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

Abstract

Late Gadolinium Enhancement (LGE) is the gold standard for myocardial scar assessment, yet conventional static LGE acquisition is prone to artifacts from respiratory and cardiac motion. We reinterpret LGE imaging as a highly accelerated dynamic multi-frame reconstruction problem. By partitioning sequential k-space acquisition into multiple temporal frames without extending scan time, we reconstruct these frames independently to improve motion robustness. To enable reliable reconstruction under extreme undersampling, we develop an inception-powered physics-guided unrolling method. The multi-branch inception design expands the receptive field per iteration, facilitating the separation of structured aliasing from fine myocardial details. Additionally, a parameter-shared fixed-point implementation reduces memory consumption. Experiments on large-scale clinical 3T (100 patients) and 5T datasets demonstrate consistent improvements over state-of-the-art baselines. Notably, even under challenging free-breathing conditions, our method robustly reconstructs high-quality images, demonstrating strong potential for clinical diagnostics.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1026_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ZhoYa_InceptionPowered_MICCAI2026,
        author = { Zhong, Ya AND Cheng, Jing AND Wang, Pan AND Zhang, Zhaonan AND Du, Lixin AND Fan, Weixiong AND Liu, Xin AND Liang, Dong AND Liu, Yuanyuan AND Zhu, Yanjie},
        title = { { Inception-Powered Unrolling Reconstruction for Accelerated Dynamic Cardiac Late Gadolinium Enhancement 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

    Authors propose an Inception-powered unrolling framework for reconstruction of accelerated dynamic cardiac LGE MRI. This method formulates the static LGE imaging as a dynamic multi-frame reconstruction problem, and combines proximal gradient descent, weight-shared equilibrium model, and inception module to develop an unrolled reconstruction model. The method is trained on the public CMRxRecon multimodal dataset and validated on two clinical datasets. The approach outperforms the comparative baselines on both testing sets, and demonstrated robustness to noise and motion.

  • 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 method propose to partition sequential LGE k-space data into multi-frames for motion-resolved dynamic reconstruction 2) This method can mitigate motion blur in static LGE

  • 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) Acquiring segmented k-space data across different cardiac cycles and correcting for motion artifacts is a common strategy for mitigating motion induced image degradation, as demonstrated in prior works, e. g. , Zeilinger at al. , European Radiology, 2022; Munoz et al. , Journal of Cardiovascular Magnetic Resonance, 2020; Fotaki et al. , Journal of Cardiovascular Magnetic Resonance, 2024.

    (2) Inception, proximal gradient descent, and weight‑shared equilibrium model have been well established, and the technical novelty is low.

    (3) I am surprised that CS outperformed MoDL in SSIM (Table 1). The author should double-check their implementation and results.

    (4) Many things are unclear, e. g. , training loss, computational overhead, inference, LGE k-sapce acquisition details and its division into multiple frames.

    (5) Table 2.Many SSIMs are incorrectly bolded.

    (6) The comparative methods, L+S-net, Learned_DC, MoDL, are published in 2019 and 2021.Please include more recent baselines.

    (7) It’s hard to indentify major differences in the results of different methods in Fig 2 and 3.

  • Please rate the clarity and organization of this paper

    Poor

  • 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.

    (2) Reject — should be rejected, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The method simply combines existing modules (inception, proximal gradient descent, weight‑sharing, unrolled framework) and does not provide any novel concepts. Many key details are missing (data details, training/test cost). Comparison with more recent SoTA is missing.

  • 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.

    Authors proposed to partition LGE k-space dataset into multiple temporal subsets for 11x accelerated reconstruction without the motion artefacts issue. Major issuses remain: 1.The concept of partitionining static k-space into multiple temporal segments for reconstruction is not new at all. See Munoz et al., JCMR, 2020; Milotta et al., JCMR, 2021; Escobar et al., JCMR, 2020; Kustner et al., Scientific reports, 2020.2.Limited novelty of methodology. Does integration of Inception into a widely used unrolled model represent any methodological or theoretical innovation? Also see 4D spatoi-temporal U-net in Kustner et al., Scientific reports, 2020; Inception‑Based U‑Net in Vafay Eslahi, Elmira. “Inception-Based U-Net in Magnetic Resonance Image Reconstruction.” (2022). 3.K-space data acquisition and temporal binning/splittign details are missing. A small figure showing the acquisition protocol would be helpful. 4.Comparative methods are outdated: L+S-Net in 2021, Learned_DC in 2021, MoDL in 2019, CS in 2007. 5.Advantages of Inception over LSTM/VIT is subjective and lacks evidence.



