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

Abstract

Learning 3D+t shape completion and generation from multi-view cardiac magnetic resonance (CMR) images requires a large amount of high-resolution 3D whole-heart segmentations (WHS) to capture shape priors. In this work, we leverage flow matching techniques to learn deep generative flows for augmentation, completion, and generation of 3D+t shapes of four cardiac chambers represented implicitly by segmentations. Firstly, we introduce a latent rectified flow to generate 3D cardiac shapes for data augmentation, learned from a limited number of 3D WHS data. Then, a label completion network is trained on both real and synthetic data to reconstruct 3D+t shapes from sparse multi-view CMR segmentations. Lastly, we propose CardiacFlow, a novel one-step generative flow model for efficient 3D+t four-chamber cardiac shape generation, conditioned on the periodic Gaussian kernel encoding of time frames. The experiments on the WHS datasets demonstrate that flow-based data augmentation reduces geometric errors by 16% in 3D shape completion. The evaluation on the UK Biobank dataset validates that CardiacFlow achieves superior generation quality and periodic consistency compared to existing baselines.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1449_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/m-qiang/CardiacFlow

Link to the Dataset(s)

MM-WHS dataset: https://zmiclab.github.io/zxh/0/mmwhs/ WHS++ dataset: https://www.zmic.org.cn/care_2024/track5/

BibTex

@InProceedings{MaQia_CardiacFlow_MICCAI2025,
        author = { Ma, Qiang and Meng, Qingjie and Qiao, Mengyun and Matthews, Paul M. and O’Regan, Declan P. and Bai, Wenjia},
        title = { { CardiacFlow: 3D+t Four-Chamber Cardiac Shape Completion and Generation via Flow Matching } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15961},
        month = {September},
        page = {89 -- 99}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The authors present a novel method for performing 3D+t four-chamber cardiac shape completion and generation through flow matching. The incorporation of a generative modeling component facilitates data augmentation, which is particularly beneficial when working with limited datasets, ultimately enhancing the shape completion performance. Additionally, the temporal encoding aspect of the model ensures that the generated shapes maintain periodic and clinical consistency.

  • 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 methodology is clearly articulated and the writing is well-structured, making it easy for the reader to follow.
    2. The use of latent rectified flows in the latent space to model the temporal relationships between frames is an innovative approach.
    3. The method establishes a robust generative modeling framework for both shape completion and generation.
  • 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 rationale for employing periodic Gaussian kernel encoding of time frames is not clearly explained.
    2. The generative modeling aspect of the framework requires more rigorous testing. While Table 1 shows improved performance from a single run, this evidence alone is insufficient to attribute the performance gain solely to data augmentation.
    3. The description of the dataset is somewhat unclear. Specifically, does the dataset include Whole Heart Segmentation (WHS) for the entire cardiac cycle for each patient? Are the time steps consistent across all patients?
    4. The paper lacks a thorough explanation of 3D+t shape completion and generation, and there is no discussion of the clinical relevance of the results.
    5. It would be beneficial to address expected shape changes throughout the cardiac cycle, volume variations measured by segmentations of the structures, and comparisons between diastolic and systolic phases. There is insufficient discussion regarding potential clinical use cases for the proposed methods.
    6. The paper lacks a literature review of current methods used for shape completion and generation of 3D+t cardiac structures, and the proposed method is not compared to existing alternatives.
  • 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.

  • 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/2025/en/REVIEWER-GUIDELINES.html
    1. It would enhance the paper to include additional experiments or a written justification, supported by references, regarding the choice of periodic Gaussian kernel encoding for time frames.
    2. In reference to section “Learnable Initial Value,” the explanations provided, along with Figures 1 and 2, could benefit from clarification. Can the authors justify the design choice of making z0 learnable, rather than using the autoencoder’s latent space as a data-informed latent representation for the sample at the initial time point (t=0)? Subsequently, the Latent Rectified Flow (LRF) could be employed to estimate the latent representations for subsequent time steps.
    3. Additional points can be found in the weaknesses section.
  • 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 — could be accepted, dependent on rebuttal

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

    Overall, while the paper presents a promising approach to 3D+t cardiac shape completion and generation, it requires further validation, clarity in methodology, and a more detailed exploration of clinical relevance and comparative analysis with existing methods to warrant a stronger acceptance.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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

    A diffusion model to generate cardiac shapes along the cardiac cycle, trained on whole heart segmentations from CT, which is later used to generate shapes from sparse (MRI) data.

  • 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.
    • Using multimodality data to overcome the limitations of MRI is clever.
    • The authors compare their method with several methods of the state of the art
  • 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.

    Overall, it is a good paper, with strong experiments. The weakest experiment is the shape completion for MRI.

    • It is not clear to me whether the original images are used as input to the model, or the segmentation masks. I

    • Just evaluating the volume curves (Fig 6). is not enough, and it seems that there is little variability in terms of timing (position of the ES , diastasis etc). The STD seems constant along the cardiac cycle, which seems not very reliable.

  • 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.

  • 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/2025/en/REVIEWER-GUIDELINES.html

    • When describing the experiments, it is confusing to keep track of which kind of data (CT or MRI) is used for evaluation and testing. Please, clarify.

    • For the MRI shape completion, missing a comparison with the classical approach of fitting 2D segmentations to a 3D atlas.

    • When doing statistics from bootstraps, the samples are not independent, which makes p-values appear more relevnat than they actually are (Table 1)

  • 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?

