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

Abstract

Accurate three-dimensional (3D) reconstruction of the four-chamber heart from routine cine MRI remains an open challenge due to sparse slice acquisition, inter-slice misalignment, and the structural heterogeneity of cardiac substructures. Although cine MRI is the clinical gold standard for cardiac assessment, its inherent 2D acquisition limits comprehensive 3D characterization of both ventricular and atrial anatomy. Existing reconstruction approaches predominantly focus on biventricular geometry and do not explicitly address multi-class anatomical completion under severe sparsity. We propose HeartFormer, a semantic-aware dual-structure transformer designed for multi-class cardiac point cloud completion. HeartFormer jointly models global cardiac topology and substructure-specific geometric priors within a unified coarse-to-fine refinement framework, enabling anatomically coherent four-chamber reconstruction from sparse and misaligned inputs. We embed HeartFormer within the first completely automated geometric deep learning pipeline for 3D four-chamber cardiac surface reconstruction from routine cine MRI. To facilitate systematic evaluation, we introduce HeartCompv1, the first large-scale public benchmark for multi-class cardiac point cloud completion, comprising 17,000 high-resolution 3D heart models. Extensive cross-domain experiments on HeartCompv1 and UK Biobank demonstrate that HeartFormer achieves state-of-the-art geometric accuracy while preserving clinically meaningful structural fidelity. Code and dataset are available at https://github.com/MultiMeDIA-Oxford/HeartFormer.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0664_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: https://papers.miccai.org/miccai-2026/supp/0664_supp.zip

Link to the Code Repository

https://github.com/MultiMeDIA-Oxford/HeartFormer

Link to the Dataset(s)

HeartCompv1 dataset: https://zenodo.org/records/20721166

BibTex

@InProceedings{MaZhe_HeartFormer_MICCAI2026,
        author = { Ma, Zhengda AND Banerjee, Abhirup},
        title = { { HeartFormer: Semantic-Aware Dual-Structure Transformers for 3D Four-Chamber Cardiac Point Cloud Reconstruction } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16889},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper introduces HeartFormer, a novel semantic-aware dual-structured Transformer that employs a coarse-to-fine framework to reconstruct anatomically coherent four-chamber structures from sparse, misaligned cine MRI data by jointly modeling global topology and substructure geometric priors. Furthermore, the authors embed this model into the first fully automated geometric deep learning pipeline for direct 3D surface reconstruction, and concurrently release HeartCompv1—a pioneering benchmark dataset comprising 17,000 high-resolution 3D cardiac models that addresses a critical gap in systematic evaluation for multi-class point cloud completion.

  • 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.

    Previous studies lacked a large-scale public benchmark specifically designed for multi-class cardiac point cloud completion. The authors constructed a dataset consisting of 17,000 high-resolution 3D heart models. Not only did they train and validate their own HeartFormer model, but the dataset will also be made public after the final version is published, providing opportunities for other researchers to use it.

  • 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.

    This paper proposes the novel HeartFormer model and releases HeartCompv1, a highly valuable, large-scale benchmark dataset. It makes solid academic contributions to the field of fully automated 3D cardiac reconstruction; therefore, I am inclined to recommend a Weak Accept. However, the authors must clarify several key issues in their subsequent response. First, the current reconstruction pipeline appears to rely heavily on upstream segmentation details. The model’s robustness when faced with defective or noisy segmentation inputs requires further validation. Second, the extent to which the large-scale synthetic dataset can faithfully capture the complex distribution and variations of real clinical cine MRI data warrants a more in-depth discussion. Additionally, the authors must carefully proofread the manuscript to correct notational errors in the figures and ensure consistent formatting of mathematical formulas throughout the text. If the authors can convincingly address these core concerns during the rebuttal phase and correct the formatting issues, I would be very willing to consider raising my final score.

  • 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?

