Paper Info Reviews Author Feedback Meta-Review Back to top

List of Papers Browse by Subject Areas Author List

Abstract

Clinical magnetic resonance imaging (MRI) protocols are spatially heterogeneous across modalities and clinical objectives, compromising multimodal joint analysis and arbitrary-view generation. To harmonize MRI resolution, we leverage the commonly acquired isotropic T1-weighted sequence in neuroimaging protocols. Our Gaussian Splatting (GS)-based shared geometry framework adopts a two-stage training strategy, in which an explicit, subject-specific Gaussian scaffold encoding anatomical geometry is first learned from the isotropic structural scan and then reused to fit appearance for target modalities acquired with sparse slices. Experiments on the UK Biobank, GBM, and ABCD datasets for through-plane super-resolution across multiple modalities (T2-weighted, FLAIR, DWI, ASL), degradation factors (x3, x5, x7), and pathological abnormalities (glioblastoma) demonstrate state-of-the-art reconstruction fidelity. The shared Gaussian geometry enables arbitrary-view generation for target modalities with strong structural consistency and further shows potential for self-supervised in-plane super-resolution. This work establishes explicit geometry-guided representations as a novel, flexible, and interpretable pathway toward retrospective multi-contrast MRI harmonization and reliable clinical reference construction. The GitHub code is available at: https://github.com/yfgao76/AtlasGS.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2065_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/yfgao76/AtlasGS

Link to the Dataset(s)

UKBiobank: https://www.ukbiobank.ac.uk/ UPENN-GBM: https://www.cancerimagingarchive.net/collection/upenn-gbm/ ABCD: https://abcdstudy.org/ (curated from https://github.com/FGA-DIKU/fomo_mri_datasets)

BibTex

@InProceedings{GaoYif_Shared_MICCAI2026,
        author = { Gao, Yifan AND Xu, Peiran AND He, Yimeng AND Li, Haoran AND Long, Ziyang AND Li, Debiao},
        title = { { Shared Gaussian Geometry for Brain MRI Spatial Resolution Harmonization Without External Training } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16881},
        month = {September},
        page = {pending}
}

Reviews

Review #1

Please describe the contribution of the paper

This work leverages Gaussian Splatting-based shared geometry across scanning modalities to perform harmonization to a target modality, with the end goal being multi-contrast MRI spatial resolution harmonization. This is achieved by leveraging topology preservation, regularization and introducing a low-resolution consistent fusion.
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 objective of the paper and the motivations are clear. The methodology is cohesive, with all necessary mathematical background provided. The overview figure is a clear visual summary of the paper’s methodology as well. Experiments are thorough, and there is a decent combination of 3 cited datasets, 4 scanning modalities, and two degradation techniques. Overall, great presentation and the methodology is compared about recent state of the art methodologies. Finally, excellent vocabulary and phrasing of the science behind the MRI acquisition and inverse problem.
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.

Some aspects of the methodology don’t seem to contribute enough to spark discussion. In the ablation study it appears that the peristance topology loss and the latent appearance does not have a significant impact compared to LR Consistent Fusion which should be explained more thoroughly. Also, functional scans such as DWI and ASL might not be the optimal target modalities for this methodology. Finally, notation in equations is not defined before usage in a few cases cases.
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 has provided an anonymized link to the source code, dataset, or any other dependencies.
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

I think the paper is really good. A few questions: I would like to hear more about whether there are any qualitative improvements introduced by the latent appearance fitting and the topology persistance loss. Is there significant overhead training time for these losses? DWI and ASL are not anatomy based reconstructions and are more functional. Could that be why DWI did not do as well? What’s the opinion on ASL doing well regardless? Other notes: 1) Table 4 should also have some bold values to indicate best scores for readability. 2) Table 3 caption, “underline” should be capital. 3) Typo: “Low-resolution inputs are we synthetically degraded” 4) FN (folded-gaussian) used in equation before definition 5) Ψ in eq 9 should be explained or the equation should not be there and the loss should be just explained with text and the citation. 6) eq 1 and intro to methodology, LR and HR should be defined before usage
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 paper is novel, with a solid motivation and great presentation of methodologies and results. The MICCAI audience would be interested in a methodology that focuses on super-resolution and multi-contrast harmonization for MRI and the work done is presented clearly, for the work to be easily cited up on in the future.
Reviewer confidence