    Good experiments, solid methodology, interesting application.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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 #3

  • Please describe the contribution of the paper

    The authors describe a “flow-matching” (ODE) technique to extract the shapes of the four chambers of the heart from cine MRI. This is applied to a large dataset from the UK Biobank.

  • 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.

    Excellent. The paper is well written, uses a relatively new framework (2022) to segment the heart and applies it to a large dataset. Almost every aspect of the methodology is clearly justified and analysed. Contains some really clever tricks as well to accelerate output generation.

    Particularly I enjoyed the treatment of time and sampling thereof.

  • 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 is very dense and hard to follow at times; there is a lot to unpack. However, the figures help digest some of the content. I think this will easily be fixable in a full journal. Had to zoom hard on some of the figures but I appreciate the effort put into them!

    (1) Would the superior results presented in Table 1 –while statistically significant– move the needle clinically? Please comment.

  • 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.

  • 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/2025/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?

    This is clever methodology that is well illustrated and easily applicable to other parts of the body. I think the audience for this paper is well beyond cardiac. The way in which the authors have approached data training, management of time, optimisation, etc. is a great example of how things should be done, in my opinion.

    Sure, it will require further clinical validation, but it’s a very strong foundational and clean method worth pursuing in other areas.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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



Author Feedback

We thank all the reviewers for their valuable comments, constructive suggestions, and recognition of our work! We provide detailed responses below to address the reviewers’ comments.

  • Method Description (R3) In Section 2.1, we introduce the flow matching algorithm and the latent rectified flow for data augmentation of 3D cardiac shapes. The 3D+t cardiac shape completion and generation are described in Section 2.2 and 2.3 respectively.

  • Choice of Periodic Gaussian Kernel Encoding (R3) As described in Section 2.3, since the heart should have a consistent shape at the start and end of a cardiac cycle, we introduce the periodic Gaussian kernel (PGK) encoding of time frame to guarantee the periodic consistency and temporal smoothness of synthetic 3D+t shapes. The effectiveness of PGK encoding is validated and reported in Fig. 6. The cycle-DSC score is reduced from 0.981 to 0.959 without PGK encoding. We will explain this clearly in the camera-ready version.

  • Learnable Initial Value (R3) The learnable initial value enables our CardiacFlow framework to learn a straight optimal transport path from noises to latent vectors for one-step generation. Although the latent rectified flow (LRF) could be employed to estimate the latent vectors of subsequent time frames based on previous time steps, it results in accumulated errors across time frames in our experiments.

  • Dataset Information (R1,R3) In this work, we only use segmentation masks as input to represent the shapes of cardiac four chambers. The original MRI and CT images are not used. For 3D cardiac shape completion, we collect 160 3D whole-heart segmentation (WHS) maps of 23 MRI and 137 CT images from multiple data sources [14,21,33,39], acquired at different time points in the cardiac cycle. We will elaborate the dataset information in the final version.

  • Evaluation on Data Augmentation (R1,R3) For 3D cardiac shape generation and completion, we mainly evaluate if LRF-based data augmentation could improve the accuracy of the shape completion network. The results in Table 1 and Fig. 3 show that the 3D cardiac shapes generated by the LRF significantly improve the performance of shape completion. Due to the limited number of data samples, these results are evaluated using a bootstrap approach, which creates 480 random test samples from 48 test subjects. We will consider cross-validation to further verify our results in future work.

  • Evaluation on Generation Quality (R3) In addition to data augmentation, we have also evaluated the generation quality of the LRF by computing the FID between synthetic and ground truth cardiac shapes. The LRF with T=100 steps achieves FID=0.0065, which is much better compared to the VAE (FID=0.037).

  • Evaluation on 3D+t Cardiac Shape Generation (R1,R3) For 3D+t cardiac shape generation, we have compared CardiacFlow to state-of-the-art methods including CHeart [25], ST-NDF [31], and vanilla LRF [5]. The vFID and cycle-DSC scores reported in Fig. 4 demonstrate the superior performance of CardiacFlow. The ablation study and visualization of volume curves are provided in Fig. 6.

  • Clinical Application (R2,R3) In this work, we focus on the methodology of 3D+t four-chamber cardiac shape completion and generation using flow matching. In future work, we plan to extend CardiacFlow for clinical validation on cardiac imaging datasets with diseases, such as HCM, DCM and PAH, to generate plausible shape models of the whole heart for shape analysis, imaging-based biomarker discovery, and digital twin modelling. We also plan to explore disease progression and prognosis by conditional generation.



Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    This paper presents CardiacFlow, a generative framework for 3D+t cardiac shape completion and generation using flow matching. Trained on whole-heart segmentations from CT and applied to sparse MRI data, the method combines latent rectified flows and periodic temporal encoding to model the dynamics of all four cardiac chambers. The approach is evaluated on a large UK Biobank dataset and demonstrates strong performance in both shape generation and completion tasks. The reviewers commend the paper for its solid methodology, clear innovation in the use of flow matching for temporal shape modelling, and thoughtful use of multimodal data to overcome limitations of MRI. The work is well-motivated and technically sound, and the experimental design is comprehensive. While the clinical relevance of the reported improvements could be further discussed, and some methodological elements (as mentioned in detail in reviewers’ comments) would benefit from clearer exposition, these are minor issues in an otherwise strong submission.



back to top