Abstract

Diffusion-weighted MRI (DWI) is widely used for assessing tissue microstructure, with echo-planar imaging (EPI) sequences being the preferred acquisition method due to their fast speed. However, EPI-based DWI is highly sensitive to field inhomogeneities, leading to susceptibility-induced distortions that compromise image quality. Traditional correction methods, such as TOPUP, estimate displacement fields from a pair of reversed phase-encoding (reversed-PE) images to mitigate these distortions. While effective, these approaches suffer from high computational cost, limiting their clinical utility. In this study, we propose an unsupervised learning method for susceptibility artifact correction in EPI. A transformer-style convolutional network enhanced with deformable convolutions is developed to estimate the displacement field from a pair of reversed-PE images, followed by image unwarping and intensity modulation to generate the distortion-free images. This approach surpasses the performance of conventional U-Net-based methods in accuracy. Additionally, a spatially weighted smoothness loss is introduced to enhance robustness against noise in the input data so that the predicted displacement fields from a pair of low b-value DWI can be applied to correct other images with different b-values and diffusion directions from the same subject, optimizing acquisition and computational efficiency. A single model was trained and evaluated on large datasets from multiple organs, acquired with diverse imaging sequences and parameters, at both 1.5T and 3T. Our results demonstrate that the proposed approach achieves generalizable high-quality distortion correction while significantly reducing processing time compared to TOPUP, highlighting its potential for clinical translation.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/4840_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{QiuShi_Unsupervised_MICCAI2025,
        author = { Qiu, Shihan and Miron, Radu and Li, Yahang and Eichner, Cornelius and Feiweier, Thorsten and Janardhanan, Nirmal and Clifford, Bryan and Darwish, Omar and Mostapha, Mahmoud and Nadar, Mariappan S.},
        title = { { Unsupervised Learning-Based Susceptibility Artifact Correction for Diffusion-Weighted MRI in Multiple Organs } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15972},
        month = {September},

}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposed to train a CNN for estimating the displacement field of EPI-DWI scans using two polarities. The advantage is the speed of the processing compared to existing conventional method such as TOPUP.

  • 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 majors strengths are: (1) significantly faster than TOPUP. (2) Unsupervised training loss comparing the corrected images from the two polarities. (3) claimed to be generalisable to different acquisitions (e.g. b-values)

  • 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.

    Based on Fig. 1, the network requires paired EPI acquisitions of different polarities, rather than correcting on either individual polarities, which means it does not work on single EPI-DWI scans. The authors should think of a method that trains the network using individual DWI scan as the input, rather than required paired, which significantly impact its application.

  • 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 provide sufficient information for 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.

    (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 novelty of the method is low and its application is limited by requiring DWI image pairs.

  • 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 #2

  • Please describe the contribution of the paper

    The authors proposed an unsupervised DL model for TOPUP-based echo planar imaging (EPI) correction, enabling computationally efficient image correction. The proposed model adapted an encoder-decoder architecture with attention modules, and a loss function to minimise the unwarped EPI image pairs and enforce smoothness in the estimated displacement field. The model achieved better correction compared with TOPUP and other U-Net structures. The smoothness loss was assessed for high b-value DWI 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.
    1. This work applied self-attention module and demonstrated improved performance in EPI image correction.
    2. This work designed a smoothness loss in field map estimation and showed its benefits to reduce noise errors.
  • 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. Novelty is incremental. The unsupervised learning strategy followed by this work has been well-developed (Duong et al, 2020; Zahneisen et al, 2020; Alkilani et al, 2024). Field map smoothness loss is not considered significantly novel, as other similar approaches have also been proposed (Zahneisen et al, 2020).
    2. The motivation to investigate high b-value DWI images is vague. The authors claimed “[P2L15] In DWI application, they have focused on low b-value images, overlooking high b-value images which suffer from lower SNR …” In the discussion, the authors argued “[P7L7] In DWI, acquiring such pairs for every b-value and diffusion direction is time-consuming, … To improve efficiency…, we estimate the displacement field from one pair of reversed-PE low b-value DWI and apply it to other single-PE images.” These arguments are confusing and need further explanation. TOPUP corrects the susceptibility-induced image distortion in EPI acquisition. The distortion is independent to the b-value in DWI signal formation, thus common practices acquire the additional reversed-PE image and TOPUP only with b=0, and apply the same correction for all other b-values. It is worth noting that eddy current-induced distortion arises with high b-values, but its correction is beyond the scope of TOPUP. (Reference: https://www.youtube.com/watch?v=fXdhCdG4S_E&list=PLvgasosJnUVl_bt8VbERUyCLU93OG31h_&index=46)
    3. Referenced methods seemed to be a bit plain. Advanced models (e.g. Alkilani et al, 2024) are available in the literature for benchmarking.
  • 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.

  • 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

    Equation 10: In the context of DWI, the letter b is reserved by default. The parameter notation should be replaced to improve clarity.

  • 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?

    This work does not provide a strong foundation for understanding the research problem and presents limited novelty in methodology. Therefore, I recommend rejection for this paper.

