List of Papers Browse by Subject Areas Author List
Abstract
Pooling multi-site T1-weighted MRI increases power but introduces scanner-driven variability that can bias analyses. We propose a 3D patch-based conditional diffusion framework for MRI harmonization that preserves fine anatomical details while aligning image contrast across sites. Our method constructs site-invariant anatomical maps using gradient and high-frequency features, which guide patch-wise diffusion synthesis together with learnable site embeddings. Processing overlapping high-resolution patches avoids latent-decoder blurring and reduces memory requirements, enabling faithful preservation of fine anatomical detail. To further enhance anatomical fidelity, we employ manifold-constrained classifier-free guidance, stabilizing diffusion sampling while maintaining subject-specific structures. We evaluate our framework on a multi-site traveling-subject dataset and a clinically acquired pathological cohort, comparing against state-of-the-art harmonization methods. Results show that our approach effectively reduces site-related intensity variability, preserves anatomical and tissue-specific information, and maintains clinically relevant brain–cognition associations. These findings demonstrate that patch-based, anatomy-guided diffusion harmonization provides a robust, high-fidelity solution for multi-site MRI, with clear practical benefits for clinical datasets and downstream analyses. Code and model weights are available at https://github.com/barna-hache/PatchLevelDiffusionMRIHarmonization.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3157_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/barna-hache/PatchLevelDiffusionMRIHarmonization
Link to the Dataset(s)
SALD dataset: https://fcon_1000.projects.nitrc.org/indi/retro/sald.html
IXI dataset: https://brain-development.org/ixi-dataset/
OASIS-3 dataset: https://sites.wustl.edu/oasisbrains/home/oasis-3/
AIBL dataset: https://adni.loni.usc.edu/aibl-australian-imaging-biomarkers-and-lifestyle-study-of-ageing-18-month-data-now-released/
SRPBS dataset: https://www.synapse.org/Synapse:syn22317079
BibTex
@InProceedings{HacBar_Anatomy_MICCAI2026,
author = { Hache, Barnabé AND Roca, Vincent AND Kuchcinski, Grégory AND Manouvriez, Dorian AND Lopes, Renaud},
title = { { Anatomy- and Site-Guided 3D Patch Diffusion for Robust MRI Harmonization } },
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 presents a 3D patch-based conditional diffusion framework for multi-site structural MRI harmonization. The framework leverages derived site-invariant anatomical maps to preserve structural details while harmonizing scanner-specific contrast. Evaluation is performed on robust, diverse datasets spanning international multi-site studies and a traveling subject study.
- 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 investigators appropriately generated a curated, robust training dataset assembled from multiple public repositories of multi-site international medical imaging studies and a robust clinical cohort of patients with Alzheimer’s disease for testing.
Methodological refinement is demonstrated in (1) performing ablation studies to optimize classifier-free guidance, (2) assessing site variability of supposedly site-invariant anatomical priors, and (3) assessing harmonization performance of clinically meaningful structure-cognition correlations. However, further methodological detail is needed as specified below.
- 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.
Although the presented 3D patch-based conditional diffusion framework is intriguing, and training and testing datasets are robust, greater clarity is needed on methodological reasoning and details throughout the paper.
Derivation of anatomical representations that are supposedly site-invariant is a critical step in the proposed framework for structural MRI harmonization. As such, greater detail should be provided on how Maximum Mean Discrepancy (MMD) tests in Ablation Study 2 demonstrate site-invariance of the anatomical conditioning map. What specific distribution pairs were MMDs calculated between? How was MMD averaged per site when MMDs were calculated on inter-site discrepancies? Presumably MMDs were calculated on same subject, same map (raw vs. anatomical map), between sites, e. g. , subject1-site1-raw histogram vs. subject1-site2-raw histogram or subject1-site1-anatomical histogram vs. subject1-site2-anatomical histogram. However, such detail should be explicitly specified for clarity and reproducibility in the paper. Visual or tabulated results, rather than single summary p-values, would also be helpful for interpretation. As is, the site-invariance of anatomical representations is questionable.
