Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Cardiac cine MRI enables visualization of the beating heart during free breathing, but severe undersampling and motion make reconstruction highly challenging. A central challenge for reconstruction is incorporating powerful priors of cardiac anatomy while remaining computationally efficient. We propose Piecewise Dynamic Diffusion Regularization (PDDR), a reconstruction method that integrates a spatiotemporal diffusion model as a generative prior within a variational reconstruction framework for cine MRI. The model employs dedicated spatial layers to encode anatomical structure and temporal layers to capture cardiac motion learned from gated cine data. PDDR leverages the dynamic prior in a piecewise manner, enabling the efficient use of spatiotemporal diffusion models for processing of long real-time sequences. Experiments on retrospectively accelerated and prospective real-time cine MRI demonstrate that PDDR outperforms classical, unsupervised, and diffusion-based methods, delivering high-quality reconstructions with substantially reduced computation time compared to state-of-the-art baselines. These results highlight PDDR as a practical and scalable solution for free-breathing, real-time cardiac MRI. Code is available at https://github.com/MLI-lab/pddr.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1880_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/MLI-lab/pddr

Link to the Dataset(s)

N/A

BibTex

@InProceedings{FürFlo_Piecewise_MICCAI2026,
        author = { Fürnrohr, Florian AND Heckel, Reinhard},
        title = { { Piecewise Dynamic Diffusion Regularization for Reconstruction of Cardiac Cine MRI } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16888},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper addresses reconstruction of highly undersampled cardiac cine MRI, with particular emphasis on real-time free-breathing acquisitions, where reconstruction is challenging due to severe undersampling, cardiac motion, and respiratory motion. The main contribution is a new method termed Piecewise Dynamic Diffusion Regularization (PDDR), which integrates a spatiotemporal diffusion model as a generative prior into a variational reconstruction framework. The key idea is to apply the dynamic diffusion prior piecewise on blocks of consecutive frames rather than on the entire video sequence at once, thereby making diffusion-based regularization computationally feasible for long cine sequences. The paper also introduces a customized separable spatiotemporal U-Net with spatial layers for anatomy and temporal layers for motion modeling. Experiments on retrospectively accelerated gated cine MRI and prospectively acquired real-time cine MRI show that PDDR outperforms classical, unsupervised, and spatial-prior baselines in reconstruction quality, while being substantially more efficient than diffusion posterior sampling and significantly faster than scan-specific untrained baselines such as FMLP and T-DIP.

  • 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 tackles an important and clinically relevant problem. Real-time free-breathing cardiac cine MRI is a meaningful target because conventional gated/breath-held approaches suppress physiological variability and can be uncomfortable or impractical. The paper is well motivated in arguing that this setting requires strong spatiotemporal priors and computationally efficient reconstruction. 2.The method is technically well aligned with the problem. Using a spatiotemporal diffusion prior is a natural way to model cardiac anatomy and motion jointly, and the piecewise regularization strategy is a practical solution to the memory/runtime bottleneck that otherwise limits diffusion-based reconstruction of long sequences. This is the paper’s strongest idea: it is not merely applying an existing diffusion model, but adapting the regularization mechanism to the long-sequence cardiac cine setting. 3.The paper combines methodological and practical contributions. In addition to the variational piecewise regularization framework, the proposed separable spatiotemporal U-Net is computationally sensible and tailored to dynamic imaging. The paper explicitly compares this design to naive 3D residual blocks and motivates it from an efficiency standpoint. 4.The retrospective quantitative evaluation is strong. On retrospectively undersampled CMRxRecon data, PDDR achieves the best PSNR and SSIM among the reported methods at 12-fold acceleration, while remaining much more practical than diffusion posterior sampling. Table 1 shows that PDDR slightly outperforms DPS in PSNR/SSIM while using much less VRAM and reconstruction time. This is a meaningful result because it demonstrates that the proposed variational piecewise approach can retain much of the benefit of diffusion priors without the heavy computational burden of posterior sampling. 5.Efficiency analysis is a notable strength. The paper does not stop at reporting image-quality metrics. It explicitly studies runtime, VRAM, block sampling strategy, sequence length, block size, and optimization steps. The block-sampling experiment is useful because it directly supports the paper’s central claim that piecewise regularization is what makes long-sequence dynamic diffusion reconstruction practical. 6.The prospective experiment improves the practical relevance of the paper. Although qualitative, the evaluation on real free-breathing OCMR acquisitions is important. The visual examples and reported runtimes suggest that the method scales to realistic long-sequence prospective data better than several alternative approaches. 7.Overall clarity is good. The paper is well structured, the method is logically presented, and the motivation for the piecewise formulation is easy to follow. The figures and tables support the narrative effectively.

