List of Papers Browse by Subject Areas Author List
Abstract
Latent diffusion models (LDMs) have recently achieved strong
performance in 3D medical image synthesis. However, modalities like cine cardiac MRI (CMR), representing a temporally synchronized 3D volume across the cardiac cycle, add an additional dimension that most generative approaches do not model directly. Instead, they factorize space and time or enforce temporal consistency through auxiliary mechanisms such as anatomical masks. Such strategies introduce structural biases that may limit global context integration and lead to subtle spatiotemporal discontinuities or physiologically inconsistent cardiac dynamics. We investigate whether a unified 4D generative model can learn continuous cardiac dynamics without architectural factorization. We propose CardioDiT, a 3D+t latent diffusion framework for short-axis cine CMR synthesis based on diffusion transformers. A spatiotemporal VQ-VAE encodes 2D+t slices into compact latents, which a diffusion transformer then models jointly as complete 3D+t volumes, coupling space and time throughout the generative process. We evaluate CardioDiT on public CMR datasets and a larger private cohort, comparing it to baselines with progressively stronger spatiotemporal coupling. Results show improved inter-slice consistency, temporally coherent motion, and realistic cardiac function distributions, suggesting that explicit 4D modeling with a diffusion transformer provides a principled foundation for spatiotemporal cardiac image synthesis. Code, configurations and models trained on public data are available at https://github.com/Cardio-AI/cardiodit.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5984_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Cardio-AI/cardiodit<
Link to the Dataset(s)
M&Ms-2: https://www.ub.edu/mnms-2/
ACDC: https://www.creatis.insa-lyon.fr/Challenge/acdc/databases.html
BibTex
@InProceedings{SeyMar_CardioDiT_MICCAI2026,
author = { Seyfarth, Marvin AND Müller, Sarah Kaye AND Ghanaat, Arman AND Ayx, Isabelle AND Fastenrath, Fabian AND Wild, Philipp AND Hertel, Alexander AND Papavassiliu, Theano AND Dar, Salman Ul Hassan AND Engelhardt, Sandy},
title = { { CardioDiT: Latent Diffusion Transformers for 4D Cardiac 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 explores whether a unified 4D generative model can learn continuous cardiac dynamics without explicit architecture decomposition. The authors propose CardioDiT, a short-axis cinematic CMR synthesis framework based on diffusion Transformer, employing a full 4D latent diffusion modeling approach. Experimental results demonstrate that this method improves inter-slice consistency, temporal motion coherence, and realism of cardiac function distribution.
- 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 evaluation indicators are relatively comprehensive, covering commonly used qualitative and quantitative CMR volume assessment indicators, and further considering physiological rationality, planar anatomical consistency, geometric consistency and scalability, which is commendable.
- 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 innovations are not clearly articulated. When applying 3D+t data from CMR to a diffusion transformer architecture, the mechanism by which this method handles spatiotemporal discontinuities and physiological inconsistencies is not clearly explained. It is recommended that the authors further clarify the design principles and innovative aspects of this method in addressing these key issues.
2.The authors point out that existing research enforces temporal consistency by decomposing spatial and temporal dimensions or introducing auxiliary mechanisms. This strategy introduces structural biases, thereby limiting the integration of the global context and leading to subtle spatiotemporal discontinuities or physiologically inconsistent cardiac dynamics. However, the experimental section does not directly compare this method with representative methods in references [11, 17], nor does it specifically analyze and verify the structural biases in these methods.
3.Regarding the choice of a phased modeling strategy: Why did the authors choose to first train a 3D VQ-GAN using 2D+t, then slice and encode the complete 3D+t volume data along the D direction, encoding each slice as a latent label, and finally stacking them to form a complete 3D+t latent volume data? Why not directly use the end-to-end approach of 3D+t to learn the latent representation and train it jointly with the 4D diffusion transformer? Please explain the rationale for this design choice.
4.The description of VQ-GAN is too brief, lacking explanation of the loss function, constraints, and specific strategies during training, affecting the understanding of the reproducibility of the entire framework.
5.Reference [15] is also based on a diffusion architecture, but this paper chooses to compare it with the 3D U-Net architecture. The basis for this choice is not clear enough. Furthermore, this paper is similar to reference [15] in its starting point and implementation method. It is recommended that the authors clearly explain the similarities and differences between this paper and reference [15] to highlight its uniqueness.
6.The comparison method in Figure 2 does not clearly indicate whether the experiment uses a public or private dataset, nor does it show the results of real images, affecting the completeness and credibility of the comparison.
