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

Abstract

Reconstructing 3D fetal brain MRI volume from 2D slices is crucial for prenatal assessment but challenged by unpredictable fetal motion and large inter-slice gap. While 3D Gaussian Splatting (3DGS) has recently enabled rapid slice-to-volume reconstruction (SVR), the optimization of independent Gaussian primitives lacks explicit geometric constraints. This unconstrained flexibility often leads to structural distortion and artifacts, limiting reconstruction fidelity. We propose G3R, Gaussian-based Geometry-Guided Reconstruction, a geometry-guided Gaussian reconstruction framework that introduces a straightforward yet effective structural prior to stabilize Gaussian optimization. Specifically, we develop the Density-Gated Prior Network (DGPN) to yield a prior volume from physics-derived density map. This prior volume is then incorporated in a two-stage Gaussian reconstruction process: (1) prior-based primitive initialization to establish stable geometry, and (2) slice-driven refinement with an annealed distillation constraint to preserve structural consistency while recovering fine details. We demonstrate that our approach matches the reconstruction speed of existing Gaussian frameworks while delivering state-of-the-art quality with 24.56 dB PSNR in simulated dataset, specifically achieving up to +1.87 dB PSNR in a single-stack stress test. These promising results highlights the significant potential of application into clinical routine of real-time fetal 3D MRI.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2298_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{CaiZhi_G3R_MICCAI2026,
        author = { Cai, Zhibao AND Lv, Yao AND Zhang, Xin AND Yang, Chaoxiang},
        title = { { G3R: Gaussian-Based Geometry-Guided Reconstruction for Fetal Brain MRI } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16895},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper introduces a structural prior to stabilize Gaussian optimization with a Density-Gated Prior Network (DGPN), which constructs a density map by integrating the scanner PSF over estimated slice poses to identify acquisition voids and guide inpainting of missing anatomy

  • 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.
    • G3R demonstrates strong performance, outperforming both NeSVoR and the unconstrained GSVR.
    • The method is effective in single-stack settings, outperforming GSVR. This demonstrates the value of the prior when spatial coverage is limited.
    • G3R maintains a fast inference speed of ~90s, outpacing implicit methods like NeSVoR (~175s).
  • 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.Limited novelty relative to existing implicit representations: While the formulation is well presented, the use of implicit neural representations for reconstruction is not entirely new, and the paper would benefit from clearer positioning with respect to prior neural field / NeRF-like approaches. 2.Computational cost and practicality; The method relies on per-subject optimization, which is computationally expensive compared to feed-forward approaches. This may limit clinical applicability, especially in time-sensitive settings. 3.Lack of evaluation on pathological cases: The evaluation appears to focus on non-pathological or standard datasets. It remains unclear how the method performs in the presence of abnormal anatomy, where reconstruction is more challenging and clinically relevant. 4.Limited analysis of robustness: There is limited discussion on how the method behaves under: severe motion , sparse slice coverage , low SNR These factors are critical in fetal MRI. 5.Comparison to strong baselines could be further clarified: While comparisons are provided, it is not always clear how the method performs relative to the latest learning-based or hybrid approaches under identical conditions.

  • 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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

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

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
    • Clarify positioning with respect to prior work: better distinguish the proposed approach from existing implicit representation and neural field-based reconstruction methods.
    • Expand discussion on computational trade-offs: provide a clearer discussion of runtime and potential strategies for improving efficiency.
    • Discuss applicability to pathological cases: include a discussion on expected performance in abnormal anatomy and potential limitations.
    • Improve robustness discussion: elaborate on performance under challenging acquisition conditions such as motion and sparse sampling.
    • Clarify experimental setup and comparisons: ensure that comparisons to baselines are clearly described and fair.
  • 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 paper presents a solid methodological contribution to slice-to-volume reconstruction by leveraging continuous implicit representations. The approach is technically sound and has the potential to improve reconstruction fidelity. However, the level of novelty relative to existing implicit representation approaches is moderate, and practical limitations related to computational cost and robustness remain. Addressing these points would further strengthen the work.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The authors have addressed the reviewer comments satisfactorily in the rebuttal. While some limitations remain, the contribution is technically sound and improved in clarity.



Review #2

  • Please describe the contribution of the paper

    This paper proposes a novel geometry-guided 3D Gaussian Splatting(3DGS) framework for fetal brain MRI SVR. This work introduces a Density-Gated Prior Network(DGPN). This network leverages physical Point Spread Function(PSF) accumulation to construct a physical density map, explicitly identifying acquisition voids.

