List of Papers Browse by Subject Areas Author List
Abstract
The diagnosis of spinal diseases is often assisted by 3D imaging techniques in clinical practice. However, precise 3D spinal assessment is limited by the high costs of 3D imaging hardware and the challenges posed by the physical differences between imaging modalities, which hinder the generalizability of models. To address these issues, we propose UniSpine-GS, an efficient, physics-aware Gaussian framework designed for novel-view projection rendering in multi-view spine imaging via a 3D-aware representation. Instead of performing explicit 3D reconstruction, our approach learns a geometry-aware Gaussian representation that ensures anatomical consistency across different views. We introduce SPWM, a structure-guided loss reweighting strategy to improve boundary fidelity and local details. We evaluate our method on the CTSpine3D dataset and a newly constructed 3D fetal ultrasound dataset, FeSpine3D. Our results demonstrate that UniSpine-GS significantly outperforms existing methods across all metrics, offering a practical and cost-effective solution for unified multi-view medical imaging. Our code is publicly available at https://github.com/orangeisland66/UniSpine-GS.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5041_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/orangeisland66/UniSpine-GS
Link to the Dataset(s)
https://github.com/orangeisland66/UniSpine-GS
BibTex
@InProceedings{CheQiu_UniSpineGS_MICCAI2026,
author = { Chen, Qiuhua AND Yu, Changning AND Huang, Na AND Sun, Chao AND Du, Bo},
title = { { UniSpine-GS: An Efficient Physics-Aware Gaussian Framework for Cross-Modality Multi-view Spine Image Synthesis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16890},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This work proposed a 3D Gaussian representation based method for accurate spine image rendering. The method incorporate physical information of imaging modality including imaging geometry and anatomical consistency. The main initiative contribution is their structure prior weight map that improve boundary fidelity and local details. The method is tested on two modalities: X-ray CT and 3D ultrasound. Experimental studies on two datasets CTSpine3D and FeSpine3D are executed and results are compared with typical and SOTA methods in the field. Clear advantage of the proposed method is demonstrated.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The proposed method has significant advange on both image quality and speed.
- 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.
Performance evaluation on different noise levels is not provided, which could be helpful to comprehensively understand the value of this method.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The method are presented with clear logic that I agree well with. The structure prior weight map is innovative and the results are convincing with advantages in image quality and speed.
- 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 SPWM idea is interesting and effective in this work. However, the testing views are only angles in the small 3.6 degree interval between training angles, which somehow limit its persuasiveness regarding generalization performance.
Review #2
- Please describe the contribution of the paper
The main contribution of this work is a Gaussian-based framework for multi-view spine imaging that replaces explicit 3D reconstruction with a geometry-aware 3D representation for novel-view rendering.
By introducing physics-aware Gaussian modeling and structure-guided loss reweighting, the method aims to preserve anatomical consistency across views while improving local boundary fidelity.
- 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 construction of FeSpine3D is a valuable contribution, especially for fetal ultrasound spine imaging where publicly available 3D datasets remain limited. 2: Experimental validation is convincing, with evaluation on more than 600 CT volumes and 100 ultrasound volumes across two modalities. 3: The proposed ACUI and SPWM modules are well motivated and show consistent improvement in reported metric.
- 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 clarification of SPWM SPWM is presented as a key contribution relative to prior Gaussian-based methods such as X-Gaussian, yet its effectiveness is not fully clarified. Since SPWM mainly functions as a loss reweighting strategy, a more explicit discussion of how it differs from baseline uniform-weighting designs would strengthen the methodological contribution. In particular, additional visualization of the weighting behavior across modalities (e.g., 2D heatmaps) would help clarify which anatomical regions receive greater emphasis during training. It would also be interesting to examine whether the reweighted regions correspond to high-error areas in the reconstruction, which could provide further insight into how SPWM improves performance.
2.Efficiency improvement is not fully explained The source of the reported efficiency improvement remains unclear. In particular, the runtime behavior differs across datasets: on CTSpine3D, SPWM increases training time (756s → 923s), whereas on FeSpine3D it reduces runtime (782s → 504s). A clearer explanation of this difference is needed, including whether early stopping or other convergence-related mechanisms are involved (is using SPWM converge faster?), and how the claimed speedup should be interpreted across modalities.
3.Missing implementation details for reproducibility Several implementation details are not sufficiently specified for reproducibility, including the train/test split strategy, how sparse views are selected, and the number of rendered views generated through the DRR operator.
4.Missing visualization of learned 3D Gaussian splats Since the method relies on a Gaussian-based 3D representation, visualization of the learned splats in a 3D manner would help readers better understand how anatomical structures are represented.
- 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 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
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 — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses a clinically relevant problem and presents a practical Gaussian-based framework for multi-view spine image synthesis across CT and ultrasound data. A notable strength is the construction of FeSpine3D and the validation on both modalities, which improves the practical relevance of the study. The overall framework is also technically reasonable, and the proposed ACUI/SPWM components appear to contribute to performance gains.