In Task 1, the dependent variable in the linear mixed-effects model is not stated – a model formula should be provided for clarity. Presumably, the site fixed effect and subject random effect predicted GM and WM volumes normalized to intracranial tissue volume. If so, why were whole tissue volumes chosen for site- and subject-effect modeling as opposed to more specific memory- or disease-related regional volumes as in Task 2? Gross whole-GM and whole-WM tissue volumes are likely less sensitive to small scanner-specific differences in tissue boundaries than small regional volumes like the hippocampus or entorhinal cortex. Normalizing GM and WM volume to total intracranial tissue volume (GM+WM) may have further reduced sensitivity to site or scanner effects. Why was normalization to whole tissue volume chosen?
Overall, it’s unclear if the proposed 3D patch-based conditional diffusion framework is a superior method or step forward for structural MRI harmonization, relative to existing published techniques. STGAN and HACA3 outperform the proposed framework in matched-subject harmonization performance (SSIM, WD, PCC). STGAN, HACA3, and basic preprocessing outperform the proposed framework in mixed-effects modeling of brain tissue volumes, albeit not consistently. HACA3 is also the best-performing method for cognitive associations in 3/8 regions assessed, whereas the proposed diffusion framework is the clear best-performer in only 2/8 regions (possibly 3/8 if including middle temporal, though it is difficult to discern a difference between STGAN and Diffusion for middle temporal in Fig. 3). For what tasks or under what circumstances would the authors recommend the proposed framework be utilized over STGAN or HACA3?
While the curated training and clinical datasets are robust in spanning multiple international sites and studies, re-training and evaluating methods on all full available datasets would increase confidence in results.
- 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?
While the presented harmonization framework is intriguing, additional methodological details are necessary, as specified in weaknesses above. Further, it is questionable that the proposed framework is a significant or impactful step forward in structural MRI harmonization, based on presented results.
- 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 sufficiently clarified details on the site-invariance MMD and mixed-effects model analyses. Importantly, evaluation scripts will be made public upon acceptance, along with code and pretrained models, so analyses can be understood and reproduced even if specific details are missing from the manuscript text.
While the proposed method does not clearly outperform other techniques universally, the authors’ rebuttal that it is the most consistent and balanced across evaluations is true and well-received.
Review #2
- Please describe the contribution of the paper
The paper presents a 3D patch-based diffusion framework for multi-site MRI harmonization. The key idea is to disentangle anatomical structure and site-specific contrast via a site-invariant anatomical map and learnable site embeddings, enabling harmonization without requiring paired or traveling-subject data. The patch-wise diffusion design is used to preserve high-resolution anatomical details while remaining computationally feasible. Experiments on both traveling-subject datasets and clinical cohorts demonstrate strong image-level harmonization performance, along with some downstream analyses such as tissue consistency and cognition-related correlations.
- 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 unpaired setting is realistic and important. Avoiding reliance on traveling-subject data is a clear practical advantage.
- The method is well-motivated. The combination of anatomical conditioning and site embedding is intuitive and makes sense for disentangling structure vs contrast.
- Experiments are quite extensive.
- Patch-based diffusion is a reasonable design choice here, especially for preserving high-frequency anatomical details.
- 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.
- Anatomical representation: This is unfortunately the biggest concern of this paper. The motivation is clear: suppress site-related characteristics while preserving the anatomical traits. From what I understand (and how the paper presents), the motivation is along the lines of simple gradient-based features. Therefore, there could be many alternatives to this. In that sense, there is little justification behind the proposed anatomical representation approach.
- 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.
(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 anatomical representation could use further justification or experimental analyses. Otherwise, the paper clearly has some strengths especially the rigorous clinical analyses.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This paper proposes a 3D patch-based conditional diffusion framework for multi-site T1-weighted MRI harmonization. The method introduces anatomy-guided diffusion by constructing site-invariant anatomical maps derived from gradient and high-frequency features. It integrates these with learnable site embeddings to guide patch-wise synthesis. By operating on overlapping high-resolution patches, the framework reduces memory consumption and mitigates blurring commonly observed in latent diffusion approaches. Additionally, the incorporation of manifold-constrained classifier-free guidance aims to stabilize sampling while preserving subject-specific anatomical structures. The method is validated on both traveling-subject and clinical datasets, demonstrating improved harmonization performance and preservation of clinically relevant associations.