  • 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 integration of the physical PSF density map into the prior generation process is highly innovative. Rather than relying purely on black-box generative models for inpainting, the framework uses physical observation confidence to explicitly gate the neural network’s contribution. This significantly improves the interpretability and reliability of the geometric prior. 2.The authors accurately identify and successfully mitigate the geometric collapse issue of standard 3DGS when applied to under-determined medical imaging tasks (sparse, motion-corrupted slices). The introduction of the distillation anchor is a straightforward yet highly effective geometric constraint. 3.The proposed method achieves a significant performance leap (+1.87 dB PSNR) in the highly challenging single-stack clinical scenario compared to the unconstrained GSVR baseline. Furthermore, it maintains the rapid inference speed characteristic of 3DGS (~90 seconds), striking an excellent balance between reconstruction fidelity and computational efficiency, which is highly desirable for intra-session clinical workflows.
  • 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.This method is an improvement for GSVR. The experimental results showed a very limited improvement, as shown in Table1.2.The authors highlight the single-stack performance as a major contribution. However, the baselines used (NeSVoR, GSVR) inherently rely on multi-view geometric intersections and are naturally expected to degenerate in single-stack settings. It should be compared to with the single-stack method. [1]Fully Convolutional Slice-to-Volume Reconstruction for Single-Stack MRI, CVPR 2024.3.The ablation study is not Sufficient and comprehensive. It lacks to verify the effectiveness of the physical density map and two-stage gaussian construction.

  • 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

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

    N/A

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

    N/A

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

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

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

    The innovation of method is not clear. It need to clairfy the motivation of design. The experimental results showed a very limited improvement. The comparison experiments and ablation study are not comprehensive.

  • 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 completely addressed my concern.



Review #3

  • Please describe the contribution of the paper

    This paper presents a 3D Gaussian based Slice-to-Volume Reconstruction method for motion-corrupted fetal brain MRI. The framework utilizes a Density-Gated Prior Network (DGPN) to generate structural priors from coarse volumes and physical density maps. By employing a gating mechanism, the network preserves original observations in high-confidence (high-density) regions while relying on network inpainting in low-confidence areas to ensure structural continuity. The Gaussian-based reconstruction follows a two-stage pipeline: (1) prior-based initialization of Gaussian primitives ; (2) detail refinement using a data-consistency term. Experimental results on simulated and private clinical datasets show that the proposed G3R outperforms existing baselines in quantitative metrics and visual quality (reduced artifacts/noise). Notably, G3R significantly surpasses GSVR in challenging single-stack scenarios.

  • 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 focuses on addressing the ill-posed challenges of motion-corrupted fetal brain MRI Slice-to-Volume Reconstruction through a clearly articulated methodology and a highly readable presentation. By introducing a Density Gated Prior Network, the authors generate a robust prior volume from corrupted input slices that effectively constrains the 3D Gaussian Splatting process, thereby successfully mitigating the structural distortions typical of underdetermined problems. This approach significantly alleviates SVR’s inherent motion artifacts, delivering superior quantitative results and visual fidelity on both simulated and private datasets compared to the proposed baselines. Furthermore, the reconstruction efficiency is on par with existing Gaussian-based techniques and is notably faster than methods based on INR.

  • 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.

    Reproducibility: The reproducibility of the proposed method is a concern. The DGPN relies on supervised pre-training, yet the manuscript lacks essential details regarding its architecture and hyperparameters. Furthermore, the specific weights for the various components of the loss function during Gaussian optimization are not specified, making it difficult for others to replicate the results.

    Fairness of Comparison: The fairness of the performance evaluation is questionable. While the proposed method introduces a supervised module, it does not demonstrate a significant numerical improvement over unsupervised baselines like NeSVoR and GSVR. The trade-off between the overhead of supervised training and the marginal gains should be explicitly discussed. Additionally, comparing reconstruction times without accounting for pre-training duration or assessing robustness across different sequences may be misleading. Furthermore, in the single-stack scenario where the method significantly outperforms GSVR, the comparison remains biased as GSVR is unsupervised. The absence of comparisons with NeSVoR, SVoRT, and other specialized single-stack SVR methods (e.g., [arXiv:2312.03102]) further weakens the evaluation.

    Redundancy in the 3DGS Pipeline: The two-stage reconstruction workflow appears unnecessarily complex. Since the loss functions from the first stage (prior initialization) are retained in the second stage (detail refinement) via linear annealing, it is unclear why these two stages cannot be integrated into a single unified process. The necessity of this multi-stage approach requires stronger justification.