7.None of the charts clearly indicate the data source (public or private datasets), nor do they explain the specific purpose of datasets such as ACDC (for training or evaluation), affecting the transparency and reproducibility of the experimental setup.
8.The description of the ablation experimental setup in Table 2 is insufficient, especially regarding the specific changes in model capacity from S to L (e. g. , number of layers, hidden layer dimension, number of parameters), which hinders the reader’s understanding of the impact of model size on performance.
9.For common applications of medical image synthesis (e. g. , data augmentation and privacy protection), it is recommended to supplement the discussion section with relevant explanations to clarify the applicability and potential value of this method in these practical applications.
10.In the discussion section, the authors mention that the framework does not depend on the temporal mechanism of a specific modality. It is recommended to further elaborate on the feasibility and potential challenges of extending it to other modalities to strengthen the argument for the generality of the method.
- 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.
(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 paper proposes a full 4D latent diffusion model, CardioDiT, based on a diffusion transformer (DiT), for synthesizing short-axis cardiac cine MRI. The core idea of this research—directly learning continuous cardiac dynamics using a unified 4D generative model—is novel and exploratory. The experimental evaluation covers commonly used volumetric parameters and commendably examines physiological plausibility and planar anatomical consistency. Preliminary experimental results also show positive signs. However, at present, the model has shortcomings in terms of innovation justification, rigor of experimental comparisons, and completeness of experimental presentation. The value of this paper would be enhanced if a revised version included direct comparisons with key baselines, elucidated methodological design choices, and improved experimental figures and descriptions. Based on this trade-off, I give this paper a “Weak Accept” rating.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
This rebuttal fails to adequately address the core issues of clinical modeling effectiveness, methodological necessity, and originality. The authors provide a insufficient mechanistic explanation of how cardiac physiological constraints are modeled, merely claiming a general improvement in global spatiotemporal consistency. The rationale for the proposed 2D+t→stacked 4D latent design remains unconvincing without comparison to direct 4D latent learning or other encoders. Furthermore, the connection to previous work [15] suggests that its main innovation lies in replacing the U-Net denoiser with a Transformer-based diffusion model, limiting its methodological contribution.
Review #2
- Please describe the contribution of the paper
This work proposes CardioDiT, a fully 4D latent diffusion model (LDM) for the synthesis of short-axis cardiac magnetic resonance (CMR) sequences. The approach combines a spatiotemporal (2D+t) VQ-VAE with a 4D diffusion transformer, enabling joint modeling of spatial and temporal dimensions. The method is evaluated against two baselines (factorized 2D+t and 3D+t LDMs) on both public and private datasets, through a comprehensive quantitative and qualitative analysis, demonstrating improved 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.
Methodological novelty: The proposed 4D diffusion transformer is well motivated for the task, enabling global spatiotemporal self-attention and avoiding space–time factorization. The results consistently support the benefits of this design. That said, the contribution should be considered in the context of closely related work by the authors, including a prior 4D diffusion approach based on a channel-merged 3D U-Net and a concurrent submission proposing a 3D variant of the transformer architecture. While the present work extends these ideas to a fully 4D setting and a different application, clarifying these relationships would strengthen the positioning of the contribution. Strong experimental design: The proposed method is thoroughly validated across three datasets (public and private), with both qualitative and quantitative evaluation. The reported metrics (FID, precision/recall, d-SSIM, AR, EF range, W2) cover complementary aspects of performance, including distributional similarity, structural consistency, and clinical plausibility. A targeted ablation study further explores the impact of model capacity.
- 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.
Incomplete novelty framing and related work discussion: The introduction motivates the limitations of space–time factorization in prior work; however, a relevant study (later introduced in Section 2.4 as a baseline) already addresses this issue using a different architectural choice (channel-merged 3D U-Net). This work is not discussed upfront, and its conceptual relationship to the proposed method remains insufficiently clarified. Given the methodological similarities (e.g., use of a VQ-VAE latent representation and joint spatiotemporal modeling), a clearer positioning of the proposed approach with respect to this prior work would help better delineate the actual contribution.
- 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?
The authors make use of a private dataset (page 3), but no explicit statement is provided regarding ethical approval or IRB compliance.
