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

Abstract

3D Cone-Beam CT (CBCT) is widely used in radiotherapy but suffers from motion artifacts due to breathing. A common clinical approach mitigates this by sorting projections into respiratory phases and reconstructing images per phase, but this does not account for breathing variability. Dynamic CBCT instead reconstructs images at each projection, capturing continuous motion without phase sorting. Recent advancements in 4D Gaussian Splatting (4DGS) offer powerful tools for modeling dynamic scenes, yet their application to dynamic CBCT remains underexplored. Existing 4DGS methods, such as HexPlane, use implicit motion representations, which are computationally expensive. While explicit low-rank motion models have been proposed, they lack spatial regularization, leading to inconsistencies in Gaussian motion. To address these limitations, we introduce a free-form deformation (FFD)-based spatial basis function and a deformation-informed framework that enforces consistency by coupling the temporal evolution of Gaussian’s mean position, scale, and rotation under a unified deformation field. We evaluate our approach on six CBCT datasets, demonstrating superior image quality with a 6× speedup over HexPlane. These results highlight the potential of deformation-informed 4DGS for efficient, motion-compensated CBCT reconstruction. The code is available at https://github.com/Yuliang-Huang/DIGS.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3968_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Yuliang-Huang/DIGS

Link to the Dataset(s)

Code to simulate CBCT datasets from clinical 4DCT will be provided at: https://github.com/Yuliang-Huang/DIGS

BibTex

@InProceedings{HuaYul_DIGS_MICCAI2025,
        author = { Huang, Yuliang and Singh, Imraj and Joyce, Thomas and Thielemans, Kris and McClelland, Jamie R.},
        title = { { DIGS: Dynamic CBCT Reconstruction using Deformation-Informed 4D Gaussian Splatting and a Low-Rank Free-Form Deformation Model } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15963},
        month = {September},
        page = {132 -- 142}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    In my view, the paper introduces a novel method called Deformation-Informed Gaussian Splatting for 4D Cone-Beam CT reconstruction. The primary technical contribution is a deformation-driven 4D Gaussian Splatting (4DGS) approach for dynamic CBCT reconstruction, which features two key innovations:

    1. Representation of Motion: The authors decompose the deformation field into a linear combination of spatial and temporal basis functions. Specifically, they use: a. A B-spline control point grid to parameterize spatial basis functions; b. A 1D B-spline to parameterize temporal basis functions;
    2. Gaussian Property Transformation: The method transforms Gaussian kernel properties through the deformation field: a. Mean position of Gaussian kernels is directly moved by the deformation field; b. Scale and rotation are derived from the Jacobian matrix of the deformation field; c. Simultaneous optimization of Gaussian attributes and spatio-temporal B-spline grid; Experimental validation on six synthetic CT datasets demonstrated the method’s effectiveness, with average PSNR significantly outperforming comparative methods like Supremo and HexPlane. Ablation studies revealed that removing the deformation information framework resulted in minimal performance improvement, while eliminating the free-form deformation representation led to a substantial performance decline.
  • 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 paper presents two strengths:

    1. the authors present a timely and innovative problem to achieve 4D medical image reconstruction with 4D GS.
    2. the authors propose a novel approach to representing GS parameters’ change through an explicit deformation field, which distinguishes their method from existing techniques.
  • 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 exhibits several significant weaknesses.

    1. the authors claim that their unified deformation-informed framework is a key innovation for ensuring consistent changes in Gaussian attributes (position, scale, rotation). However, the ablation study paradoxically reveals that removing this framework actually improved performance, casting doubt on the proposed method’s fundamental contribution.
    2. the explicitly defined deformation field fails to fully address the clinical problem of image space deformation tracking, as the Gaussian deformation field is not directly equivalent to voxel deformation. While the B-spline deformation is explicit, the corresponding voxel deformation remains challenging to obtain, since voxels are rendered through Gaussian parameters.
    3. the experimental validation is limited, with only six synthetic datasets used, leaving the method’s performance on real clinical data entirely unverified.
    4. the paper suffers from poor presentation and insufficient details. The manuscript contains only one original result image, which impedes reader comprehension, and some methodological details remain unclear, such as the specific model input configurations and the rationale behind hyperparameter selections. These limitations significantly undermine the paper’s scientific rigor and potential clinical applicability.
  • 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.