    Inconsistencies in Metrics and Ablations: There is a notable discrepancy in the ablation study: the DGR module significantly impacts SSIM but has a negligible effect on LPIPS. This phenomenon requires a detailed explanation and visual evidence to be convincing. Moreover, the ablation study is incomplete; it only focuses on the DGR module without evaluating the actual contribution of the 3DGS refinement. It remains unclear whether the final performance is driven by the 3DGS optimization or is simply a reflection of the high-quality prior generated by the pre-trained network.

  • 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 submission does not provide sufficient information for reproducibility.

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

    N/A

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

    Regarding Observation Confidence c(x) in Fig.2, there appears to be a discrepancy between the formula and its visualization. While the mathematical definition suggests a monotonically increasing trend, the curve in the figure is monotonically decreasing. Please verify whether this is a plotting error or if there is a misunderstanding of the equation.

  • 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?

    While the proposed method shows marginal improvements over existing baselines, the empirical evidence is not sufficiently compelling to justify its complexity. The technical novelty is largely confined to the DGPN module; however, despite being a supervised pre-training component, it fails to deliver a significant performance leap over unsupervised alternatives. Given these concerns alongside the previously mentioned weaknesses regarding reproducibility and fairness, I believe the current manuscript does not meet the high standards required for publication.

  • 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.

    N/A



Author Feedback

We thank reviewers for their constructive feedback and for recognizing our work’s impact. We address the concerns below. 1.Core Novelty (AC, R1, R2) Existing implicit (NeSVoR) or 3DGS-based (GSVR) methods share an unsupervised per-subject optimization paradigm, which is unstable in realistic clinical settings with strong motion or heavy noise. To bridge this gap, our core contribution is introducing explicit geometry-guided stabilization into this paradigm. Instead of relying on data-driven generative priors, we explicitly estimate observation confidence from slice acquisition process to preserve reliable anatomy and inpaint unobserved regions. Further, a two-stage optimization first establishes stable global structure, which subsequently constrains detail recovery. Consequently, G3R retains the rapid inference of 3DGS while achieving strong robustness under challenging conditions (Table 1 and Fig. 4). 2.Single-Stack Comparison (AC, R2, R3) We sincerely thank R2 for pointing us to work [1], and we apologize for its initial omission. We clarify that G3R is a general framework rather than a specialized single-stack architecture. The single-stack setting was used strictly as an extreme stress test to simulate clinical worst-case scenarios. While [1] relies on a pre-trained interpolation network to infer missing structures, G3R leverages observation confidence (Sec 2.3) to guide the structure. Our comparison shows that G3R achieves better performance than [1] under the same experimental settings. We commit to a full quantitative comparison in the final version. 3.Two-Stage Architecture (AC, R3) We sincerely thank AC and R3 for raising this critical point. We agree that a unified joint optimization is conceptually appealing, and we actually evaluated it during development. However, previous experiment shows that optimizing both global structure and fine details simultaneously from scratch is less unstable then current two-stage strategy. In such joint optimization, the Gaussian primitives quickly overfit sparse slice observations before a coherent 3D anatomy can form. We therefore decouple global geometric stabilization from high-frequency detail extraction. This instability is also consistent with the Fig.4A ablation, which already shows significant degeneration when only the final structural anchoring is removed. We will clarify this motivation and include the ablation study in final version. 4.Pre-training Overhead (R3) We thank R3 for raising the important discussion on fairness. We respectfully clarify that, under large fetal motion, unsupervised baselines (NeSVoR and GSVR) also rely on pre-trained registration modules (e.g., SVoRT) to improve robustness during slice alignment. Therefore, leveraging offline pre-trained models to process fetal data is a standard paradigm rather than unique to G3R. Clinically, per-patient processing time is the true bottleneck. Therefore, we focus our runtime comparison on the online reconstruction stage, where G3R maintains rapid inference while providing substantially improved stability under challenging conditions. 5.Ablation and 3DGS Contribution (R3) We apologize for the confusing presentation. To clarify, the final performance is not simply a reflection of the high-quality prior, but is primarily driven by 3DGS optimization. The DGPN prior alone achieves 22.75 dB PSNR (Table 2), while the final output after 3DGS reaches 24.56 dB (Table 1). This +1.81 dB gain confirms that 3DGS actively extracts patient-specific details. To better illustrate this, we will restructure our ablation table to explicitly present (1) G3R without structural distillation constraint, (2) prior w/o geometric guidance, and (3) our full model. 6.Reproducibility (R1, R3) We will detail all hyperparameters and release our code. We sincerely thank R3 for pointing out the error in Fig. 2, which will be corrected. Ref: [1] S. I. Young et al., CVPR, 2024.