  • 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 novelty is good but not fully maximal at the architectural level. The strongest contribution is the piecewise regularization framework, but the diffusion prior itself is based on a fairly standard spatiotemporal U-Net design with separable spatial and temporal processing. The paper is therefore more compelling as a reconstruction-framework contribution than as a fundamentally new generative modeling contribution. 2.The prospective real-time evaluation is mostly qualitative. Since the paper emphasizes real-time free-breathing cine MRI as a key use case, it would be stronger to include more rigorous quantitative evaluation on prospective acquisitions, even if only through surrogate or expert-based metrics. At present, the strongest quantitative evidence comes from retrospectively undersampled gated data rather than the true target setting. 3.The retrospective training/evaluation setup may not fully match the claimed deployment scenario. The dynamic diffusion prior is trained on gated breath-hold CMRxRecon data and then applied to real-time free-breathing prospective data. The paper argues that reconstruction quality remains good under respiratory motion, but the domain shift between gated training data and free-breathing deployment is substantial, and this aspect is not deeply analyzed. 4.The comparison set could be broader with respect to recent learned dynamic cardiac MRI methods. The included baselines are reasonable and cover classical, untrained, and diffusion-based methods, but the paper would be stronger if it more explicitly positioned itself against stronger recent supervised dynamic cine reconstruction methods or recent cardiac-specific learned reconstructions that are closer to the target application. 5.Some conclusions are based on relatively small prospective evidence. The prospective experiment uses 10 acquisitions from OCMR. This is useful as a proof of applicability, but still limited for making broad claims about robustness across scanners, pathologies, motion patterns, or acquisition protocols. 6.The reported metrics deserve a little more discussion. In Table 1, PDDR has the best PSNR/SSIM but not the best NMSE, where DPS appears lower. The method is still strong overall, but the paper could discuss this tradeoff more carefully rather than presenting the result as uniformly dominant. 7.Reproducibility details are decent but not exhaustive. The paper explains the optimization objective, block size, number of steps, and datasets, but more implementation detail would be useful for exact replication of the diffusion training and reconstruction pipeline.

  • 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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?

    My overall recommendation is based on the combination of a well-motivated clinical problem, a technically sensible method, and strong empirical evidence for a practically meaningful efficiency–quality tradeoff. The paper addresses real-time free-breathing cardiac cine MRI, which is a genuinely important reconstruction setting and one where standard approaches face clear limitations. The central idea of using a spatiotemporal diffusion prior in a piecewise variational regularization framework is both practical and relevant, since it directly addresses the main computational bottleneck that limits diffusion-based dynamic MRI reconstruction.

    The main factors limiting my score are that the prospective target setting is evaluated primarily qualitatively, the training-to-deployment domain shift is not deeply analyzed, and the architectural novelty of the diffusion backbone itself is more moderate than the novelty of the overall reconstruction framework. In addition, broader comparison to recent learned cardiac cine reconstruction methods would strengthen the positioning of the work. Overall, I view this as a solid and practically relevant paper with a meaningful methodological contribution, especially in how it makes spatiotemporal diffusion priors usable for long cardiac cine sequences, but with some limitations in validation breadth.

