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

Abstract

Magnetic resonance imaging (MRI) is an invaluable tool for clinical and research applications. Yet, variations in scanners and acquisition parameters cause inconsistencies in image contrast, hindering data comparability and reproducibility across datasets and clinical studies. Existing scanner harmonization methods, designed to address this challenge face limitations, such as requiring traveling subjects or struggling to generalize to unseen domains. We propose a novel approach using a conditioned diffusion autoencoder with a contrastive loss and domain-agnostic contrast augmentation to harmonize MR images across scanners while preserving subject-specific anatomy. Our method enables brain MRI synthesis from a single reference image. It outperforms baseline techniques, achieving a +7% PSNR improvement on a traveling subjects dataset and +18% improvement on age regression in unseen scanners. Our model provides robust, effective harmonization of brain MRIs to target scanners without requiring fine-tuning. This advancement promises to enhance comparability, reproducibility, and generalizability in multi-site and longitudinal clinical studies, ultimately contributing to improved healthcare outcomes.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1897_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/daniel-scholz/cacd

Link to the Dataset(s)

IXI dataset: https://brain-development.org/ixi-dataset/ ONHarmony dataset: https://openneuro.org/datasets/ds004712/versions/1.0.2

BibTex

@InProceedings{SchDan_Contrastive_MICCAI2025,
        author = { Scholz, Daniel and Erdur, Ayhan Can and Holland, Robbie and Ehm, Viktoria and Peeken, Jan C. and Wiestler, Benedikt and Rueckert, Daniel},
        title = { { Contrastive Anatomy-Contrast Disentanglement: A Domain-General MRI Harmonization Method } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15965},
        month = {September},
        page = {99 -- 109}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper propose a domain-general scanner harmonization algorithm consisting of a conditioned diffusion model and an anatomy-content-disentanglement module, which is trained by a traditional DDIM loss and two supervised contrastive losses. The experimental results on three datasets show the effectiveness of proposed method.

  • 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 idea of combining feature disentanglement and diffusion model is interesting and novel.
    2. The experimental comparison is comprehensive and sufficient.
  • 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. My major concern is the anatomy and contrast dimensions in disentanglement module in Section 3.2. If I understand correctly, an input image is embedded into a 256-length contrast vector and a 256-length anatomy vector, which are then fed into the diffusion model as condition for generating the harmonized MR image. It is fine for the contrast vector, and my question is how or why such a low-dimension anatomy vector contains the sufficient information to guide the diffusion model accurately generating the MR image while preserving subject-specific anatomy. Could the authors provide some clarification on this point?

    2. In Section 3.2, “we sample a set of Naugs contrast augmentations …”, and Naugs is set to 9 in Section 3.4. It is critical in proposed harmonization method and generally preferable for this value to be as high as possible. In fact, I believe that identical random augmentations for both input images is a more direct and sensible way. Would the authors explain this in a bit more detail?

  • Please rate the clarity and organization of this paper

    Good

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

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    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?

    Although there are some unclear points in methodology section, the basic idea is novel and interesting, and the experiments are sufficient.

  • 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 domain-general MRI harmonization framework that disentangles anatomy and scanner/contrast information using a conditioned diffusion autoencoder and contrastive learning. One of the key contributions of the work compared to existing literature is its domain generalization. Experiments are relatively thorough, including both voxel-level similarity measurements and downstream tasks.

  • 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. I like the use of domain augmentation techniques, such as GIN, to build a contrastive learning framework. It fits well with the goal of learning contrast-invariant features and offers a logical way to generate diverse examples for training.

    2. The authors are addressing a very relevant problem in MRI harmonization with some in-depth thoughts. Many existing methods struggle with domain shifts themselves, which limits their real-world performance. The paper’s focus on generalization across scanners without relying on traveling subjects is an important and valuable contribution.

    3. The paper is clearly written and well-organized. The notation is easy to follow, and the authors do a good job of explaining the motivation, laying out the assumptions, and walking the reader through the method step by step.

  • 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. I have some concerns regarding the consistency of the evaluation. The reported performance of HACA3 [27] was based on non-skull-stripped images, while the experiments in this paper were conducted on skull-stripped images using the provided pre-trained weights. This raises questions about the fairness of the comparison, especially since pre-processing steps like skull stripping can significantly affect model performance.