Very confident (4)
[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

This paper proposes a zero-shot MRI spatial harmonization framework based on shared Gaussian geometry, which enables modality-specific high-resolution reconstruction by leveraging an explicit anatomical scaffold learned from isotropic structural MRI, rather than reconstructing each modality independently without shared geometric priors. By introducing a two-stage geometry–appearance decomposition strategy, deformable Gaussian-based continuous representation, and low-resolution consistent fusion mechanism, the proposed method effectively preserves cross-modality structural consistency while improving through-plane super-resolution fidelity under sparse thick-slice supervision, providing a novel and interpretable solution for retrospective multi-contrast MRI harmonization.
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.
- Novel Geometry-Guided Harmonization Framework: This paper introduces a novel zero-shot MRI spatial harmonization framework that leverages shared Gaussian geometry learned from isotropic structural MRI to guide reconstruction of heterogeneous target modalities. By explicitly disentangling anatomical geometry from modality-specific appearance, the proposed method offers an interpretable and structurally grounded alternative to conventional interpolation or contrast-agnostic implicit reconstruction methods.
- Continuous Representation Design: The proposed deformable Gaussian-based continuous representation extends standard Gaussian splatting by modeling nonlinear through-plane anatomical continuity and modality-specific appearance separately. This design enables arbitrary-view rendering and continuous super-resolution while preserving structural consistency across modalities, representing a technically meaningful extension of Gaussian-based medical image reconstruction frameworks.
- Well-Motivated Physics-Aware Optimization Objectives: The training objective incorporates slab-aware thick-slice integration, topology-preserving persistent homology loss, and latent appearance regularization. These components are carefully designed to reflect MRI acquisition physics and anatomical structural priors.
- Multi-dataset Validation:he method is evaluated across multiple datasets (UKBB, GBM, ABCD), MRI modalities (FLAIR, T2w, DWI, ASL), degradation factors, and pathological settings including glioblastoma cases. This broad experimental setup demonstrates the generalizability of the proposed framework across diverse acquisition protocols and clinical scenarios.
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.
- Absence of Real-world Evaluation: The experimental validation is conducted primarily on synthetically degraded low-resolution inputs generated from originally high-resolution volumes, while quantitative evaluation on truly real-world clinical thick-slice acquisitions is absent. As a result, the practical robustness of the method under realistic scanner-dependent slice profiles, motion artifacts, and noise characteristics remains insufficiently verified.
- Insufficient Analysis of Failure Cases and Modality-Specific Limitations: While the paper reports broad experimental results across multiple datasets and modalities, it provides limited discussion of cases in which the proposed method underperforms or offers only marginal improvement. For example, the relatively weak performance on DWI is not meaningfully analyzed, and the paper does not clarify under what modality conditions the shared Gaussian geometry prior is reliable or when it may become a source of bias. This lack of deeper error analysis reduces the interpretability of the empirical findings.
- Terminology and Task Framing Clarity: While the paper addresses an important problem of cross-protocol MRI spatial resolution alignment, the terminology used throughout the manuscript may cause conceptual confusion within the medical imaging community. Specifically, the term “spatial harmonization” is not commonly used in the MRI harmonization literature. In general, MRI harmonization refers to reducing non-biological variability across scans acquired under different scanners, institutions, protocols, or imaging settings while preserving underlying anatomical and pathological information, rather than specifically denoting spatial resolution matching or super-resolution. Given this established convention, the current terminology may be somewhat misleading. It may improve clarity and consistency with prior literature to describe the task more explicitly as “spatial resolution alignment for multi-contrast MRI harmonization”, thereby clarifying that the proposed framework focuses specifically on resolution standardization rather than broader MRI harmonization.
- Minor Notation Clarity and Typo: Several abbreviations are introduced before being properly defined. For example, “FN” appears in Eq. (3) before its formal explanation is provided near Eq. (4), which may hinder readability. In addition, minor typographical issues are present, such as the duplicated comma following “(Fig. 3 left)” in the Qualitative Results subsection of Section 3.A careful revision of notation ordering and manuscript proofreading would improve the overall readability of the paper.
Please rate the clarity and organization of this paper

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

The authors claimed to release the source code and/or dataset upon acceptance of the submission.
Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

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

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

(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

My recommendation is based on several factors. First, the paper presents a technically solid and well-motivated framework for MRI spatial resolution alignment by extending Gaussian Splatting with deformable shared geometry modeling, modality-specific appearance fitting, and physics-aware rendering. Although the core representation is built upon existing Gaussian Splatting formulations rather than introducing an entirely new reconstruction paradigm, the proposed adaptations are meaningful and sufficiently tailored to the medical imaging setting to constitute a non-trivial methodological contribution.