- 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
1) For clarity and completeness, please include the work in [15] in the introduction and adjust the novelty framing accordingly. In particular, clarify the conceptual differences between the channel-merged 3D+t LDM and the proposed approach (beyond the obvious U-Net vs. transformer difference), and explain why the latter is expected to be advantageous. 2) The caption of fig. 2 refers to the synthesis phase (“VQ-VAE decoding”) but this is not illustrated. In fact, the inference process is not clearly described in the manuscript and should be explicitly detailed. 3) What does “label y” (Fig. 1) represent in your present scenario/task? 4) What is the reason behind the distinct number of slices used when resampling the volumes in the public and private datasets? Is the median number of slices indeed 6 for both public datasets? This seems low for clinical CMR acquisitions. 5) What is the embedding dimension of CardioDiT (variable e in section 2.2)? Is it the same as the autoencoder? 6) Please clarify the meaning of “physiologically uninterpretable” EF values for the factorized 2D+t LDM. If this refers to inconsistencies in temporal dynamics (due to independent slice synthesis), this should be explicitly stated. As written, the statement could be interpreted as referring to spatial discontinuities (as also observed in Fig. 2), which may occur in practice (due to breath-hold acquisition) and do not necessarily invalidate EF estimation. 7) Please justify the decision to merge the results of the two public datasets, and clarify how this was performed (e.g., averaging metrics across datasets vs. computing metrics jointly over all samples). 8) Please comment on the W2 results. CardioDiT appears closer to the training distribution than the real test set is, which may suggest potential overfitting or memorization effects. 9) Similarly, the reduced variability in volume curves (as suggested by the narrower IQR in Fig. 3) may indicate limited sample diversity. It would be helpful to discuss this aspect. Future work may consider complementing the analysis with diversity or coverage metrics.
Minor remarks:
- Some abbreviations (e.g., VQ, VAE, GAN) are not defined at first use.
- VQ-VAE is referred to as “VQ-GAN” in Fig. 1; please standardize the terminology.
- 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?
While the manuscript presents a well-executed study with strong experimental validation and compelling results, the novelty and positioning of the methodological contribution are not sufficiently clear, particularly in relation to closely related prior and concurrent work. A more explicit and careful framing would strengthen the manuscript and better delineate its contribution. That said, the overall quality of the analysis and experimental evaluation is solid and supports acceptance.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I update my recommendation from weak accept to accept. The manuscript presents a strong experimental study with compelling results, and the rebuttal somewhat clarifies the positioning of the contribution with respect to prior and concurrent work (one would expect the camera-ready version to follow accordingly). I also appreciate the additional methodological clarifications regarding implementation details, temporal processing, inference, and metric computation. While I still believe some concerns, particularly regarding memorization/diversity assessment and the interpretation of unreliable EF estimation, were perhaps dismissed somewhat too readily in the rebuttal, they do not detract significantly from the overall quality and validity of the work. Considering the strength of the experimental validation, the relevance of the task, and the overall impact of the results, I believe the manuscript is worthy of acceptance.
Review #3
- Please describe the contribution of the paper
The paper introduces CardioDiT, a latent diffusion framework for unconditional synthesis of short-axis cine cardiac MRI that operates on full 4D (3D spatial + temporal) volumes rather than factorized 2D+t or channel-merged 3D+t representations. A spatiotemporal VQ-GAN first encodes 2D+t slices into a compact latent space; the resulting per-slice latents are then stacked and processed jointly by a Diffusion Transformer with 4D positional encodings and FlashAttention. The paper evaluates this approach against progressively stronger baselines on public (ACDC + M&Ms-2) and private multi-scanner datasets using a clinically grounded metric suite that includes CineMA-based FID, through-plane structural similarity (d-SSIM), inter-slice alignment (ARED), and ejection-fraction distribution matching.
- 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 controlled experimental ladder from 2D+t factorized to 3D+t channel-merged to fully 4D diffusion provides a clear and well-motivated comparison structure. Few papers in cardiac generative modeling attempt this kind of progressive ablation across spatiotemporal coupling strategies, and the design itself should be useful as a template for future work in this area.
The evaluation metric suite goes meaningfully beyond FID-only assessment. The inclusion of ejection-fraction distribution matching via Wasserstein distance, through-plane alignment (ARED), and domain-specific FID computed with CineMA foundation model features rather than ImageNet-pretrained Inception reflects careful thinking about what matters for cardiac MRI synthesis.
CardioDiT-L achieves better distributional quality than the 3D+t U-Net baseline at roughly one-third the parameter count (177M vs. 578M) and less than half the VRAM, with the full pipeline fitting within 24 GB. This practical accessibility lowers the barrier for academic labs working with consumer-grade GPUs.
The scalability ablation across three model sizes (S/B/L) with explicit reporting of parameter counts, inference times, and VRAM requirements provides useful engineering guidance and demonstrates that the framework has room to grow with additional compute.
