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

Abstract

Accurate cardiac anatomy modeling requires the model to be able to handle intricate interrelations among structures. In this paper, we propose VecHeart, a unified framework for holistic reconstruction and generation of four-chamber cardiac structures. To overcome the limitations of current feed-forward implicit methods, specifically their restriction to single-object modeling and their neglect of inter-part correlations, we introduce Hybrid Part Transformer, which leverages part-specific learnable queries and interleaved attention to capture complex inter-chamber dependencies. Furthermore, we propose Anatomical Completion Masking and Modality Alignment strategies, enabling the model to infer complete four-chamber structures from partial, sparse, or noisy observations, even when certain anatomical parts are entirely missing. VecHeart also seamlessly extends to 3D+t dynamic mesh sequence generation, demonstrating exceptional versatility. Experiments show that our method achieves state-of-the-art performance, maintaining high-fidelity reconstruction across diverse challenging scenarios. Code is available at https://github.com/Scalsol/VecHeart.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0081_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Scalsol/VecHeart

Link to the Dataset(s)

N/A

BibTex

@InProceedings{CheYih_VecHeart_MICCAI2026,
        author = { Chen, Yihong AND Fua, Pascal},
        title = { { VecHeart: Holistic Four-Chamber Cardiac Anatomy Modeling via Hybrid VecSets } },
        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 study presents VecHeart, an implicit neural representation framework for the reconstruction and generation of cardiac four-chamber structures. A Hybrid Part Transformer architecture is developed to capture the intra/inter-part correlations and interactions among four chambers. The Anatomical Completion Masking and Modality Alignment techniques are introduced to tackle four-chamber shape reconstruction from incomplete point cloud with missing parts and from sparse contours. The performance is evaluated on both synthetic data and a wide range of public 3D cardiac CT/MR and cine CMR datasets.

  • 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.
    • This work proposes a novel implicit cardiac four-chamber shape reconstruction framework, enabling effective four-chamber shape completion from sparse or partially observed input.
    • The proposed VecHeart framework achieves state-of-the-art performance across multiple cardiac four-chamber shape reconstruction and generation tasks.
  • 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.
    • It would be interesting to explore if contrastive learning approaches could further improve modality alignment in the latent space.
    • Fig. 1 is somewhat small, which may affect its clarity and 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

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

    N/A

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

    N/A

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

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

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

    Overall, this is a high-quality paper with novel technical contributions and comprehensive experiments. It achieves state-of-the-art performance in multiple cardiac four-chamber shape reconstruction and generation tasks.

  • 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.

    I recommend acceptance based on the technical novelty and methodological contributions of this paper.



Review #2

  • Please describe the contribution of the paper

    The paper proposes VecHeart, a hybrid part transformer model for reconstructing four-chamber cardiac anatomy from sparse cardiac magnetic resonance-derived contours. The authors trained the model to be robust to sparsity, using additional components for anatomical masking and modality alignment. The model was trained on CT-derived reconstructions and evaluated on CT and sparse MRI settings, reporting improved reconstruction accuracy over baselines across chambers and under inputs missing anatomical structures and MRI sparsity.

  • 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.

    Strong improvement over baselines in the reported quantitative evaluation.

    Interesting methodological contribution: a conditioning approach intended to mitigate sparsity and variability in input configurations.

    The methods section appears to include sufficient detail to reproduce the approach, at least at a high level.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    Incomplete-data / missingness evaluation is not well-motivated clinically. As the authors state in the introduction, in typical CMR protocols, the information content for the atria is fundamentally limited: SAX stacks often do not cover the atria well, and LAX views provide only sparse 2D delineations. Given this, it is unclear what evidence supports the claim that there is sufficient patient-specific information to reconstruct detailed atrial geometry. In practice, completion in these regions is likely dominated by learned priors and correlations with other structures.

