List of Papers Browse by Subject Areas Author List
Abstract
Large and demographically balanced datasets are essential for reliable neuroimaging biomarkers. Full-resolution 3D brain MRI synthesis can support data augmentation in this setting, but existing approaches either incur prohibitive computational cost at volumetric scale or rely on lossy latent compression that may compromise anatomical detail. As a result, practical 3D generative augmentation often requires specialized compute infrastructure.
We propose WaveDiT, a conditional flow matching framework operating in the coefficient space of a 3D Haar Discrete Wavelet Transform. The model combines factorized spatio-depth attention with band-wise heteroscedastic uncertainty modeling derived from higher-order wavelet statistics. Predicted log-variance is integrated directly into both the flow objective and conditioning pathway, enabling adaptive precision consistent with the heavy-tailed and input-dependent variance structure of anatomical detail. This formulation supports full-resolution 3D synthesis under practical memory and time constraints on a single modern GPU.
Evaluation on a multi-site cohort demonstrates improved alignment between generated and real MRI distributions, together with enhanced downstream brain age prediction and region-level anatomical agreement relative to diffusion, latent, and wavelet-based baselines.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4757_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/sisinflab/WaveDiT
Link to the Dataset(s)
N/A
BibTex
@InProceedings{DanDan_WaveDiT_MICCAI2026,
author = { Danese, Danilo AND Lombardi, Angela AND Fasano, Giuseppe AND Attimonelli, Matteo AND Di Noia, Tommaso},
title = { { WaveDiT: Distribution-Aware Wavelet Flow Matching for Efficient 3D Brain MRI Synthesis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16890},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper studies 3D medical image generation with a wavelet-based DiT under a flow matching formulation. The authors make three main contributions. First, they introduce a 3D WaveDiT framework that models volumetric data in the wavelet domain. The design is motivated by an empirical analysis showing that different wavelet bands exhibit highly different energy distributions and statistical evolution during flow matching. Based on this observation, the method is designed to handle low-frequency and high-frequency components in a more adaptive way. Second, following this, the paper proposes an uncertainty-aware Morpheus module, to estimate band-wise uncertainty. The module conditions on frequency statistical features of the current state xt to adaptively adapt their Bayesian heteroscedastic objective. Third, to further make transformer modeling practical for 3D volumes, the paper adopts an efficient slice attention design. In shallow layers, they use window attention to capture local high-frequency structures, while deeper layers combine intra-slice attention and cross-slice attention to preserve both details and anatomical consistency across slices. They evaluate their model, termed WaveDiT-CFM, on image generation, brain age prediction using the generated volumes, and ROI analysis, and the results consistently demonstrate its strong performance.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper is generally well written and easy to follow. The motivation is clear, and the overall framework is reasonably presented. 2.The paper studies a meaningful problem in 3D medical image generation, i. e. , how to generate high-quality 3D volumes efficiently while keeping the training complexity manageable. The overall design, combining wavelet modeling, uncertainty-aware optimization, and efficient attention, is thoughtful and well motivated. 3.The method contains some interesting technical designs. In particular, the analysis of wavelet-band statistics during flow matching provides reasonable support for the band-aware design, and the uncertainty-aware conditioning is also a novel part of the framework. 4.The experimental evaluation is fairly comprehensive. In addition to image generation results, the paper also includes brain age prediction and ROI analysis on generated volumes, which makes the evaluation more convincing.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The overall framework contains several components, so it is sometimes not easy to fully understand the contribution of each individual design choice. In addition, the introduction of several advanced modules makes the paper a bit harder to follow in some parts. These aspects could be further clarified with more detailed analysis if possible. 2.The uncertainty-aware formulation is interesting, but its role could be better discussed. In particular, it would be helpful to further explain how the learned uncertainty contributes to the final improvement.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper is well written and studies a meaningful problem in 3D medical image generation, namely how to generate high-quality 3D volumes efficiently while keeping the training complexity manageable. The overall framework is technically sound, and the combination of wavelet modeling, uncertainty-aware optimization, and efficient attention design is thoughtful and well motivated. The experimental evaluation is also fairly comprehensive, covering not only image generation quality, but also brain age prediction and ROI analysis on the generated volumes. My main concerns are relatively minor. Since the framework contains several components, it is sometimes difficult to fully disentangle the contribution of each design choice, and some parts could be explained more clearly. I also think the role of the uncertainty modeling could be further analyzed. Overall, I am positive about the paper and think it is worthy of acceptance.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The paper presents WaveDiT, a generative framework for efficient full-resolution 3D brain MRI synthesis that operates in the coefficient space of a single-level 3D Haar discrete wavelet transform (DWT). The central motivation is that existing 3D MRI generative approaches face a trade-off between fidelity and computational efficiency: pixel-space diffusion models are costly for volumetric data, while latent diffusion methods reduce cost at the expense of lossy compression that may compromise anatomical detail. Wavelet-domain methods avoid learned compression artifacts, but previous works generally treat all wavelet subbands uniformly, despite their markedly different statistics.
