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

Abstract

Multi-contrast MRI synthesis is inherently challenging due to the complex and nonlinear relationships among different contrasts. Each MRI contrast highlights unique tissue properties, but their complementary information is difficult to exploit due to variations in intensity distributions and contrast-specific textures. Existing methods for multi-contrast MRI synthesis primarily utilize spatial domain features, which capture localized anatomical structures but struggle to model global intensity variations and distributed patterns. Conversely, frequency-domain features provide structured inter-contrast correlations but lack spatial precision, limiting their ability to retain finer details. To address this, we propose a dual-domain learning framework that integrates spatial and frequency domain information across multiple MRI contrasts for enhanced synthesis. Our method employs two mutually trained denoising networks, one conditioned on spatial domain and the other on frequency domain contrast features through a shared critic network. Additionally, an uncertainty-driven mask loss directs the model’s focus toward more critical regions, further improving synthesis accuracy. Extensive experiments show that our method outperforms state-of-the-art (SOTA) baselines, and the downstream segmentation performance highlights the diagnostic value of the synthetic results. Code and model hyperparameters are available at https://github.com/sanuwanihewa/D2Diff

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2110_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/sanuwanihewa/D2Diff

Link to the Dataset(s)

https://www.med.upenn.edu/cbica/brats-2019/

BibTex

@InProceedings{DaySan_D2Diff_MICCAI2025,
        author = { Dayarathna, Sanuwani and Peiris, Himashi and Islam, Kh Tohidul and Wong, Tien-Tsin and Chen, Zhaolin},
        title = { { D2Diff: A Dual-Domain Diffusion Model for Accurate Multi-Contrast MRI Synthesis } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15961},
        month = {September},
        page = {131 -- 140}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This research proposes a dual-domain diffusion model (D2Diff) designed to synthesize multi-contrast MRI images, with the goal of improving image quality and advancing medical imaging applications.

  • 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) Enhanced Feature Correlation: The model incorporates a multiscale frequency feature fusion module, which strengthens the relationship between features across different MRI contrasts, ultimately improving synthesis accuracy.
    (2) Improved Focus on Critical Regions: Through the use of an uncertainty-aware mask loss and a critic network, the model enhances its attention to key regions while enabling smooth and collaborative training within the dual-domain framework.
    (3) Validation Through Downstream Tasks: The model’s performance is validated on downstream tasks, underscoring its practical applicability and effectiveness in real-world 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.

    (1) High Computational Demands: The dual-domain network structure increases computational costs and introduces challenges in training, potentially limiting its practicality in resource-constrained settings.
    (2) Absence of Comparisons with Recent Models: The study does not include experimental comparisons with the latest state-of-the-art methods, which limits the strength of the evidence supporting the model’s advancements.
    (3) Lack of Discussion on Generalization: The research does not examine the model’s ability to generalize across diverse datasets or applications, leaving its broader utility unclear.

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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.

    (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?

    (1) Include Comparisons with Leading Methods: To strengthen the evaluation of the model, the study should compare its performance with recent state-of-the-art approaches in MRI image synthesis.
    (2) Expand Visualization Across Modalities: Adding more visual examples for healthy datasets across different modalities would provide clearer insights into the model’s synthesis capabilities.
    (3) Address Ambiguity in Blurry Lesions: While the model demonstrates slight improvements in generating clearer images for blurry lesion areas, the study should discuss whether these changes may introduce new limitations or challenges.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    This paper lacks completeness in its content and framework. The author’s response does not include comparisons with the latest methodologies or adequate visualizations of multimodal health data. Papers with such deficiencies are, in my opinion, unsuitable for publication.

    Specifically, the most recent method referenced in the study is from 2023, which does not align with the current standards. This omission suggests that the comparative analysis of the proposed method is both incomplete and unreasonable. Furthermore, the study fails to incorporate experimental comparisons with cutting-edge, state-of-the-art methods, which weakens the evidence supporting the model’s claimed advancements.