However, my confidence is limited by several issues that affect how strongly the current results support the claimed methodological contribution. In particular, SPWM is presented as a key novelty, yet its behavior and distinction from the uniform weighting-based designs are not sufficiently clarified. The efficiency claim is also difficult to interpret because runtime behavior differs across datasets without clear explanation. In addition, several implementation details remain unclear, which makes reproducibility and comparison harder to assess.
Overall, while the direction is promising, clearer analysis and discussion would be needed to fully establish the novelty and practical advantages of the proposed design. I would be interested to see whether the rebuttal can clarify these concerns.
- 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.
1.Novelty: The main novelty still appears to be an extension of X-Gaussian to new medical data, namely CT spine and fetal ultrasound, together with a reweighted loss mask. I am not fully convinced that this alone reaches the MICCAI acceptance bar. The authors could strengthen the claim by providing more evidence, such as error visualizations or deeper analysis, to show that the proposed SPWM is explicitly connected to the imaging/physical model and can better handle ambiguous boundaries and speckle-like effects.
2.Convergence: It is still unclear to me whether early stopping is used, and how convergence is determined in practice. More details on the convergence criterion would explan better the efficiency improvement,
3.3D visualization: Although raw splats may not correspond to clinically interpretable surfaces, 3D-level visualization remains useful as a sanity check. It can help determine whether the model learns a plausible anatomical 3D distribution or mainly overfits to 2D projection appearances.
Overall, while the application is interesting, I remain concerned that the technical novelty over X-Gaussian is limited, and that the current evidence does not sufficiently demonstrate whether the proposed components lead to anatomically meaningful 3D representations, rather than merely improving 2D projection-level metrics. For these reasons, I will keep my recommendation as Weak Reject.
Review #3
- Please describe the contribution of the paper
The main contribution is UniSpine-GS, an efficient Gaussian-based framework for multi-view spine image synthesis across CT/X-ray-style projection data and fetal ultrasound, built on a unified radiative forward operator and a structure-guided loss reweighting module (SPWM) to improve boundary fidelity and local detail under sparse-view supervision. The paper also introduces the FeSpine3D dataset.
- 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.
A strength of the work is the quality-efficiency trade-off. The reported results on both CTSpine3D and FeSpine3D show very strong PSNR/SSIM and substantially faster training and rendering than the listed baselines.
The proposed SPWM is also a reasonable design choice: it is simple, lightweight, and the ablation suggests it provides a consistent gain over the radiative Gaussian baseline. The inclusion of a new fetal ultrasound dataset is another positive aspect.
- 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.
My main concern is that the methodological novelty is limited. The framework is built largely on top of X-Gaussian, and the paper explicitly adopts ACUI and the radiative Gaussian formulation from that prior work. In practice, the main new element is SPWM, which makes the contribution feel more incremental than the title suggests.
I am also not fully convinced by the “physics-aware cross-modality” positioning. For ultrasound, the method uses the same DRR-style projection operator as a pragmatic proxy and explicitly does not claim a physically accurate acoustic model. This weakens the cross-modality claim, since the framework does not really model ultrasound image formation in a modality-faithful way.
- 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
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 — 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?
I lean weak reject. The paper is well motivated, technically solid, and reports strong quantitative results with excellent efficiency. However, the core method appears incremental relative to X-Gaussian, the “physics-aware cross-modality” claim is stronger than what is actually implemented for ultrasound, and the main comparison omits the most relevant direct baseline from the headline results. Taken together, these issues make the overall contribution less convincing than the quantitative numbers alone suggest, so I currently view the paper as slightly below the acceptance threshold.
- Reviewer confidence
Somewhat confident (2)
- [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 think the author partially addressed most of my concerns.
Author Feedback
We thank all reviewers for recognizing the clinical relevance of our task, the strong performance and efficiency of UniSpine-GS, and the value of FeSpine3D. We address the main concerns and will add clarifications in the revision.
1.SPWM and efficiency (R2, R3). Unlike uniform weighting, SPWM uses a structure prior from edge and high-frequency responses to emphasize anatomical boundaries and discontinuities, guiding optimization toward vertebral contours and detail-rich regions. We will add 2D SPWM heatmaps and error maps showing that reweighted regions correspond to boundary and high-error areas. SPWM has negligible overhead because it is computed by lightweight operations and used only as a pixel-wise loss weight. On CT projections, it slightly increases training time by encouraging finer refinement around dense boundaries. On fetal ultrasound, where boundaries are ambiguous and speckle-like variations are common, it guides optimization toward structural areas, leading to faster convergence.