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.

    This paper tackles an important problem in fetal brain MRI SVR and presents a technically interesting 3D Gaussian-based framework with a density-gated prior, and the reviewers agree that the method is clearly presented and shows promising reconstruction quality and speed, especially relative to GSVR and in challenging limited-coverage settings. At the same time, the reviews reveal substantial disagreement about whether the current evidence is strong enough for acceptance. The main concerns are that the novelty relative to prior implicit/3DGS-style reconstruction methods is not yet sufficiently clarified, the empirical gains over existing baselines are sometimes modest, and the evaluation may not be comprehensive or fully fair, particularly in the single-stack setting where comparisons to stronger or more specialized single-stack methods are missing. Reviewers also raise concerns about incomplete ablations, limited discussion of robustness under difficult acquisition conditions, and insufficient implementation details for reproducibility, including DGPN configuration and optimization settings. Given the substantial disagreement among the reviewers, I recommend invite for rebuttal rather than an early decision. In the rebuttal, the authors should focus on clarifying the method’s novelty relative to prior neural field and Gaussian-based SVR approaches, justifying the design choices behind DGPN and the two-stage pipeline, addressing the fairness and completeness of the baseline comparisons, especially for the single-stack scenario, and clarifying the missing reproducibility details and the contribution of the different components in the ablation study.

  • 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.

    This paper proposes G3R, a Gaussian-based geometry-guided reconstruction framework for fetal brain MRI slice-to-volume reconstruction, aiming to improve reconstruction quality under motion-corrupted and limited-coverage acquisition settings. The method introduces a Density-Gated Prior Network (DGPN) to construct geometry-aware structural priors from physical density maps and uses a two-stage 3D Gaussian optimization strategy to stabilize global anatomy before recovering finer details. The reviewers recognized that the paper addresses an important fetal MRI SVR problem and presents a technically sound and efficient reconstruction framework. In particular, they appreciated the use of scanner PSF-based density maps to identify acquisition voids, the geometry-guided prior for mitigating geometric collapse in 3DGS-based reconstruction, and the favorable reconstruction speed compared with implicit neural representation methods. The results also show promising performance, especially in challenging single-stack settings where standard GSVR is less stable. The reviewers raised concerns about the novelty relative to existing implicit or Gaussian-based SVR methods, the fairness of the single-stack comparison, the necessity of the two-stage pipeline, incomplete ablations, and insufficient reproducibility details. In the rebuttal, the authors clarified that the main contribution is not simply applying 3DGS to SVR, but introducing explicit geometry-guided stabilization through observation-confidence modeling and density-gated structural priors. They also explained the role of the single-stack experiment as an extreme stress test, justified the two-stage design as a way to avoid early overfitting of Gaussian primitives to sparse observations, and clarified that the final performance is not solely driven by the prior but further improved by 3DGS refinement. Although some limitations remain, including the need for clearer DGPN implementation details, fuller ablation presentation, and more complete discussion of pre-training overhead and robustness under difficult clinical conditions, the rebuttal sufficiently addressed the main concerns raised by R1 and R2.The remaining issues are important but can be handled through camera-ready clarification rather than constituting fundamental flaws. Therefore, I recommend acceptance.



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.

    The main concerns remain unresolved after rebuttal. The novelty over existing implicit and Gaussian-based reconstruction methods is only partially clarified, and the empirical gains over strong baselines appear modest relative to the added complexity and supervised pretraining. The single-stack result is interesting, but the most relevant specialized single-stack comparison was absent from the submitted manuscript and is only promised for the final version.

    The contribution of the density map, the two-stage optimization, and the 3DGS refinement is not fully isolated in the submitted results. Reproducibility details for DGPN architecture, loss weights, and optimization are incomplete. The rebuttal provides helpful explanations, but several key points depend on additional comparisons and revised ablation tables that were not part of the reviewed manuscript.



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.

    Key concern have been addressed in the rebuttal and both experienced reviewers now recommend “accept”. Key strengths include the methodological innovation employing Gaussian Splatting in SVR with the G3R framework, the ablation study performed and the evaluation on multiple data sets among others. While weaknesses remain such as the two stage reconstruction workflow and some inconsistencies (with some addressed and recognized in the rebuttal), in general the paper could indeed provide interesting discussions at MICCAI in a highly dynamic research field.



back to top