    To enhance the study’s clarity and impact, it would be beneficial to expand visualizations by including more examples of healthy datasets from diverse modalities. This addition would provide a more comprehensive understanding of the model’s synthesis capabilities and strengthen its overall contribution.



Review #2

  • Please describe the contribution of the paper

    The paper introduces D2Diff, a novel dual-domain diffusion model designed for accurate multi-contrast MRI synthesis. The framework integrates both frequency-domain features and spatial-domain features dual-encoder architecture. The method achieves state-of-the-art synthesis performance on both healthy and tumor MRI datasets and demonstrates clinical utility through improved downstream segmentation.

  • 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 dual-domain diffusion model, combined with the uncertainty-aware mask loss, showcases technical depth and innovation.
    2. The paper presents comprehensive experimental results that demonstrate superior performance on both tumor and healthy MRI datasets. Additionally, the authors show that the synthesized images enhance downstream segmentation performance, underscoring the model’s clinical utility.
  • 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 application of frequency-guided learning (H1) does not appear to significantly tumor segmentation, when comparing the performance of to the spatial-guided (H2) variant and the full D2Diff model. Are these performance improvements statistically significant?
    2. Figures 2 and 3 do not clearly highlight the advantages of the proposed method. Including error maps or difference images would help visually demonstrate where D2Diff outperforms baseline methods.
    3. The authors acknowledge that multi-site variability may affect segmentation performance (Section 4). For a fair comparison, inter-site variability should be addressed first. Otherwise, using the “Complete” setup as a reference is questionable, and the observed performance gains may not reflect the improvements due to better synthesis quality.
    4. The method involves two denoising networks and multiple input modalities, which likely increases computational load. However, the paper lacks details on runtime or hardware requirements.
  • 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.

  • 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.

    (4) Weak Accept — could be accepted, dependent on rebuttal

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

    The paper presents a technically novel method that achieves strong results in both synthesis and segmentation tasks. There are only minor concerns regarding the ablation study results.

  • 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

    The paper presents a dual-domain learning framework for multi-contrast MRI synthesis, integrating spatial and frequency domain information to enhance synthesis accuracy. The method employs two mutually trained denoising networks conditioned on spatial and frequency domain features, guided by a shared critic network. An uncertainty-driven mask loss is introduced to focus the model on critical regions, improving synthesis quality and diagnostic value.

  • 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 paper is methodologically sound. The dual-domain approach effectively combines spatial and frequency domain features, addressing limitations of existing methods.
    • Extensive experiments demonstrate superior performance compared to state-of-the-art baselines, with validation through downstream segmentation tasks.
  • 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 could benefit from a more detailed description of the experiments to improve reproducibility and clarity.
    • It is unclear whether the uncertainty-aware mask loss genuinely captures uncertainty or working similarly as deep supervision or merely adds more parameters. No experiment results support the claim.
  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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. The paper could benefit from a more detailed description of the experiments to improve reproducibility and clarity. (1) The paper lacks essential details in the experiment design and description, such as training hyperparameters, network architecture specifics, and other critical parameters. (2) Additionally, the parameters for the comparison methods are missing, which makes it difficult to understand the basis of comparison and replicate the results. (3) In Table 2 - d, it is unclear how the synthesis pipeline work given no modalities are used.
      (4) Figure 4 is too small to read, hindering the ability to assess the visual quality of the results. The authors may consider moving some of the method details to supplementary materials to give more room for a thorough experiment description. This would enhance the paper’s transparency and facilitate better understanding and replication of the experiments.

    2. The title of Table 1 indicates that a paired t-test was conducted to assess the statistical significance of the proposed method. Could you please clarify whether the experiments were performed for each pair of D2Diff compared to the other methods?

    3. The proposed method seems to be a 2D approach, which was not clearly stated in the paper. This is a limitation that should be mentioned, as it affects the applicability of the method to 3D MRI data.