  • 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



Review #2

  • Please describe the contribution of the paper

    The authors propose Piecewise Dynamic Diffusion Regularization (PDDR), a reconstruction method that integrates a spatiotemporal diffusion model as a generative prior within a variational framework for cardiac cine MRI. The method employs dedicated spatial layers to encode anatomical structure and temporal layers to capture cardiac motion learned from gated cine data. PDDR applies the diffusion model in a piecewise manner to random blocks of consecutive frames during optimization, enabling efficient processing of long real-time sequences that would otherwise exceed memory constraints. Evaluated on retrospectively accelerated CMRxRecon data (12× acceleration) and prospective real-time OCMR data, PDDR achieves 32.84 dB PSNR and 88.06% SSIM, outperforming classical (L+S), unsupervised (FMLP, T-DIP), and diffusion-based (DPS) baselines while requiring substantially less computation time (43s vs. 419s for DPS) and memory (5.4 GB vs. 16.3 GB).

  • 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.

    Novel and Practical Piecewise Regularization Strategy: The core innovation—applying a spatiotemporal diffusion model to random blocks of consecutive frames rather than the full sequence—is elegant and addresses a genuine computational bottleneck. This design enables the use of powerful spatiotemporal priors for long real-time sequences (up to 128 frames, 6+ seconds) that would otherwise be infeasible due to memory and runtime constraints. The piecewise approach is well-motivated and represents a meaningful contribution to scalable diffusion-based reconstruction. Effective Spatiotemporal Architecture: The separable spatiotemporal U-Net with skippable temporal integration is well-designed. By decomposing 3D convolutions into separate 2D spatial and 1D temporal layers with learnable fusion weights, the architecture efficiently captures spatiotemporal correlations while reducing memory consumption. The ablation (SDR vs. PDDR) convincingly demonstrates that the temporal modeling is critical, as the spatial-only prior underperforms L+S, while the spatiotemporal prior achieves state-of-the-art results. Strong Quantitative Results: PDDR achieves the best reconstruction quality across all metrics (PSNR: 32.84 dB, SSIM: 88.06%, NMSE: 0.097) on retrospectively accelerated data, matching or slightly exceeding DPS (which uses the same diffusion model but with posterior sampling) while being 10× faster and using 3× less memory. The improvements over classical L+S (+1.58 dB PSNR) and unsupervised methods (T-DIP: +3.11 dB) are substantial and consistent. Comprehensive Experimental Validation: The evaluation is thorough and well-structured, covering: (1) retrospective reconstruction on CMRxRecon with quantitative metrics and comparison to five diverse baselines, (2) ablation studies on block sampling strategies (random vs. sliding window) and block size Q, and (3) qualitative assessment on prospective real-time OCMR data. The experiments convincingly validate the design choices and demonstrate practical applicability. Computational Efficiency: PDDR achieves a favorable trade-off between reconstruction quality and computational cost. Compared to DPS, it is 10× faster (43s vs. 419s) and uses 3× less memory (5.4 GB vs. 16.3 GB). Compared to unsupervised methods (FMLP: 1124s, T-DIP: 445s), it is 10-25× faster while achieving higher quality. This efficiency makes PDDR practical for clinical deployment. Flexible and Scalable Design: The method offers flexibility through hyperparameters Q (block size) and K (optimization steps), allowing users to trade off memory, runtime, and reconstruction quality based on hardware constraints and clinical requirements. The sliding window sampling strategy ensures robust performance even with small block sizes relative to sequence length (e.g., Q=12 for N=120). Clear Presentation: The paper is well-written and clearly structured. The motivation, methodology, and experimental design are easy to follow. The figures (especially Fig. 1 and Fig. 2) effectively illustrate the key findings.

  • 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.