Review #2

  • Please describe the contribution of the paper

    This paper formulates LGE MRI as a dynamic multi-frame reconstruction problem to reduce motion artifacts associated with conventional static LGE. The authors introduce an inception-based, physics-guided framework for the highly accelerated dynamic reconstruction. The method is evaluated on a 3T dataset of 300 subjects and a 5T dataset of 7 subjects, with comparisons against baseline methods and ablation studies. Based on the reported results, the proposed framework appears to improve image quality relative to the state-of-the-art methods compared. The authors also evaluate motion robustness on free-breathing data and perform a noise robustness analysis.

  • 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.
    • Clear motivation and clinical relevance: The paper addresses a clinically meaningful problem in dynamic LGE MRI, namely reducing motion artifacts associated with conventional motion-compensated LGE.
    • Technical contribution: The proposed framework formulates LGE as a dynamic reconstruction problem and incorporates inception modules and physics-based priors for highly accelerated reconstruction.
    • Comprehensive evaluation: The paper includes comparisons against four baseline methods, ablation studies, and evaluation on two clinical datasets.
    • Additional robustness analysis: The inclusion of free-breathing data and noise robustness experiments further supports the evaluation.
  • 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 literature review: Several relevant prior works have formulated LGE as a dynamic reconstruction problem, including Sultan et al. [1] and Vornehm et al. [2]. The introduction would be strengthened by discussing these related studies and clarifying how the proposed method differs from them.
    • Insufficient acquisition details: The manuscript would benefit from clearer details about the PSIR acquisition protocol. Please clarify whether the PSIR data were acquired under segmented breath-held conditions or free-breathing conditions. Since PSIR is usually acquired as a single-shot image at a fixed cardiac phase, it would also be helpful to clarify whether the different motion phases shown in the figures correspond to cardiac motion or respiratory motion. A small schematic of the acquisition protocol would improve clarity.
    • Lack of limitations discussion: The manuscript would be stronger if it included a brief discussion of the limitations of the proposed framework, including key assumptions and scenarios where the method may be less reliable. For example, it would be helpful to discuss the achievable temporal resolution, how many k-space readouts are grouped into each reconstructed frame, and how far the acceleration rate can be increased before reconstruction quality or motion fidelity begins to degrade.
    • Limited reproducibility support: The code is not provided. More complete reporting of model configuration, optimization settings, hyperparameters, and evaluation details would improve reproducibility.

    [1] Sultan, Muhammad Ahmad, et al. “An unsupervised method for MRI recovery: deep image prior with structured sparsity.” Magnetic Resonance Materials in Physics, Biology and Medicine, 38.5, 2025, 859–871.[2] Vornehm M., Chen C., Sultan M.A., et al. “Multi-dynamic deep image prior for cardiac MRI.” Magnetic Resonance in Medicine, 2025;94(6):2668–2679.doi: 10.1002/mrm.70000.

  • 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

    Address weaknesses:

    • Add missing hyperparameters and training details.
    • Include a concise limitations paragraph.
    • Clarify the acquisition protocol and the meaning of the displayed motion phases.
    • Discuss relevant prior work on dynamic LGE reconstruction.
  • 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?

    Overall, this paper presents a technically interesting and clinically relevant approach to myocardial scar assessment using dynamic LGE reconstruction. Its main strengths are the dynamic reconstruction formulation, the inception network, and the breadth of experimental evaluation. My main concerns relate to unclear acquisition/experimental details, and limited discussion of related dynamic LGE reconstruction literature. These issues do not necessarily undermine the central contribution, but addressing them would improve the clarity and the positioning of the work.

  • Reviewer confidence

    Very confident (4)

  • [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 authors adequately address my initial concerns by acknowledging the missing acquisition details and limitations that need to be included in the final version. Since these issues appear resolvable through the promised revisions and do not undermine the paper’s main contribution, I stick to to accept.



Review #3

  • Please describe the contribution of the paper

    This paper addresses the persistent challenge of motion artifacts in accelerated Cardiac MRI (CMRI), specifically for Late Gadolinium Enhancement (LGE) sequences. The authors’ primary contribution is a novel framework that reformulates static LGE acquisition as a dynamic multi-frame reconstruction problem. By partitioning k-space data into temporally localized subsets, the method leverages temporal consistency to enhance motion robustness without extending scan duration. Technically, this is enabled by an inception-powered deep unrolling network, which improves multi-scale representation within the iterative reconstruction loop. This combination offers a promising approach to stabilizing reconstruction quality in the presence of patient motion, a key factor for clinical applicability.