The main methodological contribution has two tightly coupled parts:
1.Wavelet-domain conditional flow matching for 3D MRI synthesis. The model uses a conditional flow matching formulation in wavelet space rather than diffusion in voxel space or latent space. The use of wavelets preserves invertibility while reducing spatial resolution, allowing the model to synthesize full-resolution 3D MRI under practical memory and runtime constraints. The paper further extends the Hourglass Diffusion Transformer (HDiT) design to volumetric data by introducing a computationally efficient factorized spatio-depth attention mechanism. The volume is processed as a sequence of 2D slices for spatial modeling, and inter-slice consistency is restored by explicit depth attention. This avoids the prohibitive cost of full 3D global self-attention while preserving volumetric coherence. 2.Morpheus: state-aware uncertainty modeling in the wavelet domain. A key novelty is the observation that wavelet subbands exhibit strongly different and time-varying distributions along the generative trajectory: the approximation band remains relatively close to Gaussian, whereas high-frequency bands are sparse, heavy-tailed, and highly heteroscedastic. To exploit this, the authors propose Morpheus, a lightweight auxiliary network that predicts band-wise log-variance from higher-order wavelet statistics of the current noisy sample state, including mean, standard deviation, max absolute value, L2 norm, skewness, and kurtosis. The predicted uncertainty is used in two ways: 1) to define a Bayesian heteroscedastic flow objective that adaptively reweights errors according to predicted precision, instead of applying uniform MSE across all bands and locations; 2) to provide frequency-aware conditioning to the backbone, enabling the model to adapt its predictions to the current reliability and complexity of each wavelet band during both training and inference.
Experimentally, the paper shows that this combination improves both efficiency and quality. WaveDiT generates full 3D MRI in about 1 second with 10 steps on a single modern GPU, while outperforming diffusion, latent diffusion, and prior wavelet/flow baselines in:
- global distribution metrics (FID, MMD),
- downstream brain age prediction when synthetic data are used for augmentation,
- and ROI-level anatomical fidelity measured by region-wise intensity error, KL divergence, and Dice overlap.
Overall, the paper’s main contribution is a distribution-aware, wavelet-domain flow matching framework for practical and anatomically faithful 3D brain MRI synthesis, with explicit modeling of the heteroscedastic and heavy-tailed structure of wavelet coefficients.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Well-justified use of the wavelet domain. The paper makes a convincing case for why wavelets are attractive for 3D MRI synthesis: they reduce spatial dimensionality while preserving invertibility and avoiding learned compression artifacts. This is especially relevant in neuroimaging, where subtle cortical and anatomical details matter. The empirical analysis showing that the approximation band contains most energy, while high-frequency bands are sparse and heavy-tailed, provides a solid statistical basis for the proposed design. 2.Efficient architecture for full-resolution 3D synthesis. The proposed factorized spatio-depth attention is practically important. Full 3D self-attention at this resolution would be infeasible, and the paper provides a concrete complexity analysis demonstrating large computational savings. The design—local 2D neighborhood attention at higher resolution and deeper factorized spatial/depth attention—appears sensible and aligned with volumetric anatomy. 3.Strong practical efficiency. A notable strength is that the method supports full-resolution 3D generation on a single H100 GPU, with training in about 26 hours and inference in about 1 second for 10 steps. In this problem domain, practical accessibility is a meaningful contribution, since many high-quality 3D generative methods remain difficult to use outside highly specialized compute settings. 4.Comprehensive evaluation beyond standard generative metrics. The evaluation is stronger than many generative imaging papers because it does not rely only on FID-like metrics. The authors follow a multi-level evaluation protocol:
-
global distributional alignment (FID, MMD, MS-SSIM),
- a clinically relevant downstream task (brain age prediction),
- and ROI-level anatomical analysis using segmentation-based metrics. This is important because global image metrics can miss anatomically meaningful errors. The ROI and downstream analyses make the empirical case substantially more convincing.