    Limited Technical Novelty in Individual Components: While the piecewise regularization strategy is novel and practical, the individual components—variational diffusion-based reconstruction [14, 17], separable spatiotemporal architectures [1], and sliding window processing—are well-established techniques. The paper’s contribution is primarily the combination and application of these methods to cardiac MRI, rather than fundamental algorithmic innovation. The novelty is more engineering-focused than methodological. Insufficient Justification of Piecewise Approximation: The piecewise regularization strategy (Eq. 1) replaces the full regularization term with a stochastic estimate based on a random Q-sized block. While this is intuitively reasonable, the paper does not provide theoretical justification or analysis of the approximation error. Specifically: (1) Under what conditions is this stochastic estimate an unbiased estimator of the full regularization gradient? (2) How does the variance of the gradient estimate depend on Q and N? (3) Is there a principled way to choose Q and K to achieve a target approximation quality? A theoretical analysis or empirical convergence study would strengthen the contribution. Limited Comparison with State-of-the-Art Supervised Methods: The baselines are limited to classical (L+S), unsupervised (FMLP, T-DIP), and one diffusion-based method (DPS). Modern supervised deep learning methods for dynamic cardiac MRI reconstruction (e.g., [20, 25] cited in the paper, or more recent transformer-based or recurrent models) are not included. While the authors mention the challenge of obtaining ground-truth training data, a comparison on the retrospective setting (where ground truth is available) would contextualize the performance and clarify when unsupervised/diffusion-based methods are preferable. Lack of Statistical Significance Testing: The quantitative results (Table 1) report mean ± std over 5 runs with different random masks, which is good practice. However, no paired statistical tests (e.g., t-test, Wilcoxon) are performed to confirm that the improvements over baselines are statistically significant. Given the modest standard deviations (e.g., PSNR: 32.84 ± 0.08 for PDDR vs. 32.78 ± 0.06 for DPS), statistical validation is essential to rule out random variation. Prospective Evaluation is Purely Qualitative: The prospective real-time experiment (Section 4.3) provides only qualitative visual comparisons without quantitative metrics. While ground truth is unavailable for prospective data, the paper could report: (1) temporal smoothness metrics (e.g., temporal total variation, optical flow consistency), (2) downstream task performance, or (3) expert radiologist assessment of image quality and diagnostic utility. The lack of quantitative validation on prospective data limits the strength of the clinical applicability claims. Generalization to Pathological Cases Not Validated: The training and evaluation are conducted on healthy subjects from CMRxRecon and OCMR. The paper does not validate on pathological cases (e.g., myocardial infarction, cardiomyopathy, valvular disease). How well does the learned spatiotemporal prior generalize to abnormal cardiac anatomy or motion patterns? Do the reconstructions degrade for diseased hearts, or does the diffusion prior provide robustness? This is a critical question for clinical deployment. Sliding Window Sampling Strategy Underexplored: Figure 1 shows that sliding window sampling outperforms random sampling, especially for small Q/N ratios. However, the paper does not explain the sliding window implementation details: (1) Is the window moved sequentially (q = 1, 2, 3, …) or with a fixed stride? (2) How many passes through the sequence are performed in K iterations? (3) Does the method ensure that all frames receive equal regularization coverage, or are edge frames regularized less frequently? (4) How sensitive is performance to the window stride? These details are important for reproducibility and understanding the method’s behavior. Hyperparameter Selection Not Fully Justified: Several hyperparameters are introduced but not fully justified: (1) the time-dependent weight in Eq. 1—how is λ chosen, and how sensitive are results to this choice? (2) the diffusion step schedule with 0 < T’ < T—what is the value of T’, and why not use the full range [0, T]? (3) the block size Q and optimization steps K—while Fig. 1 provides some guidance, a more systematic hyperparameter sensitivity analysis would be valuable. Reproducibility Concerns: Several critical implementation details are missing: (1) What is the exact architecture of the spatiotemporal U-Net (number of channels per layer, downsampling factors, attention mechanism details)? (2) What optimizer and learning rate schedule are used for training the diffusion model? (3) What optimizer and learning rate are used for the variational reconstruction (Eq. 1)? (4) How are the sensitivity maps estimated for the prospective OCMR data? (5) What is the exact definition of the kt-Gaussian undersampling masks? Without these details, reproducing the method would be challenging. No Discussion of Failure Cases or Limitations: The paper does not present any failure cases or discuss scenarios where PDDR might struggle. For example: (1) What happens with extreme undersampling (R > 16)? (2) How does the method handle irregular heart rates or arrhythmias in real-time data? (3) Are there cases where the piecewise regularization introduces temporal discontinuities or artifacts at block boundaries? (4) How does performance degrade with increasing sequence length N? Providing failure case examples would give a more balanced view of the method’s capabilities and limitations.

  • 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 does not provide sufficient information for reproducibility.

  • 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?