    Insufficient engagement with the existing cardiac reconstruction literature. The introduction and related work do not adequately situate the contribution within recent work on (i) multi-class point-cloud based four-chamber reconstruction/completion (e.g. HeartFormer: Semantic-Aware Dual-Structure Transformers for 3D Four-Chamber Cardiac Point Cloud Reconstruction, arXiv:2512.00264) and (ii) anatomically/topologically consistent mesh reconstruction constraints (e.g. Topology-Preserving Loss for Accurate and Anatomically Consistent Cardiac Mesh Reconstruction, arXiv:2503.07874). This weakens the novelty claim and makes it harder to interpret what is genuinely new beyond architectural choices.

    Intended use is too unfocused to evaluate clinical relevance. If the goal is shape completion for downstream simulation (e.g., finite element modelling), then a strong learned prior could be appropriate. If the goal is patient-specific anatomical measurement/analysis from sparse CMR, then hallucinated structure may be misleading. The manuscript does not clearly commit to a particular use case; there is only one comment in the introduction that “These constraints severely limit their clinical utility,” but it does not elaborate further on the intended clinical use.

    Residual atrial errors appear too large for realistic use (for measurement). Even if the method improves over baselines, the reported LA/RA CD errors under missingness (e.g., ~9.2 mm and ~7.9 mm) are far above typical MRI voxel sizes (~2 mm) and likely beyond what is usable for patient-specific quantification.

    Generation / anatomical realism is not addressed. A major concern in anatomy generation is concordance with known clinical distributions and anatomical variability; the paper does not meaningfully discuss or evaluate this.

    Overall, the organisation and writing of the paper would benefit from significant improvement, which might clarify some of the issues mentioned.

  • 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 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

    The paper has an interesting methodological contribution, but the following areas require improvement.

    Clarify the problem statement and intended use case. Name the downstream clinical or technical task (or set of tasks) that motivates why a four‑chamber reconstruction from sparse MRI is needed, and what “utility” means in that setting.

    Add a clinical-use disclaimer and interpretation guidance. Explicitly state whether reconstructed anatomy in poorly observed regions (e.g., atria) is intended for patient-specific measurement, or for producing anatomically plausible shapes for downstream modelling (e.g., FE).

    Reframe and justify the incomplete-data evaluation. Since the model is conditioned on sparse point clouds (downstream of contour extraction), explicitly define what “missingness” corresponds to in practice (e.g., segmentation dropout, limited coverage, low-resolution atrial signal), and why completing that missingness is useful for the stated task.

    Make the evaluation match the claims. If the claim is clinical utility, consider adding an evaluation tied to a downstream analysis, or provide strong justification and calibration showing that reconstruction accuracy is within clinically relevant tolerances.

    Improve figures. Increase figure sizes, add zoom-ins for regions where differences matter, and ensure qualitative comparisons are legible.

    Minor issues (non-exhaustive): grammar and typos, and some statements are too general or too vague to evaluate. E.g. Section 1.“Accurate cardiac structure modeling is vital for application like clinical diagnosis…” → applications . “…limit their clinical utility.” In what task?

  • 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?

    While the paper appears to make a promising methodological contribution, the problem formulation and clinical motivation are not yet sufficiently clear. In particular, the purpose and evaluation of sparse atrial reconstruction does not convincingly support the implied claims of clinical utility. Overall, the work would benefit from a tighter framing and an evaluation that is more explicitly aligned with a concrete downstream use case. These issues are substantial and unlikely to be resolved within the rebuttal period.

  • 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.

    Rebuttals addressed most of my major concerns effectively.



Review #3

  • Please describe the contribution of the paper

    First this article extend from the single structure latent representation into a four-chamber heart geometry reconstructions to enhance the attention mechanism of each chamber and develop a full input or sparse inputs reconstruction problems. Second this article introductions some parts to multichambered and incomplete inputs for achieving the completion masking tasks that they propose.

  • 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.

    First, this article design the method by introducing the correlation of multichambered correlations and the they can captures the motion correlation between each part of the heart. Second, this article focus on the complete inputs and 3d+t generations and it seems reasonable to me to include these tasks since this is related to an issues that sometimes might encounter. Third, this article have some graphs to help understanding the detail of the articles as well.

  • 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.