  • 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.While deep unrolling networks are established in MRI reconstruction, the integration of Inception modules into the iterative loops is a novel architectural contribution. This design effectively enhances multi-scale feature representation, allowing the network to simultaneously capture global structural context and fine-grained local details during each iteration. 2.The paper proposes a clever reformulation of the problem by treating conventional static Late Gadolinium Enhancement (LGE) MRI as a dynamic multi-frame reconstruction task. By partitioning sequential k-space acquisitions into temporally localized subsets, the authors exploit temporal redundancy without extending scan time. This approach is clinically significant as it directly targets motion robustness by leveraging temporal consistency rather than relying solely on spatial priors. 3.The validation strategy is a strong aspect of this work. The authors utilized a large-scale dataset comprising 100 subjects from 3T clinical scanners, supplemented with additional 5T data. This scale is substantial for deep learning MRI studies and enhances the statistical power of the results. Demonstrating consistent improvements over state-of-the-art (SOTA) baselines across different field strengths (3T and 5T) suggests that the method is not only effective but also generalizable across different hardware configurations, which is crucial for clinical translation. 4.The paper provides a thorough ablation study that clearly isolates the contributions of key components: Architectural Impact: Comparing the full model against variants without the Inception module validates the specific benefit of multi-scale representation. Training Strategy: Evaluating the impact of the mixed-data training strategy helps understand the role of data diversity in model performance. Motion and Noise Robustness: The inclusion of synthetic noise experiments (varying SNR) and, most importantly, validation under free-breathing conditions demonstrates the method’s practical utility. Showing superior performance compared to standard single-frame static LGE imaging in the presence of motion artifacts strongly supports the claim of improved clinical feasibility.

  • 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 authors describe the method as “physics-guided,” yet the manuscript lacks sufficient technical detail regarding how the physical forward model is integrated into the network architecture. Is it just data consistency (DC) layer or there is another physics learned. Without this clarification, it is difficult to assess whether the method truly leverages physics-based constraints to ensure fidelity to acquired data, or if it relies primarily on data-driven priors. This distinction is vital for understanding the method’s robustness to out-of-distribution sampling patterns or hardware variations. 2.The proposed method relies on Inception modules within a deep unrolling framework to handle the “dynamic” multi-frame nature of the reformulated LGE problem. However, the authors do not sufficiently justify why this convolutional-based multi-scale approach was chosen over established sequence modeling architectures such as Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), or Vision Transformers (ViT). 3.The paper does not provide adequate details on the k-space undersampling (acceleration) scheme used during training and evaluation. Key parameters such as the acceleration factor, the type of mask (e.g., random Cartesian, variable-density Poisson disk, radial, or spiral), and the centering strategy (number of auto-calibration signal lines) are either missing or vaguely described. Furthermore, it is unclear if the same masking strategy was applied across both the 3T and 5T datasets. 4.The authors claim that their “parameter-shared fixed-point implementation reduces memory consumption,” but this assertion is not supported by empirical evidence.

  • 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.

    (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 presents a solid, clinically relevant contribution with strong experimental backing. The proposed method offers a tangible improvement in motion-robust CMRI reconstruction, and the evaluation is thorough enough to support the claims. The identified weaknesses are minor and can be addressed through clarifications in the final manuscript without requiring further experimentation or major revisions. Therefore, I believe the paper meets the standards for 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.

    The authors have satisfactorily addressed many critical concerns in their rebuttal. They clarified that the “physics-guided” aspect relies on explicit multi-coil data consistency within a fixed-point unrolled framework, justifying the use of Inception modules for their ability to simultaneously handle local scar preservation and global aliasing suppression under extreme acceleration, where LSTMs or ViTs may be less effective. The authors also resolved reproducibility issues by detailing the TGRAPPA-style time-interleaved Cartesian sampling strategy (with 14–22 ACS lines) applied consistently across 3T and 5T datasets. These clarifications significantly strengthen the technical rigor and validity of the work.