    My recommendation is based on the outstanding contributions of this paper in basic research. Specifically, it includes the introduction of the novel HeartFormer model and the release of the HeartCompv1 dataset. A benchmark test set containing 17,000 samples and a fully automated process were created, representing a significant advancement in this field. However, the overall score I give is mainly based on the following key limitations: Although the dataset has a strong guiding effect on subsequent research, the generation of synthetic data relies on strict transformations and Gaussian noise, which may lead to a significant domain gap compared to the complex and non-rigid changes in real clinical MRI films. Secondly, the robustness verification of this framework for defective upstream segmentation is still insufficient.

  • 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 responses have addressed most of my concerns and clarified several important points. While some limitations and open issues still remain, I appreciate the additional explanations and discussions provided by the authors. I also hope the code and dataset can be released in the future to facilitate follow-up research and improve reproducibility. Overall, I am inclined to support acceptance.



Review #2

  • Please describe the contribution of the paper

    This paper addresses multi-class anatomical completion from sparse and potentially misaligned cine MRI segmentations. The main contributions are: 1.The use of transformers for multi-class anatomical completion and design of the “dual-structure” which simultaneously represents global and local keypoints and their respective feature vectors in a semantic-aware manner. 2.The synthetic data generation which comprises point clouds. These pcds represent segmentations of cine MRI and are derived from heart CT.

  • 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.Justified application of hierarchical keypoint completion to multi-structure cardiac anatomy. This means the representation, propagation, and aggregation of global (whole-heart) and local (per-chamber) keypoints and their corresponding features within a transformer-based architecture in a coarse to fine manner for cine MRI completion. Novel is the ability to condition on the semantic information across different stages of the network. 2.Formulation of Semantic-aware Chamfer distance loss. The proposed loss formulation extends standard Chamfer distance by incorporating semantic labels. In this manner, geometrically close but semantically incorrect correspondences cannot be rewarded. This encourages inter-class separation.

  • 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.Overstated claims. The authors claim to present the “first geometric deep learning framework based on point cloud representation for 3D four-chamber cardiac reconstruction from cine MRI”. However, previous cited work PCCN already achieves this. Another work, LC-U-NET[1] (Xu et al. FIMH 2023), also achieves four chamber + valves completion from cine MRI. The latter is not cited. 2.Missing recent cine MRI shape reconstruction STOA in the introduction. The literature review part of the introduction focuses predominantly on generic, computer vision point cloud based shape completion methods and omits multiple STOA that addresses the same task of cine MRI completion such as LC-U-Net[1], ModusGraph[2], CardiacFlow (MICCAI25)[3]. Moreover, it does not place the current paper within the state of the art, neither does it motivate the current gap being addressed. The stated motivation of addressing sparse sampling, inter-slice motion as well as transitioning away from slice-based segmentation to surface paradigm is undermined by previous works which have already addressed this (both cited and uncited). It is not clear what is still missing and why this work is needed. 3.The methodology is not completely described and therefore unclear. Details of components such as SEBlock and GEBlock are missing as well as the second applicaiton of the SA GRT block on the mid point cloud is unclear. [1] Xu, Hao, et al. “Whole heart 3D shape reconstruction from sparse views: leveraging cardiac computed tomography for cardiovascular magnetic resonance.” International Conference on Functional Imaging and Modeling of the Heart. Cham: Springer Nature Switzerland, 2023.[2]Deng, Yu, et al. “ModusGraph: automated 3D and 4D mesh model reconstruction from cine CMR with improved accuracy and efficiency.” International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2023.[3]Ma, Qiang, et al. “CardiacFlow: 3D+ t Four-Chamber Cardiac Shape Completion and Generation via Flow Matching.” International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2025.

  • 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

    Would be interesting to justify and discuss the point cloud representation choice.