    First, in [Table 2] there is a systematic gap in validation. Author only reports the synthetic SAX and LAX reconstruction results of LA and RA, however for its “Real Dataset”, it didn’t show the results of LA and RA. This will inevitably leads to a validation error since synthetic LAX usually use displacement algorithm. Usually in real LAX and SAX dataset, there is always breath noises and image quality issue in LAX images, so I think using only synthetic LAX is lack of solid validation of model performance. Second, the way author used to get ground-truth meshes are likely to may lead to proxy reference. In the article they stated that “we register and fit a Statistic Shape Model to segmentation masks predicted by a publicly available method to generate ground-truth cardiac meshes.” However, this leads to inherent biases that ground-truth meshes uses the SSM model, which is a statistical model that will introduce model-based proxy. The SSM model has a potential generating bias that ground-truth meshes might be embedded in template-like anatomy. The results of using these generated ground-truth meshes lead to weakening this article’s statement of model results comparison. Third, in the ablation study, the article ablation still have a validation gap between their results and their conclusions. In the article, the authors states that “ VecHeart outperforms the baselines in both complete and missing data scenarios, validating our HPT and ACM designs”.However, the ablation experiment in “Table 1” and “Table 2” does not contain the design ablation for “HPT”, “ACM”. If the author still simultaneously contains these three parts together in the ablation study, then it will be difficult to see whether individual module are effective enough. Fourth, In the quantitative analysis part (Figure 2) the author only provide results for 1 or 2 samples. However, usually for the quantitative analysis it is important to provide the worst, medium, best result to demonstrate the generalizability of the model. It is especially in this type of missing part and SAX-only conditions, since the model received sparse information input. If the author only demonstrated selected cases, then there is a potential of cherry-picking result, which leads to gaps in validation examinations. Fifth, for the (3D+t) tasks, there is a validation gap of using only two metrics (vFID and Cycle-CD). The first reaons is that these metrics is indirect to clinical application, some more clinical-related metrics should be used such as ejection fraction and volume curve of the heart. Furthermore, the author should also compared with other published 3D+t reconstruction pipeline.

  • Please rate the clarity and organization of this paper

    Satisfactory

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

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

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

    N/A

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

    First, I am not sure how Table 1 is coded in the original Latex file, but the MICCAI 2026 template states that “If you have tables, the smallest font size allowed is 8pt.” I compared this against the front size produced using the 8pts in the templated, and the numerical number in the Table 1 appears to be much smaller compared to 8pts. I am therefore unsure whether the authors are following the MICCAI 2026 formatting requirements. If not, this may be unfair to other authors who strictly adhered to the formatting rules, as they would be more constrained by space and therefore unable to include some extra amount of their data or numerical results. In addition, the extremely small font size makes reading and comparing much more difficult as number are dense and compact.

  • 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 article provide the VecHeart this new model and focus on a reasonable clinical issue. Key factor for this score is that the ablation experiments still have a experimental gap and also the ground-truth meshes still embedded with potential proxy bias. Furthermore, there is a lack of the independent ablation of each innovative components and sensitive experiments of different loss component in their result.

  • 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.

