Abstract

To achieve accurate diagnostic outcomes, it is often necessary to acquire multiple series of magnetic resonance imaging (MRI) with varying contrasts. However, this process is time-consuming and imposes a significant burden on patients and healthcare providers. While diffusion models have emerged as a highly effective tool for image synthesis, they face challenges in handling the complexities of real-world clinical data and may distort vital information during medical image synthesis. To address these issues, we propose MRDiff, a novel diffusion model for multi-contrast MR image synthesis. MRDiff leverages the intrinsic relationship between different contrast images to derive shared anatomical information based on MR physics equations. Our approach integrates MR physics-based signal regularization for proper content feature generation and employs self-content consistency training to capture accurate anatomical structures. Experimental results demonstrate that MRDiff outperforms existing methods by generating diagnostically valuable images, highlighting its potential for clinical applications in MR image synthesis.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3896_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ShiYej_Physicsdriven_MICCAI2025,
        author = { Shin, Yejee and Byeon, Yunsu and Son, Geonhui and Jang, Hanbyol and Hwang, Dosik and Kim, Sewon},
        title = { { Physics-driven Signal Regularization in Diffusion Models for Multi-contrast MR Image Synthesis } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15963},
        month = {September},
        page = {400 -- 410}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposed a MR image synthesis model that integrates MR physics-based signal regularization for proper content feature generation and employs self-content consistency training. The proposed method outperforms existing methods in synthesized images’ quality.

  • 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 develops a novel method to apply MR physics model in the latent space, which may help regularize the latent space.

    2. The decoder to generate image y is an interesting design, which improves the generator’s performance.

  • 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 MR signal model Eq. (6) is for Spin Echo MRI. However, the dataset contain MP-RAGE data, which are GRE MRI and have a different signal model. I’m not sure how the authors handle this issue.

    2. According to Table 1, the proposed method’s improvement over cDPM is small, especially for the In-house dataset.

  • 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.

  • 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
    1. I am not sure the importance of DWT and IWT modules in the network. Even without DWT, the CNN can extract high frequency and low frequency components implicitly. Can authors provide results with all the DWT and IWT modules removed?

    2. In Fig.1, why the generator doesn’t use the MR physics module? I think the authors can use the scan parameters of image x0 to implement MR physics module for generator, which may improve the performance.

  • 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 — could be rejected, dependent on rebuttal

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

    The paper has proposed some novel methods, but some details need to be clarified.

  • 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

    The main contribution is proposal of a diffusion model that includes physics-based modules to regularize the synthesis of MR images.

  • 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 overall design of the network is sound, with technically novel composition.

    • Presented results are fairly comprehensive and convincing.

    • A wavelet decomposition is used for multi-level analysis, a physics-module links between MRI sequences.

    • Writing is detailed and clear for the most part.

    • Self-consistency training seems to be helpful.

  • 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.
    • Numerical comparisons on two datasets suggest significant albeit limited performance improvements.

    • Statistical tests are parametric based on the assumption of normally-distributed test statistics, but PSNR/SSIM/FSIM/GMSD are non necessarily normal variables.

    • A main limitation of including a physics-module in a network is that the results will be biased towards the MR sequence signal equations. For spin-echo sequences, the presented equations are mostly valid. However, for many other sequences and for micro-macromolecular interaction effects that influence even spin-echo images, these equations will fall short of accuracy. This is a major limitation that should be discussed.

    • Another important discussion point is on whether the proposed method would be capable of handling unpaired source-target images (see Ref. [14] for instance).

  • 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.

  • 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?

    This is a carefully conducted study that presents a novel approach incorporating physics-based processing into a diffusion model, and achieving significant performance benefits over baselines. Experimental scope is sufficiently large to justify the claims in the paper, and the reliability gains from physical signal equations are likely to push the envelope of MRI synthesis.

  • 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

    Authors propose MRDiff, a conditional diffusion model to synthesize a missing MRI contrast from two other available MRI contrasts. Compared to other diffusion-based synthesis work, MRDiff has a more elaborated design to guide the encoding and decoding of image features. It incorporates MR physics models to encourage the encoded features to focus more on anatomical features rather than on contrast features specific to the condition images. A content consistency loss is introduced to constrain the model to extract consistent anatomical features across different posterior samples. The method also leverages discrete wavelet transform to better capture high-frequency details. The method was evaluated in an in-house spine dataset and a public brain dataset, outperforming six recent state-of-the-art image synthesis methods in most metrics. The ablation studies show the effectiveness of the proposed components.