- 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 baseline comparison simultaneously changes the architecture class (U-Net to transformer) and the spatiotemporal coupling strategy (factorized to joint 4D), making it impossible to attribute improvements to 4D coupling specifically. A factorized DiT baseline (using separate spatial and temporal attention within the same transformer framework) would have isolated the variable the paper claims to study. Without this control, the central thesis that joint 4D modeling is the key driver of improvement remains plausible but not conclusively supported, and the evidence is equally consistent with transformers simply being more efficient than U-Nets for this task.
All quantitative comparisons use only N=50 synthetic samples with no confidence intervals, bootstrap estimates, or significance tests. At this sample size, FID and precision/recall can exhibit substantial variance, and because all metrics are computed on the same 50 generated samples they are not statistically independent. The quantitative differences in Table 1 should therefore be treated as preliminary rather than established. The paper does not explain why N=50 was chosen or whether it reflects a computational constraint.
The model achieves recall of 0.70 on the public dataset and 0.54 on the private dataset, meaning it fails to cover 30-46% of the real data distribution. For the paper’s stated applications of data augmentation and privacy-preserving sharing, systematically excluding nearly half the anatomical and pathological diversity is a serious limitation. The paper does not analyse which phenotypes or modes are missing, nor does it discuss the downstream consequences of this coverage gap.
Despite the “fully 4D” framing used throughout the abstract and introduction, the VQ-GAN encodes 2D+t slices independently along the depth axis, meaning inter-slice spatial information is discarded before the transformer stage. The 4D coupling is introduced only at the diffusion model level. The paper does not test whether a 3D-aware autoencoder would improve results or analyze how much inter-slice information the current factorized encoding loses.
The core transformer block architecture is specified only in supplementary material of a concurrent, unpublished submission that was not available for review. This means the reviewable novelty reduces to the 4D extension (positional encodings and patchification), and independent reimplementation from the main text alone is not possible. Additionally, the temporal standardization to 32 frames via cyclic repetition could inflate temporal consistency metrics by imposing artificial periodicity, yet no analysis of this effect is provided.
- Please rate the clarity and organization of this paper
Satisfactory
- 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses a genuinely important question and introduces a clinically grounded evaluation framework that the field should adopt. However, the experimental evidence does not adequately support the central claim. The most critical issue is that the baseline comparison confounds architecture class with coupling strategy: the observed improvements may reflect the advantage of transformers over U-Nets rather than the benefit of 4D joint modeling. A factorized DiT baseline would have resolved this ambiguity and was feasible within the scope of the work. The evaluation, while thoughtfully designed in its choice of metrics, is statistically underpowered at N=50 with no uncertainty quantification, and the substantial recall gap (covering only 54-70% of the real distribution) goes unanalyzed despite being directly relevant to the paper’s stated applications. The dependence on an unpublished concurrent submission for the core architecture specification further limits what can be assessed from the manuscript alone. A revision addressing the factorized transformer baseline, larger-sample evaluation with confidence intervals, and analysis of the recall gap and cyclic repetition confound would substantially strengthen the contribution.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The rebuttal addressed the cyclic-repetition concern by clarifying that temporal metrics are computed over one detected cardiac cycle rather than the padded 32 frame window, so the artificial periodicity cannot inflate temporal consistency scores. The recall discussion is also correctly reframed and the relative ordering across baselines (with CardioDiT best) still remains. The N=50 choice is now reasonably justified by sample-size matching with the validation set, and the authors explicitly commits to editing Table 1 to soften the claims.
On the central methodological critique (comparison confounds architecture class with coupling strategy) the authors do not run the factorized-DiT control but narrow the claim from “4D coupling causes the gains” to “a joint 4D transformer is an effective realization of spatiotemporal coupling”, which is what the experiments actually support.