    This paper presents a practical and well-executed contribution to diffusion-based reconstruction of cardiac cine MRI. The piecewise regularization strategy is an effective solution to the computational bottleneck of applying spatiotemporal diffusion models to long real-time sequences, and the separable spatiotemporal architecture is well-designed. The experimental validation is comprehensive, covering retrospective and prospective data with thorough comparisons to diverse baselines. PDDR achieves state-of-the-art reconstruction quality (PSNR: 32.84 dB, SSIM: 88.06%) while being 10× faster and using 3× less memory than the closest diffusion-based baseline (DPS). However, several limitations prevent an acceptance recommendation: (1) the technical novelty is incremental, as the method combines existing techniques rather than introducing fundamentally new algorithms, (2) the piecewise approximation lacks theoretical justification or convergence analysis, (3) the prospective evaluation is purely qualitative without quantitative metrics or clinical validation, (4) statistical significance testing is absent, and (5) critical implementation details (architecture, hyperparameters, sliding window strategy) are missing, hindering reproducibility. (5) The performance is minor improvement (PSNR: +0.06dB compared to DPS, NMSE: -0.028 compared to DPS). Additionally, the method is not validated on pathological cases or compared to modern supervised baselines. If the authors can address these concerns in the rebuttal—particularly by providing statistical validation, quantitative evaluation on prospective data, and more detailed implementation specifications—the paper would be suitable for acceptance. The work demonstrates clear practical value and strong empirical results, but the evaluation and technical exposition need to be strengthened to fully support the claims.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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.

    The novelty is more engineering-focused than methodological.



Review #3

  • Please describe the contribution of the paper

    This manuscript proposes Piecewise Dynamic Diffusion Regularization (PDDR) for cardiac cine MRI reconstruction. The method combines a spatiotemporal diffusion prior with a variational reconstruction objective, but applies the prior only to local temporal blocks rather than the full sequence, to reduce memory and runtime. The diffusion network uses separable spatial and temporal layers to model anatomy and motion. Empirically, the paper evaluates PDDR on retrospectively undersampled gated cine data from CMRxRecon and on a small prospective real-time free-breathing set from OCMR. The main claim is that PDDR improves reconstruction quality over classical, unsupervised, and diffusion-based baselines while being much faster than posterior-sampling diffusion methods.

  • 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.

    A piecewise regularization strategy for using spatiotemporal diffusion priors on long cine sequences. A separable spatiotemporal diffusion U-Net with learnable spatial/temporal integration. A variational diffusion-based reconstruction framework for cardiac cine MRI that is intended to scale from gated cine to longer free-breathing real-time scans.

  • 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.

    In the current literature, spatiotemporal diffusion for cardiac/dynamic MRI is no longer a fresh idea on its own. The closest recent papers already occupy much of that space. The manuscript’s genuine novelty appears to be the piecewise variational deployment of such a prior, not the overall paradigm. The paper needs tighter positioning against recent cardiac diffusion papers, The retrospective experiments on CMRxRecon do not compare against strong contemporary supervised cardiac cine methods such as CineVN, nor against challenge-leading CMRxRecon-style methods. Given that CMRxRecon has become a major public benchmark and recent papers report strong supervised performance on cardiac cine, omitting these baselines makes it difficult to judge the manuscript’s actual standing. Reporting mean ± std over random masks is useful but insufficient by current standards for strong claims of superiority. The manuscript lacks paired significance tests, confidence intervals across cases, or any clinically meaningful equivalence analysis. Recent cardiac reconstruction papers with translational ambitions are more rigorous here. There is some analysis of block sampling and block size, but not enough on architecture choices, temporal-layer contribution, noise schedule choices, optimization hyperparameters, or sensitivity to Q and K in realistic prospective settings. For a method aimed at practical real-time CMR, the absence of functional endpoints, reader confidence, or at least segmentation-based downstream fidelity means the clinical/practical value remains unproven. This is especially important because recent cardiac cine papers explicitly evaluate functional measures, not just image fidelity.

  • 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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?

    In short, I think the work is promising and could become a solid paper with better positioning and stronger evidence, but in its current form I place it slightly below the acceptance bar.