  • 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

    The authors have proposed several novel and intriguing technical questions. However, the current research work remains incomplete. Specifically, the experimental results do not fully align with the stated motivation, and the paper would benefit from improved presentation to enhance reader comprehension. To address these concerns, the authors should focus on more rigorous experimental validation, clarify the connection between their proposed methodology and research objectives, and develop a more transparent and detailed manuscript that effectively communicates their innovative approach.

  • 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 paper exhibits several significant weaknesses.

    1. the authors claim that their unified deformation-informed framework is a key innovation for ensuring consistent changes in Gaussian attributes (position, scale, rotation). However, the ablation study paradoxically reveals that removing this framework actually improved performance, casting doubt on the proposed method’s fundamental contribution.
    2. the explicitly defined deformation field fails to fully address the clinical problem of image space deformation tracking, as the Gaussian deformation field is not directly equivalent to voxel deformation. While the B-spline deformation is explicit, the corresponding voxel deformation remains challenging to obtain, since voxels are rendered through Gaussian parameters.
    3. the experimental validation is limited, with only six synthetic datasets used, leaving the method’s performance on real clinical data entirely unverified.
    4. the paper suffers from poor presentation and insufficient details. The manuscript contains only one original result image, which impedes reader comprehension, and some methodological details remain unclear, such as the specific model input configurations and the rationale behind hyperparameter selections. These limitations significantly undermine the paper’s scientific rigor and potential clinical applicability.
  • 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.

    My primary concern has not been satisfactorily addressed. The conceptualization of the B-spline DVF as being independent of the Gaussian representation is controversial. The manuscript itself states that the parameters for the B-spline DVF and the Gaussian attributes are jointly optimized. This joint optimization implies a strong coupling, making it difficult to argue for the DVF’s independence from the specific Gaussian entities it deforms during the learning process. If the authors maintain that the learned DVF accurately reflects the underlying physical deformation (and is thus applicable beyond the specific Gaussian instances), they should provide visualizations of the deformation field itself, applied to representative anatomical structures or a reference CT, to substantiate this claim of consistency and accuracy.



Review #2

  • Please describe the contribution of the paper

    The manuscript presents a novel approach for dynamic CBCT reconstruction using 4D Gaussian splatting, with a strong focus on motion modeling via a deformation field.

    The method is evaluated on two datasets: the XCAT phantom and real patient data from the 4DLung dataset. Performance is compared against a classical reconstruction approach that estimates respiratory motion from projections and applies motion-compensated FDK (MC-FDK) and a HexPlane-based 4D Gaussian splatting method.

    Ablation studies explore the effects of omitting either the coupled deformation or the regular grid points, but there is no experiment removing both. The proposed approach achieves better quantitative and qualitative results compared to the baselines while also outperforming the existing Gaussian splatting method in terms of computational efficiency.

