporting LaTeX rows. The caption says the collapse mechanism is expected to affect K-fold instances but was not independently verified for each one against the zero-residual baseline.Source: Table 7 · p. 23, Appendix C · pp. 23–24

Per-territory metrics

  • Table 9 reports CAE3D cohort mean±std per atlas territory on the held-out test set (N=32), with metrics computed within each atlas mask on normalized data.Territory names follow the Liu et al. (2023) atlas registered to MNI 2 mm.Source: Appendix C · p. 24, Table 9 · p. 24

Residual 3D tips collapse

  • Residual 3D tips collapsed to approximately zero predicted residual across three distinct checkpoints, and a zero-residual baseline reproduced Table 2 metrics to floating-point precision.The shared cosine noise schedule can produce a slightly negative cumulative-product term near its tail; resulting NaNs are floored to zero during DDIM sampling. TIPS stabilization did not prevent the failure. The mechanism was verified for the fixed-test three-seed instance, but not independently for each K-fold checkpoint.Source: Appendix C · pp. 24–25

Patch 3D variance caveat

  • Patch 3D’s fixed-test three-seed run showed near-zero variance because the same pretrained VAE was reused, whereas K-fold retrained a fold-specific VAE and showed nonzero variance.The shared VAE reconstruction dominated decoded full-volume output at the patch/stride configuration, obscuring differences between diffusion-stage seeds at the reported precision; this caveat is specific to the fixed-test run.Source: Appendix C · p. 25

LLM Use Disclosure

Claude (Anthropic) assisted with manuscript prose and code, while the authors reviewed and validated the scientific content and take full responsibility for it.

LLM Use Disclosure

  • Claude (Anthropic) assisted with drafting and editing prose and code for implementation, training/evaluation scripts, and figure generation; authors reviewed and validated results, figures, and claims and take full responsibility.The assistance was consistent with the MLHC LLM Use Policy.Source: LLM Use Disclosure · p. 25

Related papers

Loading 2608.14758v1…