- 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 paper presents a well-motivated and technically sound framework that combines patch-based processing with conditional diffusion, which is a meaningful extension for high-resolution MRI harmonization tasks. The use of overlapping patches is particularly practical, as it addresses both memory constraints and the loss of fine anatomical detail often observed in full-volume or latent-space methods. 2.The introduction of anatomy-guided diffusion through gradient and high-frequency-derived maps is a novel and interesting design choice. This explicitly enforces structural consistency across sites and provides a principled way to disentangle anatomical information from scanner-specific contrast variations. 3.The incorporation of manifold-constrained classifier-free guidance is a thoughtful addition that improves sampling stability while maintaining subject-specific features. This reflects careful consideration of the trade-off between the strength of harmonization and anatomical fidelity.
- 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 paper does not sufficiently quantify the contribution of the proposed anatomical map components. While gradient and high-frequency features are claimed to guide harmonization, no ablation studies isolate their individual effects (e.g., removing gamma perturbation, gradient, or frequency components). This makes it difficult to assess whether the performance gains stem from the anatomical guidance or from the diffusion framework itself. 2.Although the method achieves strong global performance, the regional analysis reveals inconsistencies. Specifically, certain clinically important ROIs, such as the posterior cingulate and entorhinal cortex, show slightly worse correlations with MMSE compared to baseline methods, suggesting that the harmonization may not uniformly preserve region-specific information. This raises concerns about its reliability for localized analyses. 3.The novelty of the framework is somewhat incremental relative to recent diffusion-based harmonization approaches. Prior work on conditional diffusion for medical image translation and harmonization (e.g., diffusion-based MRI synthesis and style transfer methods) has already explored conditioning mechanisms and structural guidance, and the current work primarily extends these ideas through patch-based processing and feature engineering rather than introducing a fundamentally new formulation.
- 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?
This paper addresses an important problem in multi-site MRI analysis and proposes a technically solid and reasonably well-evaluated framework. The combination of patch-based diffusion and anatomy-guided conditioning is practical and shows clear empirical benefits, particularly in preserving global consistency and clinical associations. However, the lack of detailed ablation studies weakens the ability to attribute performance gains to specific components of the method. In addition, the observed regional inconsistencies in clinically relevant ROIs raise concerns about robustness, particularly for applications that require localized accuracy. Finally, while the method is well-engineered, its conceptual novelty relative to existing diffusion-based harmonization approaches is somewhat limited. Overall, this paper is promising but would benefit from deeper analysis and clearer validation of its key design choices, which leads to a weak reject recommendation.
- 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
Author Feedback
Dear Reviewers, Area Chairs, and Program Chairs,
We thank you for the careful reading and constructive feedback. We address the main points below.
To preserve anonymity, the GitHub repository is not included in the submission but is already prepared and will be made public after the review process. It contains the full training and inference code, pretrained weights, the exact train/test splits data, and all evaluation scripts used in the paper.
1) Clarification of the anatomical map / site-invariance analysis. We agree that this is a central step and should be described more explicitly. In Ablation Study 2, MMD was computed between every pair of acquisition sites (si ,sj), not between matched subjects. For each site, we pooled the histogram feature vectors from all scans acquired at that site, and computed the RBF-kernel MMD separately for raw-image histograms and anatomical-map histograms. Thus, each value corresponds to one site-pair comparison. The “per-site” summary is the average of that site’s MMD values against all other sites. The paired t-test compares raw vs. anatomical-map MMD on the same site pairs, so the reduction directly reflects lower inter-site discrepancy for the anatomical map. For brevity, we do not reproduce the full pairwise table here. Our anatomical map is intended as a simple low-level prior that suppresses scanner contrast while preserving structure, inspired by recent work on anatomical priors (10.1016/j.neuroimage.2025.121297). Gradient magnitude and high-frequency content are chosen because they are less tied to site-specific intensity statistics than raw signal, and random gamma perturbation further reduces leakage of contrast cues into the conditioning map.
2) Mixed-effects model in Task 1.For each method, we fitted a linear mixed-effects model with site as a fixed effect and subject as a random intercept: tissue_ratio ~ site + (1 | subject) The dependent variable is the GM or WM tissue ratio. These ratios are normalized by total intracranial volume, defined as GM + WM + CSF, to reduce inter-subject size differences and make comparisons across sites and harmonization methods more comparable.