Second, the paper demonstrates broad empirical validation across multiple datasets, modalities, degradation factors, and pathological cases, and the proposed method shows strong performance particularly under severe anisotropic degradation. The explicit geometry-guided formulation is also practically appealing, as it addresses retrospective multi-contrast MRI harmonization without requiring external training data, which may offer meaningful clinical utility in heterogeneous real-world neuroimaging cohorts where acquisition protocols vary substantially across scanners and institutions.

At the same time, several weaknesses prevent a stronger recommendation. The experimental evaluation is conducted primarily on synthetically degraded data rather than truly real-world low-resolution clinical acquisitions, leaving practical robustness insufficiently validated. In addition, the shared-geometry assumption does not appear to hold uniformly across all modalities, as evidenced by the relatively weaker performance on ABCD DWI, yet this limitation is not adequately analyzed.

Overall, despite these limitations, I believe the paper addresses an important and clinically relevant problem, presents a technically sound and reasonably novel solution.
Reviewer confidence

Somewhat confident (2)
[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

1.This approach offers a novel method for addressing spatial heterogeneity in MRI protocols. 2.This flexible strategy allows for adaptation to various modalities. 3.These improvements significantly enhance the quality of reconstruction. 4.The method demonstrates state-of-the-art reconstruction fidelity for through-plane super-resolution across multiple modalities (T2w, FLAIR, DWI, ASL), degradation factors, and pathological abnormalities (glioblastoma) without external training data (zero-shot). 5.This highlights the practical applicability of the framework in medical imaging analysis.
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.It introduces GS to MRI resolution harmonization, employing a novel two-stage training strategy where a geometric scaffold is first learned from isotropic T1w scans, then used to fit the appearance of other modalities. 2.The zero-shot approach operates within a single subject without external training data, leveraging existing clinical T1w images as a basis for knowledge transfer, which is highly practical for clinical use. 3.It demonstrates state-of-the-art reconstruction fidelity across diverse datasets, critically preserving tumor lesions accurately and mitigating interpolation artifacts, thus proving its clinical reliability. 4.Capabilities like arbitrary-view generation, potential for in-plane super-resolution, and an interpretable Gaussian-based representation offer significant extensibility to various future applications in medical image analysis.
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 technically interesting and the overall idea is promising. However, I believe the manuscript has several important weaknesses that should be more clearly acknowledged. 1.The term zero-shot is used, but it’s fundamentally a per-subject optimization. While it doesn’t require external training data, it still requires a full, high-resolution isotropic T1w scan and subsequent training for each new subject. This is a crucial distinction. It’s not a pre-trained model that can be instantly applied to any new patient. This significantly limits its scalability in a high-throughput clinical setting. 2.The paper claims “zero-shot adaptation of volume appearance” but this adaptation takes 20 minutes per subject. This is not “adaptation” in a practical sense. It’s per-subject training. 3.The method assumes that different MRI modalities can share the same geometry while differing mainly in appearance. However, modalities such as FLAIR, DWI, and ASL reflect substantially different signal mechanisms, not just intensity contrast. It remains unclear whether the proposed appearance representation is expressive enough to capture these modality-specific properties. 4.The inclusion of linear and cubic interpolation is reasonable, but these are relatively weak baselines compared with a learned multimodal reconstruction framework. Additional comparisons with stronger learning-based cross-modal reconstruction or super-resolution methods would make the empirical validation more convincing. 5.The proposed 2D+t Gaussian framework includes many interacting parameters and dynamic components, which makes the model highly flexible but also potentially difficult to optimize and control. The manuscript does not sufficiently discuss regularization, optimization stability, or sensitivity to noise and initialization.
Please rate the clarity and organization of this paper

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

The authors claimed to release the source code and/or dataset upon acceptance of the submission.
Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

N/A
Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
This is an interesting and technically promising paper, and the shared-geometry design is appealing. The results are encouraging, particularly for pathology-preserving reconstruction. However, several points should be clarified or strengthened. 1.the term “zero-shot” may be somewhat misleading, since the method still requires subject-specific optimization at test time with nontrivial computational cost. A more precise description would improve clarity. 2.the paper provides limited discussion of robustness in realistic clinical settings, such as motion corruption, protocol variation, and severe anatomical deformation. This reduces confidence in the practical applicability of the framework. 3.the technical novelty relative to prior Gaussian Splatting-based medical imaging methods could be positioned more clearly. The manuscript would benefit from a more explicit statement of what is fundamentally new in the proposed approach.
1. the baseline comparisons are somewhat limited. While interpolation baselines are useful references, stronger learning-based comparisons would make the empirical validation more convincing. 5.the paper would be strengthened by a clearer discussion of optimization stability, model complexity, and generalizability beyond brain MRI. Overall, the paper is promising, but the manuscript would benefit from clearer positioning of the contribution and a more explicit discussion of limitations and practical scope.
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 — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

