Abstract

Susceptibility-induced distortions in diffusion MRI (dMRI) have significantly hindered the study of human brain pathways. Despite the availability of various B0-image-based and FOD-based distortion correction methods, significant distortions persist in certain brain regions, such as the brainstem. In this paper, we present a novel unsupervised learning-based distortion correction framework, named Multi-Modal Distortion Correction Network (MMDC-Net). To aggregate the complementary information in B0 and FOD images, we model distortion correction as a multi-modal optimization problem. We further establish a multi-modal learning paradigm for distortion correction, including modality-aware attention, intensity correction and gradient optimization. First, a Modality-Aware Attention Module (MAM) is designed to extract correction-related features from each modality. We introduce a differentiable intensity correction unit to correct the intensity bias in B0 images. Finally, PCGrad is integrated to mitigate gradient conflicts when optimizing similarity losses across the two modalities. Our method was evaluated on the HCP dataset with RL-LR phase encoding and the HCP-Aging and HCLV datasets with AP-PA phase encoding. The results demonstrated that MMDC-Net achieved superior and robust distortion correction performance compared to other state-of-the-art methods.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6218_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{FenJia_Joint_MICCAI2026,
        author = { Feng, Jianhui AND Li, Xiang AND Qiao, Yuchuan},
        title = { { Joint Multi-modal Learning for Susceptibility-Induced Distortion Correction in Diffusion MRI } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16888},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This study proposes a deep learning method for correcting susceptibility-induced geometric distortion for EPI-based diffusion MRI. It takes the same input (i. e. , a bliped-reversed pair of b=0) as FSL TOPUP, but introduces a multi-modal approach with both b=0 and FOD image derived from dMRI to constrain the optimization, showing improvement over TOPUP and existing DL methods.

  • 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 method is technically sound. Introducing FOD into distortion correction is a natural yet powerful design. 2.Evaluations include multiple datasets with well-designed metrics. I like the fact that the authors paid special attention to important brain structures that are interesting to neuroscientists but susceptible to distortions. 3.The paper is well-written, and figures are scientifically decent.

  • 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.

    Several technical details warrant clearer justifications and clarifications: 1.The integration of FOD image is helpful for distortion correction, but it also introduces another layer of complexity. Specifically: (1) Does this add requirement for the number of diffusion directions of the dataset to give robust FOD estimation? How would the method perform on data with limited directions (e. g. , 6 for DTI fitting that is common in clinical practice). (2) The original workflow for dMRI preprocessing is: topup using b=0 -> eddy correction -> downstream analyses; in the paper, what the authors proposed is: eddy for one PE -> FOD fitting -> eddy for another PE -> FOD fitting -> distortion correction using proposed MMDC-Net -> eddy for the entire dataset -> downstream analyses. This looks rather complicated. It is worth quantifying or at least discussing this additional burden. 2.Without FOD, the loss function is essentially very similar to FSL TOPUP using only b=0 as input but with a neural network optimization. However, looking at the results, b=0 only MMDC-Net is consistently outperformed by TOPUP. It’s worth finding out why.

  • 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.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

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

    The paper is well-written with sound methodology, but further justifications and clarifications are needed before final acceptance.

  • Reviewer confidence

    Very confident (4)

  • [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 addressed my comments by acknowledging the limitation. The paper is acceptable for publication.



Review #2

  • Please describe the contribution of the paper

    The paper proposes an unsupervised multi-modal distortion-correction framework for brain dMRI that jointly uses B0 and FOD images under reversed phase-encoding. The key contribution is a unified one-step learning framework that estimates a shared deformation field from both modalities, together with modality-adaptive feature weighting, a differentiable B0 intensity-correction unit, and PCGrad-based multi-objective optimization.

  • 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 motivation is reasonable. Combining B0 and FOD information for distortion correction is a natural idea. In brain dMRI distortion correction, B0 information can be insufficient in certain regions such as the brainstem, whereas FOD may provide additional structural guidance. This motivation is not arbitrary and is well connected to prior work in this area, including FODReg (2019), which used FOD to estimate the distortion field in brainstem connectome imaging, and DrC-Net (2022), which further developed FOD-based correction in a deep learning framework.

    Another strength is that the method is fairly comprehensive in design. The proposed FDM module attempts to combine both B0 and FOD information, and the use of PCGrad is also a reasonable design choice in a multi-objective setting. In addition, the paper includes extensive comparisons and ablation studies, which help support the empirical evaluation.

  • 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.