  • 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.
    • Novelty: The method uses MR physics models to encourage the encoded features to function like tissue parameter maps, such as T1 and T2, which are intrinsic properties of the organ and are shared information across different imaging contrasts. It is an interesting way to separate tissue properties, which is more important for synthesis of new image contrast, from the acquisition-dependent contrast information (which is more specific to input images themselves). Such physics-driven design may have the potential to better guide the AI models and improve explainability. The content consistency loss is also an interesting design in the diffusion framework. It encourages the encoded features to focus more on the input conditions rather than the random noises. It has the potential to be applied to other conditional diffusion models and may be helpful to reduce variations across posterior samples.
    • Evaluation setup: The authors did comparison to six SOTA methods in two datasets, along with ablation studies, which effectively show the superior performance of the proposed designs.
  • 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.
    • Applicability to different input contrasts: The method necessitates the MR signal equations for the condition images. However, a simplified signal equation is not always readily available. For example, for T2 FS TSE in the in-house spine dataset, it may be harder to directly have an analytical equation. A separation of water and fat signals may be needed. For some other more complicated MR sequences, such as 3D T2 FLAIR, a dedicated simulation may be needed to have an accurate description of the physics, which is challenging to be incorporated into a diffusion framework. (But we don’t know how accurate the physics models need to be in this 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.

  • 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

    Thanks for the authors’ work! Here are some thoughts and comments to share:

    • Currently the MR physics module is introduced only in the decoder path, not in the generator path. It would be interesting to see whether adding MR physics module before the generator would be helpful. In that case, the MR parametric features will explicitly be the shared features between encoder and decoder/generator. I understand that the authors didn’t do this possibly due to the lack of appropriate MR physics models for the output contrasts in their experiments (e.g., T2 FS TSE). (For T1 MPRAGE, there are simplified signal models.)
    • Eq. 7 shows a convolution layer after applying the signal equations. What kind of convolution is used (e.g., kernel size and how information in the two channels are exchanged)? Will this convolution break the physical relationship?
    • In Fig 2, I may prefer cDPM in terms of details. The proposed method looks to have an artifact at the position of the right arrow. (But I am not a radiologist.)
  • 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 incorporation of MR physics-based signal regularization and the self-content consistency loss into diffusion frameworks is novel based on my knowledge and may be inspiring to other researchers. The evaluation setup is reasonable, and the paper is clearly presented.

  • 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 reviewers for their thoughtful and constructive feedback, and for their positive evaluation of our work.

We agree with the reviewers’ concern regarding the limited availability of accurate MR signal models for certain target contrasts (e.g., GRE-based MP-RAGE, 3D T2 FLAIR, fat-suppressed TSE). Our design applies the MR physics module only in the decoder path, where it serves as a regularizer conditioned on the input contrast. This allows the model to benefit from known MR physics without imposing potentially invalid assumptions on the target contrast, for which accurate signal models are often unavailable or analytically intractable. We would also like to clarify that GRE-based sequences such as MP-RAGE were used only as target contrasts in our experiments, and not as inputs to the physics module. Therefore, the mismatch in signal modeling for MP-RAGE does not affect the model’s validity. We will make this design choice and its motivation more explicit in the final version.

We also acknowledge the reviewers’ comment regarding the modest numerical improvements over cDPM, especially on the In-house dataset. While the quantitative gains may appear small, our method yields notable improvements in preserving fine anatomical details, as observed in the qualitative results. In contrast, cDPM tends to produce overly smooth outputs due to strong denoising, which can artificially increase global metrics such as PSNR or SSIM. However, this comes at the cost of blurring clinically relevant structures and failing to preserve the intrinsic noise characteristics of MR images. Our model better balances denoising with structural fidelity, and we will highlight these findings with additional visual comparisons in the final version.

We appreciate the reviewer’s suggestion regarding the applicability to unpaired source-target image settings. While our current framework relies on paired supervision, it can be extended to accommodate unpaired data through architectural modifications such as adversarial learning or cycle-consistency constraints. This is a promising direction that we plan to explore in future work, and we will briefly discuss this in the revised manuscript.

Once again, we thank the reviewers for their time and thoughtful feedback.




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