Abstract

Modern diffusion MRI sequences commonly acquire a large number of volumes with diffusion sensitization gradients of differing strengths or directions. Such sequences rely on echo-planar imaging (EPI) to achieve reasonable scan duration. However, EPI is vulnerable to off-resonance effects, leading to tissue susceptibility and eddy-current induced distortions. The latter is particularly problematic because it causes misalignment between volumes, disrupting downstream modelling and analysis. The essential correction of eddy distortions is typically done post-acquisition, with image registration. However, this is non-trivial because correspondence between volumes can be severely disrupted due to volume-specific signal attenuations induced by varying directions and strengths of the applied gradients. This challenge has been successfully addressed by the popular FSL Eddy tool but at considerable computational cost. We propose an alternative approach, leveraging recent advances in image processing enabled by deep learning (DL). It consists of two convolutional neural networks: 1) An image translator to restore correspondence between images; 2) A registration model to align the translated images. Results demonstrate comparable distortion estimates to FSL Eddy, while requiring only modest training sample sizes. This work, to the best of our knowledge, is the first to tackle this problem with deep learning. Together with recently developed DL-based susceptibility correction techniques, they pave the way for real-time preprocessing of diffusion MRI, facilitating its wider uptake in the clinic.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/3783_paper.pdf

SharedIt Link: pending

SpringerLink (DOI): pending

Supplementary Material: https://papers.miccai.org/miccai-2024/supp/3783_supp.zip

Link to the Code Repository

github.com/CIG-UCL/eddeep

Link to the Dataset(s)

N/A

BibTex

@InProceedings{Leg_Eddeep_MICCAI2024,
        author = { Legouhy, Antoine and Callaghan, Ross and Stee, Whitney and Peigneux, Philippe and Azadbakht, Hojjat and Zhang, Hui},
        title = { { Eddeep: Fast eddy-current distortion correction for diffusion MRI with deep learning } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
        year = {2024},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15002},
        month = {October},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes the use of a pix2pix image translator and registration model to correct distortion induced by eddy currents in diffusion MRI.

  • Please list the main strengths of the paper; you should write about 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 generally well-written. The use of the quadratic distortion model (with 10 degrees of freedom) to model the geometric distortion is sound.

  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.

    The translated image looks worrisome to me as the structure details and contrast are wiped out. A lower MAE (in Fig.2) could simply mean that the translator networks better predict the average image when trained on more images or with extra data augmetation.

    Additionally, the translator was trained on undistorted images but is used to translate distorted images during inference. Although you claimed that augmentation techniques would bridge this train-inference gap, I did not see evidence to support that claim.

    The claim that “the first to tackle this problem with deep learning” is not accurate. There seems to be a similar DL registration-based for distortion correction in DWI, and the differences should be discussed: Bian, Zhangxing, Muhan Shao, Aaron Carass, and Jerry L. Prince. “DrDisco: Deep Registration for Distortion Correction of diffusion MRI with single phase-encoding.” In Medical Imaging 2023: Image Processing, vol. 12464, pp. 292-296. SPIE, 2023.

    There is no comparison with previous approach, such as synb0-disco.

  • 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.

  • Do you have any additional comments regarding the paper’s reproducibility?

    The code is claimed to be available, but the link leads to a non-existent repository.

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html

    Address my concerns in “weakness” section.

  • 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

    Weak Reject — could be rejected, dependent on rebuttal (3)

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

    see above.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    N/A

  • [Post rebuttal] Please justify your decision

    N/A



Review #2

  • Please describe the contribution of the paper

    Provides a deep-learning method for correcting diffusion MRI data with respect to eddy-current distortions and motion. The approach is split into two parts: one that makes contrasts similar across diffusion-weighting directions and b-values, and one that predicts correction parameters. The performance is as good as current the state-of-the-art approach (eddy from FSL).

  • Please list the main strengths of the paper; you should write about 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 very clearly written!

    The approach has the potential to accelerate a time-consuming pre-processing step in diffusion MRI.

  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.

    Computation times not reported.

    Potential weaknesses:

    Does the approach use for a flexible set of b-values and directions?

    How would the approach respond to image artifacts (e.g. dropouts, motion between slices et c)?

    Are there other edge cases where the approach is expected to produce inferior results?

  • Please rate the clarity and organization of this paper

    Excellent

  • 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.

  • Do you have any additional comments regarding the paper’s reproducibility?

    Data not available, but the code is.

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html

    I found the paper to be very clearly structured, however, the figures are generally hard to read.

    Figure 1 - spell out abbreviations.

    Figure 2. Images seem all to be b = 2000 images, and not b = 650, 1000, and 2000 as the axis labels say (CSF should not be fully suppressed at b 650).

    Figure 3. What are the different images showing? What is the take-home message from this figure?

    An alternative approach is to try to predict the exact contrast of the high b-value data, e.g. (Nilsson, et al. PloS one 10.11 (2015): e0141825). Wouldn’t this enable more accurate registrations?

  • 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

    Accept — should be accepted, independent of rebuttal (5)

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

    The paper is clearly written and successfully addresses a problem in the field.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    N/A

  • [Post rebuttal] Please justify your decision

    N/A



Review #3

  • Please describe the contribution of the paper

    This work proposes Eddeep, the first deep-learning-based method for correcting eddy current distortions in DWIs. Eddeep is composed of a translation network, for shifting a DWI to a reference gradient strength and (non-)direction, and a registration network, for estimating rigid and quadratic distortion terms. Eddeep is trained and evaluated on an in-house 2mm diffusion dataset. This model is shown to produce highly accurate distortion parameter estimates given only a handful of training subjects, even compared to the FSL-eddy tool itself.