    A major concern is that much of the novelty appears to be compositional. The method mainly integrates several existing ingredients, such as FOD-based supervision, Jacobian-style intensity modulation, and PCGrad, rather than introducing a fundamentally new principle for distortion correction.

    Related to this, based on the paper’s description, the FDM module generates a two-channel attention map to perform spatially adaptive reweighting of B0 and FOD features. This seems closer to a modality weighting or attention-gating mechanism. In that sense, describing it as “feature disentanglement” feels somewhat overstated.

    Another issue is the scope implied by the title. The title suggests a more general framework for diffusion MRI distortion correction, but all experiments are conducted on brain datasets, and the second modality used here, FOD, is itself highly brain-specific. As a result, the empirical scope appears narrower than the title suggests.

  • 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

    1.Have the authors verified whether the FOD branch truly does not require a similar form of intensity correction, or is this mainly an empirical assumption? The current presentation feels a bit too absolute. The optimization of the FOD branch may not necessarily follow the same Jacobian-based intensity compensation as the B0 branch, but the paper does not sufficiently justify this design choice. More specifically, the relationship between FOD and this type of intensity modulation does not appear as direct or as physically grounded as it is for B0.2.The ablation study needs clearer explanation. It is currently unclear whether “+FDM,” “+IC Unit,” and “+PCGrad” are added independently or cumulatively. In particular, Table 2 is difficult to interpret because it is not clear whether the “+IC Unit” and “+PCGrad” rows are built on top of “+FDM” or evaluated as separate variants. Please clarify the exact component combination used in each row. 3.If the ultimate goal is distortion correction in diffusion MRI, it would strengthen the paper to show whether the learned shared deformation field also leads to direct improvement in high-b DWI or other diffusion-derived measures beyond FOD.

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

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

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

    My overall score is based on the fact that this is a solid and fairly complete paper, but with limited novelty. The motivation is reasonable, the method is well implemented, and the experiments are comprehensive enough to support the empirical findings. However, the main contribution feels largely compositional, combining known ingredients such as B0/FOD fusion, intensity correction, and PCGrad, rather than introducing a fundamentally new idea for distortion correction. I also think some wording is slightly overstated, especially around “feature disentanglement,” and the title implies a broader scope than the brain-specific evidence actually supports. Overall, this feels like a competent MICCAI paper, but not one that stands out strongly enough for a higher score.

  • 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.

    After reading the rebuttal, I am maintaining my score of 4 (Weak Accept). While the authors have adequately addressed my technical questions, the paper remains a borderline case. Specifically, I appreciate that they acknowledged the overstatement regarding “feature disentanglement” and agreed to rename the FDM to “Modality-Aware Attention Module” — this was my primary concern. The clarifications on the cumulative ablation study (Table 2) and the physical reasoning behind the FOD intensity bias are also reasonable. However, despite these corrections and the solid implementation, the overall methodological novelty is somewhat incremental, and the limited scope of the title remains an issue. It is a solid, well-implemented contribution suitable for MICCAI, but it sits right on the acceptance threshold.



Review #3

  • Please describe the contribution of the paper

    This paper presents a method for susceptibility-induced distortion correction in diffusion MRI. The method proposes to utilise complementary information in the B0 and estimated fibre orientation distribution to estimate a distortion field to correct distortions.

  • 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.
    • Paper is well written and motivation is clear and the method well explained.
    • Each of the components of the proposed method seems sensible and well-motivated.
    • The proposed method shows good performance compared to the reference methods, outperforming the widely used standard topup and other comparator DL methods.
    • Method shows good generalisability onto unseen data with same phase-encoding direction, comparable to existing DL SoA.
    • Ablation study is nice to demonstrate the benefit of each of the components.
  • 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.
    • My main thought is around the use of the FOD in the uncorrected images and whether that is fundamentally sound.
    • As I understand it, susceptibility distortions mean that signal that “should” be in certain voxels get accumulated or spread into other voxels which would mean that the FOD you estimate in those distorted voxels is wrong. Then, simply spatially transforming the SH coefficients with the estimated distortion field would not result in the correct FOD being reconstructed. So I think the blip up and blip down “corrected” FODs will not be the same even if the distortion could be perfectly corrected. I think this why FODreg and DrC-Net first use topup+eddy, to get initial FODs that are already close to correct and the FOD step is more of a refinement.
    • That said, it seems that the FOD method may be “good enough” since it seems to help.
    • Figure 3 is quite small and requires a lot of zooming in to parse.
    • What do the asterisks mean in Table 2?
    • In Section 3.1, HCP, HCP-Aging and HCLV acronyms are introduced without definition
  • 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.