-
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The uncertainty model is relatively coarse. Morpheus predicts band-wise log-variance from global summary statistics per channel rather than finer-grained spatial uncertainty. Given that heteroscedasticity is argued to vary strongly across space, especially near tissue boundaries, a spatially resolved uncertainty model might be more faithful to the stated motivation. The current design is elegant and lightweight, but potentially limited. The paper does not deeply analyze whether band-wise uncertainty is sufficient compared with more localized alternatives. 2.Evaluation remains limited to one modality and one task setting. The study focuses only on T1-weighted brain MRI. This is a reasonable starting point, but it limits generalizability. It is unclear whether the gains would transfer to:
- other MRI contrasts,
- pathological populations,
- or other 3D medical imaging modalities such as CT. The authors acknowledge this limitation.
3.Potential concern about downstream evaluation interpretation. The brain age prediction improvement is interesting, but the causal interpretation should be made carefully. Better augmentation performance does not necessarily imply that the generated images are clinically realistic in all relevant respects; it may partly reflect task-specific regularization or data balancing effects. The paper is reasonably cautious, but this point is worth emphasizing.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses an important and practical problem: efficient, anatomically faithful full-resolution 3D brain MRI synthesis. Its strongest aspects are:
- a well-motivated wavelet-domain formulation,
- a novel state-aware uncertainty module (Morpheus) that is integrated into both the loss and conditioning,
- an efficient factorized spatio-depth transformer architecture,
- and a strong empirical evaluation that goes beyond standard image-generation metrics by including brain age prediction and ROI-level anatomical analysis.
I found the paper technically solid and empirically convincing. The method appears to deliver a meaningful improvement in the quality-efficiency trade-off, especially under a low-step sampling regime and practical single-GPU constraints. The ablations also support the central design choices.
The main reasons I do not score it higher are:
- some aspects are incremental relative to recent wavelet-based MRI generation papers, especially WDM [Friedrich et al., 2024] and FlowLet [Danese et al., 2025];
- several building blocks are adapted from prior work, so the novelty is more in the integration and domain-specific design than in a fundamentally new generative principle;
- the validation is currently limited to T1-weighted brain MRI, without expert human assessment of realism.
Overall, however, I think the paper makes a useful and timely contribution to medical image generation, and the empirical evidence is strong enough to justify acceptance.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This paper proposes WaveDiT, a conditional flow-matching framework for efficient full-resolution 3D brain MRI synthesis in the wavelet domain. The method combines a 3D Haar DWT representation, a factorized spatio-depth DiT backbone, and a Morpheus module for state-aware heteroscedastic uncertainty modeling. The goal is to improve full-volume 3D MRI synthesis under practical memory and inference constraints.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper addresses an important problem in scalable 3D brain MRI synthesis, where voxel-space diffusion is expensive and latent compression may lose anatomical detail. The wavelet-domain formulation is well motivated, and the analysis of band-wise statistics provides a reasonable basis for the Morpheus uncertainty module. The paper also includes many relevant baselines, ablations, efficiency comparisons, and downstream evaluations such as brain age prediction and ROI-level anatomical analysis.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The main weakness is the evaluation protocol. Although the paper is framed as a 3D medical image synthesis method, it does not report commonly used image synthesis metrics such as PSNR, SSIM, or LPIPS, making it difficult to assess voxel-level fidelity and perceptual image quality. The authors argue that global volumetric metrics can be biased by background voxels, but this does not fully justify omitting foreground-masked or ROI-based PSNR/SSIM/LPIPS comparisons. In addition, the downstream brain age prediction result may not directly reflect synthesis fidelity, since improved BAP performance could come from distributional or demographic augmentation rather than anatomically accurate image generation. Finally, the model is still relatively large, around 142M parameters, and requires an H100 for training, which somewhat weakens the claim of broad accessibility despite its faster inference.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This work presents a technically interesting and reasonably well-motivated framework for efficient 3D brain MRI synthesis. The wavelet-domain modeling, factorized spatio-depth attention, and heteroscedastic uncertainty module are coherent, and the comparisons with diffusion, latent, and flow-based baselines are fairly strong. However, I am not fully convinced by the evaluation protocol. For a medical image synthesis paper, the absence of standard image-quality metrics such as PSNR, SSIM, and LPIPS is a significant limitation, even if additional distributional and downstream metrics are provided. Overall, the method is promising, but the evaluation should be more complete and the claims about synthesis quality should be better supported.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank all reviewers for their thoughtful and constructive feedback.