    N/A



Author Feedback

[R1] We appreciate the advice and will explore it. [R2] Formulation, motivation: Our motivation, as detailed in the first two paragraphs of the Intro, is to propose a better multi-part 3D latent representation to address the limitations of prior implicit methods: restriction to single-part modeling or reliance on complete surface and slow test-time optimization. Accurate multi-part modeling is important and a prerequisite for downstream tasks like index analysis and cardiac simulation. By introducing HPT, ACM, MA, VecHeart achieves SOTA performance across widely studied benchmarks, including better representation capacity with faster inference (Tab 1 Top) and better visible part reconstruction from sparse slices (Tab 2). “limit their clinical utility” means prior implicit methods’ performance and speed limitations on the fundamental shape representation problem. Regarding the learned prior: even standard SAX MRI is sparse and necessitates anatomical priors. Incorporating prior is common in SOTA methods [19,32,33,34,37]. [R2] Missingness Evaluation: The missingness experiment is designed to probe the model’s ability in extreme scenarios without requiring architecture modification, rather than to show immediate clinical relevance. This evaluation highlights the model’s ability to produce plausible outputs on unobserved parts under severe data constraints, rather than representing its only intended use. We acknowledge the reconstruction of unobserved anatomy is not perfect, but this shows VecHeart’s potential for application like conditioned generation (e.g., via volume or pathology). The model’s SOTA reconstruction performance on visible parts, which is a well-established and standard setting, shouldn’t be overshadowed by its currently imperfect handling of extreme tasks, where it shows potential despite not being fully mature. We shouldn’t be penalized for exhibiting new abilities absent in previous methods. Regarding “usable”: CD is computed as the standard sum of pred-to-gt and gt-to-pred squared L2 distances, and is thus not directly comparable to the 2mm threshold. Using the comparable mean surface distance instead yields 3.02/2.51 mm, indicating strong potential despite these parts being completely missing. [R2] Literature: Requiring comparison with unpublished preprints is inappropriate, especially HeartFormer was released only 3 months before deadline. Moreover, given the fact that we do things these unpublished methods do not fully cover, we don’t see how they diminish our work’s novelty. [R2,R3] 3D+t Generation Experiment: We clarify that this experiment evaluates mesh sequence generation via diffusion model, not reconstruction (3D+t reconstruction is in Tab 2, Real SAX). Following the setting of CardiacFlow, the current SOTA for 3D+t generation, VecHeart shows better volume preservation and periodic consistency. We agree by including analysis with clinical-related distributions and metrics can further strengthen the evaluation. [R3] Real Evaluation of LA/RA: It’s impossible to evaluate 3D LA/RA quality as we can’t get 3D GT from LAX, which is just a 2D image. Also, the real dataset we use only has SAX. Thus synthetic evaluation is the only viable proxy for 3D LA/RA accuracy. [R3] GT mesh: 1060 shapes are directly derived from expert CT annotations, ensuring no proxy bias. SSM is only used to obtain bi-ventricle GT for real SAX MRI. Our evaluation against SSM GT is zero-shot (no SSM bias during training), yet we still achieve SOTA. Ours-finetune proves adapting to SSM bias indeed improves metrics. [R3] Ablation Study: Ours-Sep serves as an ablation for HPT, proving modeling inter-part relationship is beneficial. Excluding ACM can keep performance on visible parts, but it won’t produce any meaningful results on missing parts, which we mention in the main text as: “cant extend to missing-input”. We will clarify. [R3] Visual Results: Due to space limit, more visualization will be provided in a possible journal extension.



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.

    I totally agree with the reviewer #2 that the motivation/fomulation of this study is not quite convincing. The authors should explain this clearly in the rebuttal stage.

  • 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 motivation is still not fully convincing to me. The paper claims that multi-part latent representation improves clinical utility, but the link between reconstruction accuracy and downstream clinical tasks such as index analysis or cardiac simulation is not sufficiently demonstrated. Therefore, the work appears more like an incremental improvement over existing implicit/shape-prior methods rather than a clearly motivated methodological advance.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The rebuttal clarified the motivation and intended use as a multi-part latent representation rather than a clinical end-application, reframed the missingness study as a probe of new capabilities, and resolved the evaluation concerns, leading R2 to move from a reject to accept while R1 maintained acceptance. On balance, the methodological contribution is good enough, and I therefore recommend acceptance.



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.

    Reviewers generally agreed that the proposed VecHeart framework provides meaningful methodological contributions, particularly in modeling multi-part cardiac structures under sparse and incomplete input settings, and achieves strong performance across multiple reconstruction and generation benchmarks. The main concerns were related to the clinical motivation and interpretation of the missingness experiments, the evaluation of anatomically unobserved regions, and the clarity of the intended downstream use case. The rebuttal clarified that the missingness setting is primarily designed to evaluate representation robustness and conditional completion ability rather than immediate clinical deployment, and also addressed concerns regarding proxy bias, ablation design, and the role of anatomical priors.



back to top