    (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?

    Overall, this is a well-presented paper demonstrating a method that improves upon the state-of-the-art with a well-motivated and implemented method. The proposed distortion correction approach is evaluated well and shown to perform as well as, or better than all comparator methods.

  • 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 responses to my questions were satisfactory and their proposed changes will fix my main comments.



Author Feedback

We thank the reviewers for their insightful comments. Reviewer#1 1.Does incorporating FOD add requirements for the number of diffusion directions? How would the method perform on data with limited directions? FOD estimation requires sufficient directions; fewer directions degrade reconstruction precision and correction performance. To address this, integrating deep learning to estimate robust FODs from limited-direction data represents a valuable future direction. 2.The proposed preprocessing workflow is complicated. What is the additional computational burden? While prior eddy current correction and FOD reconstruction requires extra computation compared to B0-based methods, this is a necessary step for all FOD-based methods. Fast deep learning tools for these steps now greatly reduce this burden. We will discuss this computational trade-off in the discussion of the revision. 3.Why is the b=0 only MMDC-Net consistently outperformed by FSL TOPUP? TOPUP uses iterative optimization with a multi-resolution pyramid, achieving excellent instance alignment at the cost of extremely long running times. Conversely, our B0-only baseline is a single forward-pass network lacking iterative multi-scale refinement, explaining the performance gap.

Reviewer#2 1.The novelty appears compositional, and the term “feature disentanglement” is overstated. Our primary contribution lies in the joint multi-modal modeling of the susceptibility-induced distortion correction issue and the specific modules designed to resolve modality conflicts. We agree the term may be overstated. To accurately reflect its function of recalibrating regional feature contributions, we will rename FDM to the “Modality-Aware Attention Module” in the revision. 2.The title implies a broader scope, but the method and experiments are highly brain-specific. Susceptibility-induced distortions are severe in the brain (e.g., near brainstem/orbitofrontal tissue-air boundaries) and critically impact downstream connectome analyses. Most existing susceptibility distortion correction works are exclusively validated on brain datasets. Future work will extend our framework to other anatomies, such as the spine, to comprehensively evaluate its generalizability. 3.Does the FOD branch truly not require Jacobian-based intensity correction? As visually evident in our data, the intensity bias (signal pile-up or loss) observed in raw images like B0 does not manifest in the corresponding regions of the FOD images. As shown in the comparison of the same spatial location in Fig.1(c), the FOD exhibit no intensity distortion compared to B0.4.Are the components in the ablation study (Table 2) added independently or cumulatively? They are added cumulatively. We will update the Table 2 caption to state: “Components are added cumulatively to the baseline.” 5.Does the method improve high-b DWI or other diffusion metrics beyond FOD? The estimated deformation field is applied to all raw images. Our evaluation already uses Fractional Anisotropy (FA)—a tensor-derived metric calculated from corrected diffusion images. The improved FA MSD scores (Tab.1, Fig. 2) explicitly prove our method successfully corrects raw DWIs and enhances downstream diffusion metrics beyond FODs.

Reviewer#4 1.Is computing and transforming FODs directly on uncorrected, distorted images fundamentally sound? This is valid and widely accepted. Because local diffusion characteristics are consistent, raw DWIs and B0 images share the identical spatial distortion. And the voxel-wise FOD estimation introduces no extra deformation. Consequently, the uncorrected FODs preserve the original distortion and can be used to estimate the deformation field. 2.Fig.3 is too small. What do the ‘’ mean in Tab.2? Acronyms in Sec 3.1 are undefined. We will correct these by: 1) enlarging Fig. 3 to show key results; 2) clarifying in Tab.2 that ‘’ denotes baselines lacking proposed modules; 3) defining all dataset acronyms (HCP, HCP-Aging, HCLV) in Sec. 3.1.




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.

    The paper is solid overall, with sound methodology, good experimental validation, and generally positive reviewer consensus. However, concerns about limited novelty, potential overstatement of contributions, and clarity around key claims (e.g., feature disentanglement and generalizability) should be addressed before final acceptance. The rebuttal should specifically clarify the true methodological novelty, temper overstated claims, and better justify the broader applicability of the approach beyond the current brain-specific setting.

  • 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 reviewers recommended acceptance. The authors have adequately addressed the reviewers’ questions and concerns. The paper is therefore recommended for acceptance.



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.

    All reviewers state that their concerns were addressed in the rebuttal.



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.

    Reviewer consensus.



back to top