  • 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 use of hierarchical keypoint based shape completion and introduction of semantic aware modules are interesting and justified here. However, the motivation and explanation of what previous papers are lacking or failing is non existent and what exactly this paper addresses is unclear. Overclaiming novelty relative to existing literature is as well a big factor for my decision. Extensive comparison to single-class point cloud completion methods is also insufficient, by design these will fail in multi-class settings. Moreover, claiming STOA in cine MRI reconstruction, although comparison with only one prior work that specifically address cine MRI reconstruction. Lastly, the paper could benefit from improvements in methodolgy description which are sometimes unclear.

  • 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.

    I thank the authors for their answers. However, the following concerns remain and are the reason of my final decision.

    • Comparison to STOA: In my opinion, the claim that point cloud representation is superior to voxelgrid-based methods deserves more nuance. Voxelgrid methods such as CardiacFlow naturally capture volumetric topology, achieve competitive reconstruction accuracy and do not require an additional mesh extraction step. A direct quantitative comparison with CardiacFlow remains necessary to support the state-of-the-art claim in four-chamber cardiac reconstruction from cine MRI. Otherwise the proposed method achieves the same task just with another representation. The fact that this comparison is not available significantly weakens the paper. – Overclaiming & missing STOA: PCCN claims that it can be used directly only by changing the number of classes used. If the extension to adapt PCCN to 4 chambers is trivial, this should be clearly stated.
    • The proposed dataset consists of simulated point clouds rather than real cine MRI data. While useful as an initial benchmark, it is not a significant contribution of this paper.
    • Reliance on previous segmentations has not been addressed, since segmentations inaccuracies usually occur spatially structured, rather than through isotropic perturbations. In my opinion, the tested scenario is not clinically realistic.
    • SEBlock and GEBlock still not clarified.



Review #3

  • Please describe the contribution of the paper

    This paper proposes a method to reconstruct six cardiac substructures from sparse, misaligned point clouds obtained from 2D cine MRI, addressing the limitation of prior methods that focus solely on biventricular reconstruction.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The paper is generally well written, and enough experiments and ablation studies are reported to support the claim.

    Paper outperformed several existing baselines.

    The paper has released a dataset comprising 17,000 high-resolution 3D heart models.

  • 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 paper would benefit from reporting model size, computational cost, and inference time, all of which are important for practical deployment.

    The first sentence of sub-section 2.3 can be improved for better readability.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

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

    The paper is well written, clearly motivates the problem, presents a sound methodology and experiments, and demonstrates improved performance compared to existing methods.

  • 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

We wholeheartedly thank the Area Chair and the three Reviewers for their constructive feedback to improve the quality and readability of the manuscript. We will incorporate all suggestions in the revised manuscript.

R1-Heavy reliance on segmentation: We highly appreciate the insightful suggestion. Motivated by this comment, we analyzed the robustness of HeartFormer under defective and noisy input conditions. Specifically, we simulated incomplete point clouds with 10% and 20% missing points and introduced localized perturbations by adding Uniform noise within [-3, 3]mm to 10% and 20% of randomly selected points, mimicking segmentation inaccuracies commonly encountered in clinical practice. Compared with the standard setting (chamfer distance:1.585), HeartFormer maintained stable performance under missing-point conditions (10%:1.588; 20%:1.599) and noisy perturbations (10%:1.611; 20%:1.646), demonstrating its robustness and stability under challenging clinical-like conditions. We will discuss this in the revised manuscript.

R1-Variations of real clinical cine MRI: The synthetic dataset incorporates spatial perturbation and slice misalignment simulation to mimic real clinical cine MRI acquisition, based on six prior studies (Xu et al.,2019; Tarroni et al.,2020). We will include detailed dataset generation protocol in the revised manuscript. The generalizability of HeartFormer to real clinical data is demonstrated through its evaluation on 116 real clinical cine MRI, achieving strong quantitative and qualitative performance. The dataset included subjects aged 47-78.8 years (65.5±7.4), both sexes (M/F:83/33), and diverse cardiovascular conditions, including myocardial infarction (37.1%), hypertension (39.7%), ischemic heart disease (41.4%), atrial fibrillation (13.8%), and heart failure (6.0%), acquired from two imaging centers (Cheadle and Newcastle).