    On the positive side, the paper tackles a clinically and technically relevant challenge in cardiac MRI reconstruction, and the proposed piecewise diffusion regularization strategy is practically motivated. I also appreciate that the work considers runtime and memory, not only image metrics. However, the main reasons for my score are the following. First, the paper’s broader framing is not sufficiently novel relative to recent literature. Spatiotemporal diffusion for cardiac/dynamic MRI has already been explored in recent papers, so the manuscript’s contribution is more incremental than it initially appears. Second, the baseline comparisons are not strong enough for the claim level, especially given the availability of recent strong cardiac cine reconstruction methods and public benchmarks. Third, the prospective validation is limited and mostly qualitative, which makes it difficult to assess real clinical or practical impact. Finally, the paper would benefit from stronger statistical analysis and a more detailed robustness study.

  • 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.

    N/A



Author Feedback

Firstly, we thank all reviewers for the constructive feedback, and their recognition of the practical relevance of PDDR for making diffusion-based reconstruction computationally feasible for long cine sequences.

Novelty: To the best of our knowledge, no other diffusion-based reconstruction method for dynamic MRI (1) applies variational reconstruction, all previous methods use computation heavy sampling, (2) provides a principled way for handling long video sequences efficiently, and (3) applies spatiotemporal diffusion priors to free-breathing cardiac MRI. As none of the reviewers provided further references to prior work, we only clarify the distinction to dSTDM [27], which uses 2D priors on x-t and y-t slices and an expensive sampling approach.

Reproducibility: Hyperparameters are set to λ=0.05, T’=0.4*T as the MRI measurements provide initial information making the full range used in unconditional sampling unnecessary, and for best regularization coverage sliding window is applied with stride Q. Furthermore, after acceptance we publish the full code, including hyperparameter, model, and evaluation configurations.

Significance: For results in Tab. 1, we ran paired t-tests for statistical analysis. The improvements of PDDR over DPS in terms of PSNR/SSIM are weakly significant (p<0.1), and VRAM/Time are highly significant and of exceptionally practical value as they enable the use of diffusion models within a reasonable compute budget. All other reported metrics are significant with p<0.01. Baseline selection: Supervised methods require exact knowledge of the forward operator (sampling pattern, acceleration factor) during training. Matching the setup, supervised methods perform well, but fail to generalize to variations at inference. As in real-time cardiac MRI it is practically impossible to acquire fully sampled ground truth, training of the methods is based on binned references and simulated sampling patterns which do not align with the actual acquisition trajectory of the data. This mismatch makes generalization to real-time reconstruction difficult. We therefore decided not to include supervised baselines, as the main goal of the paper is reconstruction of real-time, free-breathing MRI. Diffusion models, on the other hand, are trained unconditionally without knowledge of the inverse problem. They flexibly adapt to different choices of the forward model at inference. Furthermore, strong performance on the CMRxRecon benchmark does not directly relate to our case. Although, the retrospective experiment is similar in spirit, CMRxRecon adds an ACS of 16 center lines to undersampling masks, which is inapplicable in real-time acquisitions and simplifies the reconstruction problem greatly. In this setting, PDDR also shows significantly improved performance.

Quantitative prospective validation: For the prospective reconstructions, we provide the signal-to-error ratio (SER [dB], see [13]) quantifying the estimation error with respect to holdout frequencies, and the motion smoothness metric temporal total variation (TTV), amending the qualitative evaluation shown. We report the mean over the discussed prospective testset. As reference, we provide the values of SDR (SER:6.19/TTV:110.4), which suffers from strong motion artifacts. PDDR achieves strong hold-out estimation (SER:16.05) and balanced temporal smoothness (TTV:31.4). Confirming the qualitative discussion, FMLP produces overly smooth solutions (SER:14.45/TTV:7.65), while T-DIP achieves similar reconstruction quality (SER:16.03/TTV:33.7). Note that FMLP and T-DIP are optimized per sample with early stopping based on SER performance. L+S reconstructions show similar signal estimation capabilities (SER:16.12) and lower temporal variation (TTV:18.7), but the perceived noise level is much higher. Low-rank enforcing can lead to lower motion dynamics.