3) Comparison with STGAN / HACA3 and when our method is most useful. We agree that no single method dominates every metric. Our claim is not that diffusion is uniformly best, but that it is the most consistent method in our experiments and provides a strong balance between harmonization and preservation of biologically meaningful signals. In the tissue-level analysis, our method is the only one that consistently achieves both low marginal R² and high conditional R² across GM and WM, whereas STGAN and HACA3 are stronger in one tissue but less consistent in the other. In the cognition analysis, our method achieves the highest global mean absolute correlation and improves over the preprocessed baseline. This is also consistent with Fig. 2, where our method shows fewer over- and under-segmentation errors than the Anatomy Target, suggesting more robust anatomical preservation. At the same time, we acknowledge that STGAN and HACA3 can be stronger in specific metrics or regions.
4) Nature of the contributions / ablations. We appreciate the request for more component-level justification. The manuscript already includes three targeted analyses: guidance ablation (CFG vs. CFG++), the site-invariance test for the anatomical prior, and downstream clinical preservation analyses. Together, these support the full design: patch-based high-resolution synthesis, anatomy-guided conditioning, and manifold-constrained guidance. We also performed an additional internal ablation of the anatomical representation components: removing the gradient component caused a marked SSIM drop (-0.28 drop), indicating that gradient information is the main contributor; removing the high-frequency component did not significantly reduce SSIM, suggesting a smaller contribution. Other components could also merit further investigation.
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 proposes a 3D patch-based conditional diffusion framework for multi-site MRI harmonization conditioned on an anatomical map and site-specific embeddings. Evaluation includes traveling-subject data and a clinical cohort with cognition-related downstream analyses. The reviewers all acknowledge that the problem is practically important, that the combination of anatomical conditioning with site embeddings is a sensible disentanglement strategy, and that the evaluation scope is substantial.
All three reviewers recommend Weak Reject (3), but the AC finds that these rejections are based on somewhat different reasons. R1 provides the most detailed and consistent critique (and the AC weighs it more into the current decision): the MMD-based evidence for site-invariance of the anatomical maps is ambiguously described (which distributions are compared, how results are averaged, and only p-values are reported), and the linear mixed-effects model formula for Task 1 is not stated. More importantly, R1 argues that the proposed method does not clearly outperform STGAN and HACA3 across the reported metrics; STGAN and HACA3 win on matched-subject SSIM/WD/PCC, the baselines can win on mixed-effects tissue volumes, and HACA3 is best for cognitive correlations in more ROIs than the proposed method. This undermines the significance claim. R2 argues that the construction of the anatomical map is under-justified: it amounts to gradient-style feature engineering and the paper does not explain why this particular construction is preferred. R3 notes that there is no component ablation isolating the contributions of the gradient, high-frequency, and gamma-perturbation parts, that certain clinically important ROIs (posterior cingulate, entorhinal cortex) show worse MMSE correlations than baselines, and that the novelty is incremental relative to existing conditional-diffusion harmonization work.
Most of R1 and R2’s concerns can be addressed through clarification and re-analysis of existing results (MMD pair specification, mixed-effects formula, discussion of when to use the proposed method vs. STGAN/HACA3). R3’s request for finer ablations of the anatomical-map components is the one concern that would typically require new experiments, but it is not on its own decisive. Given the uniform but moderate rejection score and the substantial amount of work the paper represents, I recommend inviting the authors to respond in rebuttal before a final decision. While preparing the rebuttal, the authors should carefully read reviewers’ comments and the MICCAI rebuttal guidelines.
- 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.
All three reviewers opened at weak reject in the initial review but they all gave different reasons. The rebuttal did a good job clarifying the site-invariance (MMD) analysis and the mixed-effects model, and reframing the contribution as the most consistent and balanced method rather than the best on every metric, which moved R1 to accept. The harder concerns, though, are only partly resolved on the submitted manuscript: R2’s point that the anatomical map is under-motivated gradient feature engineering and R3’s point that the paper lacks a component ablation and shows regional inconsistencies are answered mainly by a new ablation run during the rebuttal.
Weighing the one engaged accept (which I weigh the most in the initial review) against two non-response reviewers, and the authors overall effective rebuttal given the 4000 limit, I lean toward accept.
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 originally received consistent “weak reject” reviews with clear avenues for rebuttal. The extensive work that was put into the methodology and experimentation was recognized by the AC. I believe the authors have adequately addressed the concerns that led to weak rejects, including both novelty and clarification of the contrast-agnostic space. I recommend that this paper be accepted.
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 authors have adequately addressed the reviewers’ questions and concerns. The paper is therefore recommended for acceptance.