R2 (W1) “Several components … difficult to disentangle individual contributions.” Each component corresponds to a distinct design axis isolated in the ablations: Voxel-space (representation), w/o Morpheus (uncertainty), RFM/OTFM/CFM (flow objective). The architecture mirrors two distinct observations: local HF structure (neighborhood attention) and cross-slice consistency (spatio-depth attention). We will make this mapping explicit.
(W2) “Role of uncertainty.” Morpheus contributes through both loss reweighting (Eq.1) and conditioning, introducing band-wise precision aligned with the heteroscedastic statistics of wavelet subbands. Its removal degrades global (FID, MMD), downstream (BAP), and ROI metrics simultaneously (Tables 2b, 3), indicating a structural rather than cosmetic contribution.
R3 (W1) “Band-wise uncertainty is coarse.” The dominant heteroscedasticity is between bands (Table 1), spanning up to 90x in kurtosis, which motivates a band-wise formulation. Dense voxel-wise uncertainty at 224^3 would require a decoder comparable to the backbone, undermining efficiency. Spatial selectivity is nonetheless captured implicitly through:
- conditioning on higher-order moments,
- local neighborhood attention focusing on high-frequency regions,
- AdaRMSNorm modulation, which translates band-wise signals into spatially adaptive activations.
(W2) “Evaluation limited to T1w brain MRI.” We agree, and this is explicitly acknowledged in Sec. 4.T1 is the standard benchmark in the existing 3D MRI generative literature, enabling direct comparison with prior work. We will extend the framework to other modalities.
(W3) “BAP gains may reflect distributional effects rather than anatomical realism.” The ROI metrics in Table 3 are designed for exactly this concern. Unlike BAP, iMAE, KLD, and Dice are computed on independently segmented regions and are insensitive to demographic balancing. WaveDiT-CFM is best across all ROI metrics together with BAP.
R4 (W1) “Absence of PSNR/SSIM/LPIPS is a limitation.” We appreciate this point and clarify our methodological position. WaveDiT addresses age-conditional synthesis, not reconstruction: each sample is drawn from p(x|age) with no one-to-one correspondence to any real volume, so paired PSNR/SSIM/LPIPS, are not well-defined in our setting. The closest paired analogue already in Table 2 is intra-set MS-SSIM, designed precisely to avoid this issue.
To engage the concern directly, we conducted an intra-cohort, age-matched study over 500 within-cohort pairs. The real reference is PSNR 15.93 ± 1.48 dB, SSIM 0.715 ± 0.030, LPIPS 0.204 ± 0.035.The aim is not to maximise these metrics but to match the real cohort’s own intra-similarity distribution, so deviations acquire a diagnostic interpretation: positive ΔPSNR with negative ΔLPIPS signals mode collapse (pairs too self-similar); negative ΔPSNR with stable SSIM/LPIPS points to additive noise or artefacts; strongly negative ΔLPIPS with otherwise normal PSNR/SSIM indicates perceptual collapse.
WaveDiT-CFM is closest to the real reference on all three metrics: ΔPSNR = +0.76 dB, ΔSSIM = +0.004, ΔLPIPS = −0.020.By contrast, MOTFM exhibits mode collapse (ΔPSNR = +5.10, ΔLPIPS = −0.077), MD homogeneous collapse (ΔPSNR = +4.13, ΔLPIPS = −0.044), and MLDM perceptual collapse (ΔLPIPS = −0.105) despite a “normal” PSNR/SSIM.
(W2) “BAP improvement may not reflect synthesis fidelity.” See R3(W3): ROI metrics (Table 3) measure anatomical consistency independently of demographic effects, and WaveDiT-CFM leads across all of them.
(W3) “142M parameters and H100 weaken accessibility.” WaveDiT runs in ~12 GB VRAM at batch size 1 (e.g., 4060ti). The H100 was used only for hardware parity across baselines, since WDM, MLDM, and BS exceed 24-48 GB at full resolution. At 142M parameters WaveDiT is the lightest full-resolution 3D MRI generator among the compared models.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
This paper presents a technically interesting and well-motivated framework for efficient 3D brain MRI synthesis. The problem is meaningful, and the proposed combination of an HDiT backbone with the Morpheus state-aware uncertainty scheduler is promising. Also, the paper is well-written, and the evaluation is fairly comprehensive. The main concerns are the absence of standard image-quality metrics, such as PSNR, and the need for a more detailed analysis of the individual model components. The authors should carefully address these points, along with the other concerns raised by the reviewers, in the final version.