Author Feedback

Thank all reviewers and the Meta-reviewer for the constructive comments. 1.Novelty and design rationale: We clarify that our framework is a task-specific design for ultra-high-accelerated (R>11) dynamic LGE reconstruction rather than a simple combination of existing modules. Therefore, our primary novelty lies not in introducing isolated network components, but in establishing a dynamic reconstruction methodology specifically tailored for highly accelerated PSIR-LGE acquisition. 1) From Static Recovery to Dynamic Reconstruction. Conventional motion-compensated LGE methods fuse multiple shots into a single static image. In contrast, we reconstruct each PSIR shot as an independent frame without explicit motion compensation or shot fusion, while exploiting temporal redundancy across shots during reconstruction. This reformulation converts motion from an artifact source into a resolvable reconstruction dimension. 2) Task-Specific Multi-Scale Prior for Dynamic LGE. LGE reconstruction under R>11 simultaneously requires the preservation of small structures like hyperenhanced scar and the suppression of globally distributed aliasing artifacts. The Inception module provides complementary receptive fields for these two competing requirements, while LSTM or ViTs may struggle to simultaneously preserve local features and suppress global aliasing under extreme acceleration. 3) Fixed-Point Unrolling for Dynamic Reconstruction. For high-resolution 2D+t reconstruction, naive deep unrolling becomes computationally prohibitive due to the large memory consumption associated with storing intermediate variables across iterations. We therefore adopt a fixed-point implementation with Jacobian-free backpropagation. It enables memory-efficient iterative reconstruction for dynamic data within practical GPU memory constraints. The physics-guided component corresponds to explicit data consistency using the multi-coil forward model. Unlike recent DIP-based dynamic LGE methods requiring subject-specific optimization during inference, our supervised framework enables fast and efficient inference. 2.Acquisition and implementation clarification: The reconstructed “phases” do not correspond to conventional cardiac phases or respiratory phases. Instead, they are generated from independently acquired PSIR shots collected across adjacent R-R intervals. We employed a time-interleaved Cartesian undersampling strategy (TGRAPPA-style sampling). Time-averaged central ACS lines (14–22 depending on orientation) were used for coil sensitivity estimation. The same undersampling strategy was applied to both 3T and 5T datasets. Typical imaging parameters were TR/TE(ms)=4.3/1.82 (4CH), 5.2/2.13 (3CH/2CH), and 4.2/1.73 (SAX), with 5–8 shots depending on acquisition orientation. The model was trained using a combined MAE and perceptual loss with equal weighting. Due to page limits, these details were omitted; code will be released upon acceptance. 3.Experimental clarification: 1) To better show the reconstruction quality of myocardium, we tightly cropped the reconstructed images with only cardiac region remaining. Our method performs better in artifact suppression and detail preservation of myocardium, and shows less phase-inconsistency-related distortions in the full image. The arrow annotations to Fig 2 and full images of Fig 3 will be added to highlight our advantage. 2) The counter-intuitive observation (CS outperforming MoDL in SSIM) was primarily due to an inconsistency in the intensity normalization protocols during the initial metric calculation. We have now re-evaluated all baselines using a unified, strictly consistent normalization framework. The corrected results now show consistent superiority of deep learning methods over traditional CS. We apologize for this oversight and have updated Tables 1–2 accordingly. 4.Discussion: Current reconstruction quality may degrade under more aggressive acceleration, which will be discussed in the revised manuscript.



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.

    The manuscript addresses a clinically relevant problem in accelerated dynamic LGE MRI and reports promising results on a large 3T cohort with additional 5T validation. However, the novelty over prior motion-resolved/free-breathing LGE reconstruction and recent dynamic MRI reconstruction methods is not clearly established, and the technical contribution appears to mainly combine known components. The rebuttal should focus on clarifying the specific methodological novelty, explaining the LGE acquisition and k-space-to-frame partitioning, justifying the baseline choices, and providing the missing implementation details needed for reproducibility. The authors should also clarify the reported baseline behavior and temper claims that are not directly supported by the submitted experiments.

  • 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.

    The paper addresses an important clinical problem and reports promising reconstruction results. However, the rebuttal does not fully resolve concerns about novelty relative to prior dynamic/motion-resolved LGE and existing unrolled reconstruction methods. The clarification that the reconstructed frames are PSIR shots rather than conventional cardiac or respiratory phases also weakens the paper’s original framing. In addition, the acknowledged normalization error in the submitted quantitative comparisons reduces confidence in the reported performance gains.



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 paper addresses a clinically relevant problem and presents a technically interesting framework for motion-robust dynamic LGE reconstruction with reasonable experimental validation. Reviewers were split: two recommended acceptance while one recommended rejection, based on limited methodological novelty, outdated baselines, and missing implementation details. The rebuttal larified the key design rationale, acquisition protocol, and corrected a normalization inconsistency in the baseline comparison. The remaining concern regarding novelty over prior dynamic LGE reconstruction work is noted, but does not outweigh the overall merit.



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.

    The clinical motivation and evaluation are pretty strong. The 3T/5T validation with free-breathing and noise robustness experiments is quite thorough. The combination of model components represents a new task-specific contribution.



back to top