  • Reviewer confidence

    Somewhat confident (2)

  • [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 proposed a deep learning based method to correct susceptibility artifact in DWI/EPI MRI.

  • 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 achieves correction performance comparable to the standard TOPUP while reducing inference time drastically, demonstrating clear clinical applicability.
    2. The spatially weighted smoothness loss adapts to signal strength, reducing noise propagation when applying displacement fields estimated from low b-value images to high b-value acquisitions.
  • 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. While distortion metrics and image similarity are reported, the study lacks validation on downstream applications such as tractography or quantitative diffusion modeling to further demonstrate clinical impact.

    2. Although efficient post-acquisition, the method still depends on acquiring two reversed-PE images. So eventually it may just just act as a faster TOPUP with risk of overfitting.

  • 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.

    (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 importance of its applications

  • 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 #4

  • Please describe the contribution of the paper

    The paper describes an unsupervised deep learning method for correcting susceptibility-induced distortions in EPI DWI across multiple organs. The authors propose a novel ‘Def-Convformer’ network architecture (using deformable convolutions within a Convformer block) to estimate the displacement field from reversed phase-encode (PE) image pairs. A key component is a spatially weighted smoothness loss function designed to improve robustness against noise, enabling the displacement field estimated from low b-value images to accurately correct high b-value 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 Def-Convformer architecture is well-suited for spatially varying distortions, and the spatially weighted smoothness loss intelligently handles noise transfer, particularly relevant for DWI. Evaluation on a large, diverse dataset strongly supports the method’s generalizability. The described method achieves distortion correction accuracy comparable to the established TOPUP method and outperforms U-Net based DL approaches, especially on challenging high-resolution scans. As this method offers significantly faster processing time, it makes it clinically viable. Lastly, the method successfully demonstrates the ability to estimate the field from higher-SNR low b-value images and apply it effectively to lower-SNR high b-value images which is a crucial capability for clinical DWI.

  • 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.

    There are certain weaknesses that the authors need to address. The study assumes negligible motion between reversed-PE scans and between low/high b-value scans, which may not always hold in practice. While achieving comparable results, TOPUP shows slightly better quantitative metrics in the presented comparison (Table 2), though the DL method is much faster.

  • 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 provide sufficient information for 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

    This is a strong contribution addressing a significant practical issue in DWI. Adding a brief discussion on potential future work incorporating motion correction would be beneficial.

  • 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 novel and effective deep learning solution for EPI distortion correction with significant practical advantages. The Def-Convformer architecture and the spatially weighted smoothness loss are well-motivated and demonstrate strong performance on a large, diverse dataset. The method achieves accuracy comparable to TOPUP while being drastically faster, and its proven ability to correct high b-value images using fields derived from low b-value images is a key strength for clinical DWI. Despite minor weaknesses regarding reproducibility and motion handling, the strengths and potential clinical impact warrant acceptance.

  • 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

Dependence on acquiring two reversed-PE images [R1, R2]: Our method estimates the displacement field from only one pair of reversed-PE low b-value DWI and applies it to correct the whole series of acquired (single-polarity) DWI images. There is no need to acquire pairs of DWI images for each desired diffusion encoding (i.e., b-value and diffusion direction). Therefore, the scan time won’t be increased a lot during application, with only one additional fast EPI scan. The acquisition of this initial reversed-PE pair provides more reliable information for the displacement field estimation in comparison to directly estimating the displacement from a single-polarity image. Risk of overfitting [R1]: We did observe that in some low-SNR cases TOPUP may over-correct the noises in the input images, and the estimated field causes a noise transfer issue when applied to other DWI images with different b-values (Fig. 3). Our method adds a spatially varying smoothness loss to mitigate such over-correction of noises in the input images, and it has been evaluated that the estimated field from low b-value images can be robustly applied to high b-value images to achieve good correction performance. The motivation to investigate high b-value DWI [R3]: As mentioned by Reviewer 3, the displacement field can be estimated from the b=0 images and applied to correct all other b-values. However, existing publications only evaluated their approaches on the input low b-value images and didn’t validate the correction performance of such application of estimated field from low b-value to other b-values. We included this evaluation in our study and enhanced our method by adding a spatially varying smoothness loss to mitigate the noise transfer issue. Similarity metrics in Table 2 and comparison with TOPUP [R4]: While higher PSNR/SSIM or lower NMSE in Table 2 indicate higher similarity between the corrected low b-value images, it should be noted that methods with high similarity metrics in Table 2, such as TOPUP and model 1, tend to suffer from noise transfer issue in application to other b-values or diffusion directions due to over-correction of noises in the inputs (as shown in Fig. 3). We specifically tuned our method to achieve more balanced performance than TOPUP, with comparable similarity metrics in correction of input low b-value images and much less noise transfer in applications to other b-values. Motion handling [R4]: A brief discussion of motion was included in the last paragraph of the Discussion section. More evaluation and investigation on motion handling will be a future direction of our research. Validation on downstream applications [R1]: This conference paper focuses more on technical development and initial validation of the DL-based susceptibility artifact correction approach. Further validation on downstream applications will be included in future extension into a journal paper.




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