  • 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. Relevance
      • The study addresses an important area of research, contributing to advancements in dynamic imaging and motion-compensated reconstruction.
    2. Promising Qualitative Results
      • While some aspects of the quantitative evaluation could be improved, the qualitative results appear promising and support the effectiveness of the proposed approach.
    3. Methodological Interest
      • The method is designed with a focus on explainability and physical robustness, which is an important consideration in medical imaging applications.
    4. Ablation Studies
      • The inclusion of ablation studies provides insights into the contributions of different components of the method, helping to understand their impact.
    5. Commitment to Open Science
      • The promise of making the code publicly available after publication enhances reproducibility and potential future contributions to the field.
    6. Diverse Datasets
      • The use of two types of datasets (XCAT and simulated 4D CT) adds value by demonstrating the method’s applicability across different data sources.
  • 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. Joint Optimization
      • The manuscript does not clearly explain how the joint optimization is performed. Including an algorithm or a schematic would greatly improve clarity.
    2. Weak Introduction
      • The distinction between dynamic CBCT and 4D CBCT is unclear. The authors state that dynamic CT reconstructs at each projection without respiratory phases but then introduce prior-based approaches and cite a 4D CBCT paper from 2012, creating confusion.
      • The references are poorly chosen. Some are not fitting, others are insufficient, and some are placed inappropriately.
    3. Language and Clarity
      • The manuscript contains numerous grammar and spelling mistakes, along with debatable word choices, which hinder readability.
    4. Lack of Novelty
      • The proposed method appears to be a combination of two existing approaches without substantial innovation.
      • The coupling method for deformation is not well-motivated. At the very least, the authors should reference continuum mechanics to justify their approach.
    5. Results Presentation
      • The results table is overly dense and does not highlight the best scores, making it difficult to interpret.
      • The ablation study suggests that excluding the physical coupling of parameters yields better performance. While the authors argue that their full method imposes consistency, this claim is neither quantified nor demonstrated with an example.
    6. Ground Truth and Metrics
      • There is no mention of how the ground truth is generated for the simulated dataset.
      • It is unclear whether the reported metrics are calculated on the entire image or within a specific mask.
  • 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

    N/A

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

    (3) Weak Reject — could be rejected, dependent on rebuttal

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

    The manuscript introduces an approach for dynamic CBCT reconstruction using 4D Gaussian splatting with motion modeling. While the method is relevant and demonstrates promising qualitative results, its novelty is questionable, and key methodological aspects lack clarity. The introduction requires restructuring, results presentation should be improved, and the explanation of the deformation method needs stronger justification.

  • 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.

    Several of the initially raised weaknesses have been constructively addressed in the rebuttal, strengthening the manuscript’s clarity. The authors have committed to adding a schematic diagram to improve methodological transparency and to revising the results table to enhance interpretability by highlighting key scores and reporting averages. They have also clarified important ambiguities regarding ground truth generation and evaluation regions. Furthermore, they acknowledged the need to better explain the role of physical parameter coupling in the ablation study and expressed their intent to revise the discussion accordingly. Overall, these revisions have addressed my main concerns, and I have revised my score in favor of accepting the paper.



Review #3

  • Please describe the contribution of the paper

    This paper presents a 4D Gaussian Splatting (4DGS)-based approach for 4D CBCT reconstruction. The key contribution is a low-rank B-spline motion model to predict the time-dependent parameters of the Gaussian kernels. Experiments on synthetic datasets demonstrate faster training and improved reconstruction quality compared to the baseline methods, highlighting the potential clinical value of the proposed approach.

  • 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. This is the first (or a concurrent) work to apply 4D Gaussian Splatting (4DGS) to 4D CBCT reconstruction.
    2. The proposed low-rank motion model demonstrates both high efficiency and goof reconstruction quality. While the design draws some inspiration from existing RGB methods, the authors clearly articulate their motivation and provide a thorough discussion, making the work technically novel within the medical imaging domain.
    3. The paper is well-structured and easy to follow.
  • 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. In the ablation studies (Table 1), disabling the deformation-informed framework (DI) results in higher PSNR and lower RMSE, which somewhat contradicts the claim in Section 3.2 that the unified motion model outperforms separate models. More explanation regarding the advantages of the proposed design is expected.
    2. The experiments are conducted on synthetic datasets—real CT images paired with synthetic X-ray projections. While it’s understandable that public real-world datasets are limited, evaluating the method on real projections would further strengthen the work.
    3. Only one baseline method (Supremo) is included. Although there are limited open-source 4DCT methods, comparing against some 3DCT methods by reconstructing each phase independently could better highlight the value of the proposed 4D module.
  • 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

    Minor weaknesses

    1. There is only one figure. It is recommended to add a pipeline figure, which would greatly help readers grasp the overall framework more quickly.
    2. The term DVF in Section 3.2 is not explained, leaving readers unfamiliar with the concept.
    3. There are some formatting issues—punctuation marks such as periods or commas are missing at the end of some equations and sentences (e.g., Equations 3, 4,7, Fig 1. etc.). Addressing these would improve the overall clarity and presentation. It is also recommended to include an average metric at the end of Table 1 to provide a clearer overall comparison across methods.
  • 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?