This paper presents a novel and clinically relevant application of Gaussian Splatting for MRI resolution harmonization. The idea of using shared Gaussian geometry from isotropic T1w MRI to guide the reconstruction of other anisotropic modalities is technically interesting, and the two-stage training strategy is elegant. The experimental results are strong overall, particularly in preserving tumor pathology and reducing interpolation artifacts, which supports the potential impact of the method. I recommend Weak Accept rather than a stronger score for several reasons. First, the claimed “zero-shot” setting still requires subject-specific optimization with substantial computational cost, which limits practical clinical usability. Second, the paper does not evaluate robustness to motion, anatomical deformation, or other common real-world clinical challenges. Third, the manuscript provides limited discussion of optimization stability, regularization, and robustness of the relatively complex 2D+t Gaussian framework. Finally, while the application is novel, the core technical components build on prior GS-based medical imaging work, and the baseline comparisons could be strengthened with more competitive learning-based methods.
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 reviewers and area chairs for their careful review and valuable comments. In this response, we would like to clarify several unclear phrasings and claims in our submission, and discuss details and potential future directions based on our experimental findings and ablation studies. To Reviewer #2: 1.The latent appearance and topology persistence losses in our method mainly contribute to improved visualization quality and suppression of speckles and unrealistic point clouds caused by the Gaussian Splatting (GS) representation. Specifically:

The latent appearance loss significantly reduces random speckles and noise from the T1-guided scaffold, especially in clinically important ROIs (white matter, basal ganglia, etc.).

The topology persistence loss helps reduce point clouds with limited contribution to the shared geometry and improves robustness to small anatomical shifts across modalities. We’d like to include an expanded ablation table with additional metrics reflecting geometry preservation, anatomical fidelity, and speckle/noise suppression. 2.For the DWI and ASL experiments, generally we noticed that GS-based methods are more sensitive to deformable distortions in diffusion and perfusion imaging. Some of these artifacts can be alleviated when the modalities are non-rigidly registered to structural modalities in other organs (e.g., prostate imaging). We did not include these experiments in the submission because they are not part of the core methodology of this paper. However, we consider distortion management an important future direction toward ultra-high-resolution and functional harmonization. We also thank the reviewer for pointing out the typos and organizational issues. These comments are very helpful for improving the Methods section. To Reviewer #3: We agree with your comment regarding the absence of real-world evaluation, and we will consider this an important direction for future work. We used the term “spatial harmonization” to emphasize the potential ability of our method to align multimodal imaging data with arbitrary spatial resolutions onto a shared structural reference, which extends beyond conventional “super-resolution” or “registration.” Your comments regarding notation clarity are also very valuable. To Reviewer #4: 1.Our method is designed for a common clinical setting in which structural isotropic T1w scans are routinely available. We agree with your concern regarding the justification of the term “zero-shot.” We used this phrasing to emphasize that our method naturally adapts to most brain MRI modalities without requiring external training data for the target modalities. 2.Regarding the assumption of “shared geometry,” we believe that TSE-based FLAIR generally shares the same geometry because it is also a structural imaging modality. We did observe structural bias in diffusion and perfusion imaging (see response 2 to Reviewer #2). However, in brain MRI this bias is relatively limited, and our method remains competitive with the baselines. We will gladly include further discussion of this issue as a future direction. 3.Most components of our GS algorithm are derived from MedGS, which we gratefully acknowledge as one of the baselines. From our perspective, MedGS is among the earliest works introducing Gaussian Splatting methods into medical imaging. We compared against most SOTA INR- and GS-based methods that do not require external training data. In contrast, many state-of-the-art deep learning methods are designed mainly for in-plane reconstruction or require a fixed super-resolution factor, which is not fully compatible with our setting. 4.In this work, we did not perform extensive ablation studies on hyperparameters because, for most subjects, 10k iterations were sufficient for all loss terms to converge adequately. We would like to evaluate more settings and include this analysis in future work, which may also help reduce the processing time per subject.

Meta-Review

Meta-review #1

Your recommendation

Provisional Accept
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.

Based on the positive consensus from all reviewers regarding the method’s practical applicability, I recommend acceptance.

back to top

Shared Gaussian Geometry for Brain MRI Spatial Resolution Harmonization Without External Training

Author(s):