R1-Proofreading: We have very thoroughly proofread the manuscript to rectify notational inconsistencies (e.g., consistent notation for $F_{glo}$, among others).

R2-Overstated & Incomplete Literature Review: We agree that our original statement was overstated and will explicitly discuss recent cine MRI shape reconstruction methods, including LC-U-Net[1], ModusGraph[2], and CardiacFlow[3], in the revised Introduction. Specifically, PCCN focuses only on biventricular reconstruction rather than four-chamber cardiac reconstruction, whereas both LC-U-Net[1] and CardiacFlow[3] perform whole-heart completion using a voxelgrid-based framework. In contrast, the proposed HeartFormer performs semantic whole-heart reconstruction based on efficient point cloud representation, preserving the whole anatomical details without considering the empty spaces in voxelgrid representation and thus facilitating higher surface resolution with flexible topology while training on limited computational resources and fast inference. In contrast to ModusGraph[2], our method does not require predefined mesh topology or template deformation.

For fair comparison, we primarily evaluated against SOTA point cloud completion methods. Single-class methods were evaluated only in single-class (whole-surface) setting (Table 1), while the multi-class PCCN was evaluated in the multi-class setting (Table 2). The key contribution in the proposed HeartFormer is high-resolution, topology-flexible semantic surface completion from sparse cine MRI using point cloud representation.

R2-Methodology Details: Due to page limit, we could not include detailed descriptions of SEBlock and GEBlock modules in the original manuscript. We will include their architectural details in the revised manuscript. The midpoint cloud follows the same processing pipeline as the coarse point cloud, which is second application of the SA-GRT.

R3-HeartFormer contains 5.688M parameters with 87.036 GFLOPs, and achieves an average inference time of 0.0358s per sample on a single GPU. The first sentence of Section 2.3 has been revised to improve readability.



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.

    All three reviewers acknowledged that this paper does have some merits, particularly a new dataset including 17000 high-resolution cardiac models that addresses a critical gap in systematic evaluation and multi-class point cloud completion.

    However, the reviewers also raised some major concerns. More discussions are needed to clarify (1) if the reconstruction performance heavily rely on the upstream segmentation models, (2) if the large-scale synthetic dataset can faithfully capture the complex distribution and variations of real clinical cine MRI data, (3) some methodological and experimental details, such as the SEBlock and the GEBlock, the SA GRT block, and the model size and computation cost as well as the inference time. In addition, reviewers also suggest that some relevant works are missing in the literature review. Besides, reviewers also raised some problems in formatting and mathematical formulas.

    In the rebuttal, please carefully study the comments raised by the reviewers, and provide responses to them, particularly above-mentioned major concerns.

  • 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.

    After rebuttal, two reviewers suggest accepting this paper while the other reviewer still have some concerns. Based on the rebuttal as well as the recommendations of the two reviewers, the paper can be accepted.

    It is suggested that the authors sufficiently consider the comments raised by the reviewers and further improve the paper 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.

    Reject

  • Please justify your recommendation.

    Although the rebuttal raised R1 to accept, R2 remained unconvinced and maintained reject post-rebuttal, noting the absence of a direct quantitative comparison to voxel-based CardiacFlow that is essential to substantiate the SOTA claim, the synthetic-only nature of the proposed dataset, and the unaddressed clinically-unrealistic segmentation-robustness setup. Given that the reviewer’s fundamental concerns went unresolved, the recommendation is therefore Reject.



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 paper have mixed score after rebuttal. Two reviewers have indicated that the rebuttal has successfully addressed the majority of their concerns. To address the concern about the overclaim raised by another reviewer, the author has promised to soften it in the final version. So I suggest accepting the paper.



back to top