    2. The choice of downstream tasks (age prediction and site classification) may not fully capture the effectiveness of harmonization. In particular, site classification can be influenced by population differences or cohort-specific biases, not just scanner-related variation. I recommend some more anatomy focused tasks, like segmentation. This would better reflect whether anatomical information is preserved and scanner-induced variations are effectively removed.

    3. Related to the previous point, the method operates on 2D slices, but it’s unclear whether the model can produce spatially consistent 3D volumes. In neuroimaging research, maintaining 3D coherence is crucial, and showing some form of slice-wise consistency (e.g., smoothness across slices, or consistency in anatomical structure) would significantly strengthen the applicability of the method for real-world use.

  • Please rate the clarity and organization of this paper

    Good

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

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    These are not critiques, but rather suggestions and comments the authors may consider for future work.

    1. While the paper demonstrates promising performance, the current evaluation includes only two comparison methods. This is a little bit limiting in my opinion. There are several recent efforts in domain-generalized harmonization that approach the problem from similar angles. I recommend exploring and comparing with these methods [1-3] in future work.

    2. The use of synthetic paired images via contrast augmentation for contrastive learning is a smart idea, and it fits well within the scope of single-modality harmonization. There are also related works in single-contrast harmonization [3-4] that share similar goals but take different perspectives, which might offer useful insights.

    That said, from a broader viewpoint I want to take a step back. I believe the future of harmonization should be trending towards incorporating multi-contrast information (such as T1, T2, and FLAIR), instead of focusing on a single modality. This is why I found the approach in HACA3 particularly appealing. These methods aims to maximize the use of available data across the clinical workflow. A multi-modal harmonization model would be better to handle variability in input quality, sequence availability, and diverse acquisition protocols. While the current paper takes an important step toward contrast-invariant learning within a single modality, I’d encourage the authors to consider extending this idea into a more flexible, multi-contrast framework in the future.

    1. To further improve the robustness and usability of the proposed method, I suggest two specific directions:
      • Analyze the structure of the learned latent space to better understand how anatomical and contrast features are separated. This could include latent traversals, cluster visualizations, or mutual information measures.
      • As the method operates on 2D slices, demonstrating slice-to-slice consistency and producing coherent 3D reconstructions would increase its practical value for neuroimaging applications. Show how the method performs in real-world tasks, such as longitudinal tracking or population-level analysis, would further validate its clinical utility.

    [1] Xu et al., SiMix: A domain generalization method for cross-site brain MRI harmonization via site mixing, NeuroImage, 2024.

    [2] Beiazze et al., Harmonizing Flows: Leveraging normalizing flows for unsupervised and source-free MRI harmonization, Medical Image Analysis, 2024.

    [3] Wu et al., Disentangled latent energy-based style translation: An image-level structural MRI harmonization framework, Neural Networks, 2025.

    [4] Zuo et al., Disentangling a Single MR Modality, MICCAI, 2022.

  • 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 presents a well-motivated and clear designed method for domain-general MRI harmonization. The authors showed promising results and practical relevance. Despite some limitations in evaluation (see weaknesses), the core ideas are novel, clearly presented, and represent a meaningful contribution to the field.

  • 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 #3

  • Please describe the contribution of the paper

    We present a novel method that utilizes a conditioned diffusion autoencoder combined with a contrastive loss and domain-agnostic contrast augmentation to harmonize MR images across different scanners, while preserving the subject-specific anatomy. This approach allows for the synthesis of brain MRI from a single reference image. Our method outperforms existing baseline techniques, achieving a 7% improvement in PSNR on a traveling subjects dataset and an 18% improvement in age regression on unseen scanners.

  • 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 paper introduces a novel approach that leverages a conditioned diffusion autoencoder, combined with a contrastive loss and domain-agnostic contrast augmentation, to harmonize MRI scans across different scanners while maintaining subject-specific anatomical features. This innovative formulation enhances the robustness of the model across diverse scanning environments.
    2. A key strength of the method is its ability to synthesize brain MRIs from a single reference image, eliminating the need for paired or multiple scans. This capability significantly simplifies the harmonization process, making it more feasible for real-world clinical applications.
    3. The approach outperforms existing baseline methods, demonstrating a 7% improvement in PSNR on a traveling subjects dataset and an 18% improvement in age regression for scans from unseen scanners. These strong performance gains underscore the effectiveness of the proposed method, particularly in generalizing to new, previously unseen data.
  • 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. The current method operates on 2D slices due to the computational cost associated with processing 3D data. Although the model can be adapted for 3D volumes, this limitation restricts its scalability and applicability to larger datasets and 3D volumetric tasks, such as those involving full-brain or organ-specific segmentation. This reduces the method’s potential for real-world clinical applications where 3D MRI data is commonly used for diagnosis and treatment planning.
  • Please rate the clarity and organization of this paper

    Satisfactory

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

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

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

    (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?