Finally, we wanted to remind the reviewers of the tight 8-page limit, which made presentation of additional results difficult.



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.

    This paper addresses an important and clinically relevant problem in real-time free-breathing cardiac cine MRI reconstruction, and all three reviewers acknowledge its practical motivation, promising empirical performance, and meaningful efficiency advantages. In particular, the piecewise diffusion regularization strategy is viewed as the main contribution, since it makes spatiotemporal diffusion priors feasible for long cine sequences while retaining strong reconstruction quality. At the same time, the two weak-reject reviews make clear that the current manuscript is not yet fully convincing in its present form, mainly because the novelty needs to be positioned more carefully against recent diffusion-based and supervised cardiac MRI reconstruction methods, the prospective free-breathing evaluation remains largely qualitative, and the evidence for superiority would be stronger with more rigorous statistical analysis, clearer implementation details, and broader validation. Importantly, however, these concerns do not fundamentally dispute the value or promise of the approach; rather, they point to issues of positioning, justification, and experimental support that may be clarified or strengthened in rebuttal. I therefore recommend invite for rebuttal. In the rebuttal, the authors should focus on clarifying the precise novelty of the piecewise framework relative to recent cardiac diffusion literature, justifying the current baseline selection, discussing the practical significance of the reported gains, and addressing the reviewers’ concerns about prospective validation, statistical testing, and reproducibility 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.

    This paper proposes Piecewise Dynamic Diffusion Regularization (PDDR), a variational reconstruction framework that applies a spatiotemporal diffusion prior to local temporal blocks, aiming to make diffusion-based reconstruction feasible for long real-time free-breathing cardiac cine MRI sequences. The reviewers recognized the clinical relevance of the problem, the practical motivation of addressing the memory/runtime bottleneck in diffusion-based dynamic MRI reconstruction, and the meaningful quality–efficiency trade-off achieved by the proposed piecewise regularization strategy. The paper also provides retrospective quantitative evaluation, prospective free-breathing experiments, and efficiency analysis comparing runtime and memory usage. The reviewers raised concerns about the level of novelty, baseline selection, prospective validation, statistical support, and reproducibility details. In particular, the novelty is mainly in the piecewise deployment of a spatiotemporal diffusion prior rather than in a fundamentally new diffusion architecture, and the prospective evaluation remains limited. However, these concerns do not fundamentally undermine the value of the approach. The core contribution is practically meaningful: PDDR makes spatiotemporal diffusion regularization applicable to long cine sequences while retaining strong reconstruction quality and substantially reducing computation compared with diffusion posterior sampling. In the rebuttal, the authors clarified the novelty relative to existing dynamic MRI diffusion methods, justified the exclusion of supervised baselines for the real-time free-breathing setting, provided additional implementation details and hyperparameter choices, added statistical testing, and reported quantitative prospective metrics using SER and TTV. Although some limitations remain, such as the lack of pathological validation and reader-based clinical evaluation, the rebuttal sufficiently addressed the main concerns regarding positioning, reproducibility, statistical analysis, and prospective evaluation. Therefore, I recommend acceptance.



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.

    I think the authors have adequately addressed the reviewers’ comments.



Meta-review #3

  • 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.

    I recommend acceptance. The paper addresses an important and practical problem in cardiac cine MRI reconstruction, especially for long real-time free-breathing sequences where fully sampled supervision is difficult and diffusion-based posterior sampling can be computationally prohibitive. The main contribution, piecewise dynamic diffusion regularization, is a sensible and useful way to make spatiotemporal diffusion priors feasible for long cine reconstruction.

    The initial reviews raised valid concerns about novelty, baseline selection, statistical support, prospective validation, and reproducibility. The rebuttal addressed several of these points reasonably well by clarifying the distinction from prior diffusion MRI work, providing key implementation details, adding statistical testing, and giving additional prospective surrogate metrics. While the method is more of a practical reconstruction-framework contribution than a fundamentally new generative model, the efficiency gains over diffusion posterior sampling are substantial and clinically relevant, and the retrospective and prospective results support the usefulness of the approach.



back to top