  • 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 — could be accepted, 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 due to its innovative dual-domain learning framework and promising results, which demonstrate superior performance compared to state-of-the-art baselines. However, the paper lacks detailed descriptions of the experiments, and it is unclear whether the uncertainty-aware mask loss genuinely captures uncertainty or merely adds more parameters.

  • 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 have addressed several of my concerns. However, their response on whether the uncertainty loss effectively targets uncertain regions remains unconvincing, as it relies solely on a single qualitative example in Figure 1. Nonetheless, I voted to accept the paper, as it is methodologically sound and likely to be of interest to the community.



Author Feedback

We thank all the Reviewers. R1-Significance of H1:To clarify, H1 plays a crucial role in enhancing the model’s performance by effectively capturing global intensity variations across contrasts. H1 alone performs slightly lower than H2 which is expected as H2 focuses on fine-grained anatomical features more relevant to local lesion characteristics. D2Diff leverages the complementary strengths of both, achieving substantial improvements over either alone. Notably, H1 alone outperforms all methods, demonstrating high structural similarity, even for challenging task like T1ce synthesis. R1-Error Maps:We will add the error maps for the results in Fig.2 and 3. R1-Inter-site variability:We acknowledge the Reviewer’s valid concern. However, our intent is not to treat the “Complete” setup as a performance ceiling, but rather as a reference that reflects actual multi-site data. As in Table2, D2Diff achieves better performance than the “Complete” setup in several cases, indicating model’s robustness in mitigating artifacts and contrast inconsistencies. Similar results from SynDiff[14] also support that diffusion-based methods enhance robustness of synthesis. R1,R3-High computational Demand:D2Diff does not substantially increases the computation load. To keep overhead low, it employs two shallow generators. Compared to SOTA model like SynDiff[14], which follows a conventional diffusion-GAN architecture, D2Diff has 68.586M parameters, while SynDiff has 67.465M. Also, D2Diff’s average sampling time is 310.10ms, only slightly above SynDiff’s 303.70ms, showcasing that despite the use of two denoising networks D2Diff does not introduce a substantial computational burden. We will clarify this in the paper. The hardware requirements for training were included in our shared git repo. R2-Experimental Details:Due to space limit, we have included essential experimental details along with citations to the architectural references. Other training hyperparameters were included in our shared Git repo. For SOTA comparisons, we used training settings from their official repositories and will add these details to our GitHub. R2-Table 2(d),Fig.4:To clarify, In the table, an ‘X’ indicates that the corresponding contrast has been replaced by a synthetic contrast, while a ✓ denotes the use of the actual contrast from the test dataset. Line d, all ‘X’ means that the segmentation model receives all 4 synthetic contrasts as input. We will allocate more space for Fig.4. R2-Uncertainty loss:The mask loss captures uncertainty, as demonstrated by U1 and U2 in the sample case in Fig. 1, highlighting areas of high uncertainty (yellow pixels), particularly around tumor boundaries, improving lesion delineation. Our ablation study validates the significant impact of this. R2-t-test:We performed a paired mean t-test between D2Diff and the second-best method. We will clarify this in the paper. R2-2D model:Our model uses 2D data and we agree that this should be properly mentioned. We will clarify and acknowledge it as a limitation. R3-Generalizability:We agree that comprehensive evaluation is essential for generalizability. However, we evaluated our model on both multi-site pathological and healthy datasets, demonstrating its generalizability across wide range of anatomical variations and imaging conditions. Our results also support its potential utility in downstream clinical applications requiring reliable multi-contrast imaging. We will include this discussion in the paper. R3-Evaluation:We present a comprehensive comparison with 6 SOTA methods, each across 7 synthesis tasks. This includes SynDiff[14], a recent SOTA known for its strong performance and reproducibility, supported by validation in recent benchmark surveys [7]. This ensures a fair and reliable assessment of our model. R3-Lesion Ambiguity: We agree that our method performs better in lesion synthesis addressing blurry regions. We will discuss these limitations in terms of clinical accuracy in the paper.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • 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

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



back to top