2.Novelty vs. X-Gaussian (R3). X-Gaussian focuses on X-ray novel-view synthesis, while UniSpine-GS addresses cross-modality sparse-view spine synthesis across CT/X-ray-style projections and fetal ultrasound. Based on broad literature review and experiments, we leverage radiative modeling and ACUI for projection-consistent representation, but further introduce SPWM for task-specific structure-aware supervision. SPWM is especially useful for ambiguous ultrasound boundaries; on FeSpine3D, it improves PSNR from 39.91 to 40.82.X-Gaussian is already in our ablation study; we will highlight it more clearly as the direct baseline.
3.“Physics-aware” concern (R3). We clarify that our goal is not to simulate the full physical mechanisms of CT/ultrasound or exact acoustic ultrasound formation. Instead, we explore a unified rendering strategy for medical volumetric view synthesis. Medical Vision Generalist shows that CT, X-ray, MRI, and micro-ultrasound can be handled in a unified image-to-image framework. Similarly, our unification is based on projection-consistent representation learning, not exact modality-specific physics. For ultrasound, the DRR-style operator is a pragmatic differentiable projection proxy. Extensive experiments on CTSpine3D and FeSpine3D show this strategy works effectively for both CT/X-ray-style projections and fetal ultrasound.
4.Reproducibility and data availability (R1, R2, R3). We have organized the source code, configs, data-generation scripts, and will add a project-page link to these materials in the camera-ready version. To avoid view overlap, each volume uses 50 training views uniformly sampled over 180° at 3.6° intervals (0°, 3.6°, …, 176.4°), and 50 testing views as held-out midpoints (1.8°, 5.4°, …, 178.2°). Projections are rendered with TIGRE under known cone-beam geometry. We will also clarify optimizer settings, densification and pruning, loss configuration, and SPWM hyperparameters.
Regarding data availability, CTSpine3D is publicly accessible. For FeSpine3D, we will provide detailed dataset descriptions, acquisition information, preprocessing steps, and representative data statistics on the project page. Access to FeSpine3D will be provided to qualified researchers upon reasonable request under appropriate data-use agreements.
5.Noise-Level Performance (R1). We appreciate the reviewer’s constructive comments. This is an interesting exploration, and we will add visualizations under different noise levels in the revised version.
6.3D visualization (R2). We thank the reviewer for raising this point. Raw 3D Gaussian splats are less suitable because UniSpine-GS targets multi-view 2D projection synthesis, not explicit 3D anatomical surface recovery. In radiative projection imaging, Gaussians are implicit representations optimized for rendering and do not necessarily correspond to clinically interpretable surfaces. We will add projection-domain SPWM heatmaps and error maps.
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 has received a range of scores from the three reviewers, reflecting divergent opinions. As such, the authors are encouraged to submit a thorough rebuttal to clarify and defend their work. In particular, attention should be directed to the following key concerns: the methodological novelty, especially regarding the introduction of SPWM which appears to represent an incremental contribution, a rigorous justification of the central claim around “physics-aware cross-modality” positioning, and the absence of relevant state-of-the-art baseline models for comparison. Given the limited space available for rebuttal, the authors are advised to prioritize the most critical points and allocate their response strategically to maximize the effectiveness of their reply.
- 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 post-rebuttal discussion improved the paper, and the main concerns are now mostly about scope, clarity, and the strength of the evidence rather than any fatal flaw. One reviewer still remains unconvinced about novelty and would prefer additional analysis, but the other post-rebuttal opinions are positive and indicate that the authors sufficiently addressed the major questions raised. Overall, I view the remaining concerns as limitations that do not outweigh the paper’s technical soundness and practical value. I therefore 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.
Accept
- Please justify your recommendation.
This paper introduces UniSpine-GS, an efficient physics-aware 3D Gaussian Splatting framework for cross-modality, multi-view spine image synthesis across CT and fetal ultrasound datasets. To capture fine-grained vertebral boundaries from sparse views, a Structure Prior Weight Map (SPWM) is proposed to dynamically reweight the reconstruction loss. The paper received split post-rebuttal recommendations (two Accepts, one Weak Reject). Reviewers consistently praised the strong clinical motivation, the exceptional parallel training efficiency (7–15 minutes), and the real-time inference rendering speed (>113 fps) alongside superior quantitative improvements over NeRF baselines. While the negative recommendation persists due to concerns regarding incremental novelty over X-Gaussian and narrow angular testing intervals, these are counterbalanced by the framework’s outstanding quality-efficiency trade-off and the value of the newly introduced FeSpine3D benchmark dataset. The authors successfully resolved baseline and optimization queries during the rebuttal, convincing the majority of the panel that the work is well-suited for publication.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper received mixed assessments, with two reviewers supporting acceptance after rebuttal and one reviewer maintaining a negative recommendation. In my view, the rebuttal addresses most of the major concerns sufficiently and the proposed framework shows clear application potential.
I therefore recommend acceptance. For the camera-ready version, the authors should further clarify the efficiency and convergence behavior raised by R2, especially why SPWM affects training time differently across datasets and how convergence or early stopping is determined in practice.