Author Feedback
Novelty & baselines(R1/Q1,Q2,Q5;R2/Q1;R3/Q1). We agree related work should be clearer, will add [15] to the introduction and adjust novelty framing. [17] models only ED/ES phases; [15] uses the same latent encoding but a channel-merged LDM with a 3D U-Net denoiser, i.e. our 3D+t baseline. CardioDiT instead models the latent cine volume as spatiotemporal patches with global self-attention. This offers direct long-range slice-time coupling instead of indirect receptive-field growth, improving consistency and fidelity. 4D coupling & architecture(R3/Q1;R1/Q2). Do gains stem from 4D modelling or architecture? Our comparisons increase coupling progressively: 2D+t via inter-slice conditioning, 3D+t within the U-Net class, and CardioDiT via global attention over latent spatiotemporal patches. We do not claim that every gain isolates “4D coupling” independently of architecture; rather, a joint 4D transformer is an effective realization of such coupling. Moving from 2D+t to 3D+t improves consistency but reduces fidelity. Concurrent work (supplement) suggests that replacing U-Nets with transformers improves diffusion fidelity, here translated to 4D. An in-depth analysis of factorized transformer modules [11] is future work. Implementation & temporal processing(R1/Q4,Q8;R2;R3/Q5). Clarification on implementation and temporal metrics: We will revise Fig. 1/caption and Sec. 2.1-2.2 to explain inference/VQ decoding. The supplement details DiT architecture; code/configs will be public. The label y is zero as no labels were available; embedding dimension e matches the autoencoder. Temporal metrics used one cardiac cycle with detected key frames, so temporal consistency is not a metric artifact. Statistical Robustness(R3/Q2). We agree uncertainty estimates would strengthen interpretation and will temper claims. N=50 per model matches the 50 validation cases, ensuring equal real/synthetic sample sizes. Notably, feature metrics use 50 × 32 time-point vectors, not only 50 volume-level vectors. This improves stability relative to one feature vector per volume, and trends are consistent across data pools/ablations. Using the same 50 samples per model supports comparable metrics. We acknowledge that Tab. 1 differences are preliminary. Future work will synthesize larger cohorts and report significance where applicable. Scope. Several suggestions concern further analysis. We emphasize that the manuscript already provides a broad evaluation within the 8-page limit: multiple distributional and functional metrics, diverse private/public data pools, visual analysis, and ablations. We will expand the discussion on augmentation/privacy and on extension to other dynamic modalities. 4D modelling(R1/Q3;R3/Q4). Concerns regarding the “fully 4D” claim and the absence of end-to-end latent representation: “Fully 4D” refers to the DiT operating on the latent 3D+t volume. The 2D+t VQ-GAN follows cine CMR acquisition: anisotropic slices with coarse through-plane resolution but fine in-plane temporal detail. To our knowledge, no established 3D+t/4D VQ-GAN exists for this setting. Systematically comparing latent encoders is beyond scope. Diversity(R2;R3/Q3). Recall indicates imperfect coverage: Recall is a limited feature-space estimate, not direct phenotype diversity and should not be read as exactly 30% absent phenotypes. Within this limitation, CardioDiT has the best recall among compared methods. W2 was computed on validation samples, so it does not imply memorization and may reflect EF differences between splits. Clarifications(R1/Q1,Q6,Q7;R2). Public datasets were merged before training and evaluated jointly over all validation samples. Volumes were cropped to preserve cardiac content. Charts use public samples or samples generated from models trained on them. The 2D+t baseline shows discontinuities from independent slice synthesis, not breath-hold artifacts; hence, EF estimation is unreliable. All clarifications will be added in the camera-ready upon acceptance.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- 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.
Reviewers agree the paper tackles an important problem in 4D cardiac MRI synthesis and includes a strong empirical evaluation across datasets with clinically relevant metrics. Two reviewers rated it Weak Accept, while one raised concerns about novelty and experimental attribution. The rebuttal should clearly position the contribution relative to closely related prior/concurrent work, justify whether gains stem from full 4D modeling rather than architecture changes, and clarify statistical robustness and implementation details.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper presents a well-motivated framework for 4D cardiac MRI synthesis. Reviewer opinions diverged after rebuttal: R1 moved to reject; R2 and R3 moved to accept. The central confound, that gains may reflect the transformer-over-U-Net advantage rather than 4D joint modeling per se, was not resolved experimentally, and the authors narrowed their claim accordingly. The authors are expected to incorporate the reframed novelty positioning, and add the clarified implementation and temporal processing details in the final version.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The rebuttal narrows the claim appropriately and two reviewers moved to Accept. R1 remains Reject, but the main remaining concern is attribution of gains rather than fatal unsoundness. Camera-ready should explicitly soften “fully 4D proves superiority” claims and state that the factorized-transformer control is future work.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
This paper proposes CardioDiT, a full 4D implicit diffusion framework designed to address spatiotemporal consistency in cardiac MRI generation. However, the main conclusion that highlighting the role of full 4D modeling for improvement of anatomical and temporal consistency, is not well supported according to the current experiments. This is a non negligible issue that was not resolved in the rebuttal, leading the paper unacceptable for publication.