    I recommend accepting the paper because the problem the author addresses is highly relevant and real-world. The proposed method offers a promising solution to the challenging issue of MRI scanner harmonization, which is critical for improving the reproducibility and comparability of medical imaging data across multiple centers. The innovative approach and strong performance on various datasets further support its potential impact on clinical research and practice.

  • 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 #4

  • Please describe the contribution of the paper

    The manuscript describes a method for MRI harmonization across scanners. The main idea is to use diffusion autoencoders (DiffAE) (CVPR 2022) for conditional image generation for mapping images from a source to a target domain. DAEs pair a denoising diffusion implicit model (DDIM) with an additional encoder whose outputs are used to condition the image generation of the DDIM. The novelty of this manuscript is in the training of the encoder such that the encoder’s outputs are disentangled into anatomy and contrast information. This is achieved using a contrastive learning setup. The method is evaluated for scanner harmonization (in terms of reconstruction quality of images of another scanner and in terms of fooling a trained classifier to detect the scanner type) as well as for a downstream age regression task.

  • 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. An important problem of MRI harmonization is tackled in the paper. The proposed method builds on recent work on diffusion autoencoders, which takes an important step towards getting meaningful latent representations from diffusion models. This work further builds on diffAEs by optimizing the training of the encoder to get disentangled shape and intensity representations.
    2. Evaluations on direct MRI harmonization as well as on downstream tasks validates the usefulness of the proposed method.
    3. The paper is clearly written, well structured, and easy to follow.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    This paper has no major weaknesses in my opinion. As minor points,

    1. Please rotate the images by 90 degrees in the visualizations. Brain axial images are usually visualized that way. See figure 2 of citation [5] for reference.
    2. Please include statistical significance tests in quantitative results.
  • Please rate the clarity and organization of this paper

    Good

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

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    The paper proposes a principled extension to diffAEs and contains solid validation on MRI harmonization tasks.

  • 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 proposes a principled extension to diffAEs and contains solid validation on MRI harmonization tasks.

  • 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 sincerely thank the reviewer for their positive evaluation of our work and their insightful feedback: We are pleased that our approach was regarded as “interesting and novel”, that our “experimental comparison [was] comprehensive and sufficient” (R1), and that the manuscript was described as “clearly written, well-organized, and easy to follow” (R2, R4). We also appreciate the recognition of our method’s substantial performance gains by our method, including the reported “7% improvement in PSNR” and “18% improvement in age regression” (R3). Below, we address the major points raised: Anatomy & Contrast Disentanglement Clarification (R1): Combining the stochastic encoding via DDIM inversion and the anatomy vector facilitates faithful reconstruction of the anatomical structures. Preliminary experiments on the influence of the initial noise map x_T and z_a indicate that the primary contribution to the anatomical representation originates in z_a, while finer anatomical details are encoded within x_T.

We agree that generally N_augs should be as large as possible to force meaningful contrast representations, similar to SimCLR. In our framework, the same set of intensity augmentations is applied to all subjects within a batch (e.g., both subjects in Figure 1), thereby introducing contrast variations that simulate scanner differences. Anatomical variability is introduced through the inclusion of different subjects in each batch. Evaluation of HACA3 (R2): For a fair and direct comparison, HACA3 inference is performed on full, non-skull-stripped images while evaluation metrics are computed only within the brain mask across all baselines. 3D & Anatomy-Focused Evaluation (R2, R3): We thank the reviewers for their valuable suggestions and agree that segmentation-based tasks, 3D spatial-consistency analyses, and multi-contrast data represent promising directions for future projects to further improve our approach. Presentation Improvements (R4): To align with standard neuroimaging conventions, we will rotate all axial brain images by 90° in the final manuscript.



Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    N/A



back to top