    [+] This paper is the first to apply 4D Gaussian Splatting (4DGS) to 4D CBCT reconstruction. The authors introduce necessary adaptations to make 4DGS suitable for CT tasks. [+] Experimental results demonstrate good reconstruction efficiency and quality, highlighting the method’s practical value. [-] Some aspects of the ablation study are confusing and would benefit from further clarification. [-] Experiments are limited to synthetic datasets and include only one baseline method.

  • 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 have adequately addressed my concerns.



Author Feedback

We thank all reviewers for their constructive feedback. R2 acknowledges ours as the first (or among the first) work applying 4DGS to dynamic CBCT. R2 and R3 appreciate our technical novelties, and all reviewers highlight the clinical relevance, promising reconstruction quality, and computational efficiency of our approach.

We address the main concerns below:

  1. Motivation for deformation-informed (DI) framework (R1,R2,R3)
    Most 4DGS methods render dynamic images through changing the Gaussian parameters independently. However, voxel-wise deformation vector fields (DVF) are often required in Radiotherapy, e.g. for dose warping or tumour tracking. If changes of Gaussian parameters are modelled independently, it is impossible to derive a voxel-wise DVF that is consistent with the rendered dynamic images, i.e. one cannot define a transformation that maps a voxel-grid representation of one dynamic image to another. Therefore, inspired by continuum mechanics and works on physics based GS, our DI framework transforms all the Gaussian parameters through a unified DVF that can also be applied to voxel grid representations.
  2. Ablation study (R1,R2,R3)
    Our first ablation study assesses the impact of enforcing consistent updates on all the Gaussian parameters, compared to allowing each parameter to evolve independently. It is unsurprising that the decoupled method reconstructs the dynamic images slightly better, as there are more parameters to fit and the Gaussians have more flexibility to match the projection data. The important conclusion is that imposing consistent updates only has a minor impact on reconstruction metrics, but requires fewer parameters, reduces training time, and more importantly enables consistent voxel-wise DVFs to be generated. We now realize this was not clear in our original paper and will endeavor to make it clearer in our revised paper if accepted.
  3. Verification on real patients (R2,R3) We agree that patient data is valuable. However, evaluation is difficult without known ground truth, so we use realistic simulations that preserve anatomical motion (e.g., sliding) and scanner geometry. Future study will explore strategies to quantify performance on real patients.
  4. Presentation (R1,R2,R3) We will revise the formatting issues, highlight best scores, report averages in Table 1, revise language and check carefully for mistakes, and clarify details as requested.

For more specific comments of each reviewer,
To R1:
7.1. All parameters are optimized jointly by Adam optimizer (section 4.2). We will include a schematic diagram if space allows. 7.2. “4D CBCT” typically denotes phase-based 4D images, so we use the term “dynamic” to refer to projection-based 4D images. Ref24 shows how a prior motion model can be constructed from phase-based images, which can later be used to fit projection-wise motion. However, we acknowledge that this may not be the most appropriate reference and will verify all citations to ensure accuracy.
7.4. We are the first to introduce an explicit B-Spline Free-Form-Deformation representation within the 4DGS framework that significantly improves efficiency, and the first to apply a physics-based DI framework to image reconstruction problems while prior works focus on generative motion. 7.6. The ground truth is created by warping reference images with known DVFs (section 4.1). All evaluations are within the reconstruction FOV. We will clarify the details as requested. To R2: 7.3 Our method aims to generate an image for each projection, so 3D methods cannot be used to provide a comparison as they cannot reconstruct an image from single projection. To R3:
7.2 The underlying B-spline DVF is defined independently of the Gaussians and is continuous in space and time. Therefore, it can be queried at arbitrary spatial locations to produce voxel-wise DVFs.
7.4 Hyperparameters are selected to be consistent with prior works (Ref 9,27–section 4.2) to ensure comparability.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • 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”.

    N/A

  • 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    The rebuttal sufficiently addresses the points raised by the reviewers.



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



back to top