  • Please list the main strengths of the paper; you should write about 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.
    • Eddeep is, to the best of my knowledge, the first deep-learning-based method for eddy current distortion correction in dMRI. While there exist many DL methods for the closely-related problem of susceptibility correction, this is the first to focus on eddy current correction. Despite being the first, Eddeep performs well in predicting eddy current and head motion parameters given relatively little training data.
    • The construction of Eddeep is carefully designed and meaningfully constrained. Rather than proposing an unconstrained model that predicts corrected images directly from DWIs, Eddeep wisely limits the manner in which the original data can be altered. For example, selecting a more limited cGAN pix2pix architecture, to limit spatial distortion from the translation network. Similarly, only predicting the quadratic and rigid transformation terms limits the model to only necessary transformations.
    • The paper is well organized and self-contained. The problem is succinctly defined, and the proposed method is justified, with small insights given throughout.
  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
    • The primary motivation for Eddeep, at its core, seems to be about saving time in post-processing. Unlike susceptibility correction, eddy current (and between-volume head motion) correction does not require special data acquisition parameters. Furthermore, this work specifically uses the FSL-eddy (the tool, noted as FSL-eddy to prevent confusion) output parameters as the ground truth, so it can never outperform FSL-eddy. So, the only gap for Eddeep to fill is that of lower inference time. This does not seem like the most compelling motivation, at least as laid out in the paper.
    • While the experiment results seem compelling (particularly Figure 3), it is not clear how good these performances are relative to any other method. While this is the first deep-learning eddy correction method, other non-deep-learning can be compared. At the same time, it is difficult to understand what a “poor” eddy current correction looks like on a resulting DWI. Put another way, it is unclear how the performance metric in Figure 3, high or low, translates into image features.
    • Given that the experiments are performed on one dataset using the same sequence between subjects, it is unclear how well Eddeep generalizes to out-of-distribution DWI sequences, scanners, DWI resolutions, or subjects. Is Eddeep capable of zero-shot generalization?
  • Please rate the clarity and organization of this paper

    Very 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.

  • Do you have any additional comments regarding the paper’s reproducibility?

    The only dataset used in this paper’s experiments was collected by the authors (or their collaborators) and is, I assume, not available to the public.

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
    • I believe the manuscript could be improved by providing a deeper motivation for Eddeep. What fundamental limitation of FSL-eddy motivates creating a new method? If the computational time of FSL-eddy is the only motivating factor, what is practically infeasible given this time cost?
    • The hand-drawn style of Figure 1, while somewhat aesthetically pleasing, makes details and variable notation difficult to understand.
    • There are several typos and unclear verbage throughout the manuscript. For example
      • page 2, “…but not yet for eddy-current ones because…”
      • page 8, “…eroded (iterations)…” (missing number of iterations?)
      • Figure 3 title, “mean std accross [sic] volumes”
  • 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

    Accept — should be accepted, independent of rebuttal (5)

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

    This work proposes a novel model for a sparsely-studied problem. The model is constructed in such a way that requires a deep understanding of the problem domain, and the experiment results, while somewhat limited, show a high-performing model. While I am not totally sold on the paper’s overall motivation, I wish more papers contained the insights and methodological care provided by these authors.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    N/A

  • [Post rebuttal] Please justify your decision

    N/A




Author Feedback

We thank the reviewers (R1, R3, R4) for their constructive critiques. R1 highlights the significance of the problem tackled “the potential to accelerate a time-consuming pre-processing step in diffusion MRI”. R4 recognises the novelty “the first deep-learning-based method for correcting eddy current distortions in DWI”, the thoughtful approach “carefully designed and meaningfully constrained […] that requires a deep understanding of the problem domain”, and performance “show a high-performing model”. All reviewers highlight the clarity of the submission, especially R1 “very clearly written”.

The only major concern was raised by R4 but this was a result of a misunderstanding. The “synb0-disco” and “DrDisco” methods R4 mentioned, as R3 pointed out, solve a different problem (correction of susceptibility-induced distortion). As R3 pointed out, this submission is the first deep-learning-based approach to correct eddy-current-induced distortion - “a problem sparsely studied”.

Minor points:

1) R1 suggested the inclusion of computation times. For the dataset of size M (n=32 for training and n=8 for validation), using a GPU NVIDIA GeForce RTX 4090 with a batch size of 4, it took about 4 hours to train the translator and 16 hours to train the registration model. Inference takes a split second.

2) R1 and R3 inquired about generalizability: In the present form, Eddeep is robust to some extent to brain shapes and orientations and to some unseen b-values thanks to the spatial and contrast augmentations preformed at training. It cannot however be considered capable of zero-shot generalization to any DW acquisitions. Future work may consider additional training data augmentation employing realistic simulated DW data to improve generalizability.

3) R1 suggested that Fig. 2 might have included images with b-values different from the values reported. After double-check, we can confirm that the indicated b-values are correct. For each displayed volume the intensities have been scaled by their maximal value. At b=650, the brightest voxels are now located in the cortex, making the CSF voxels appear dark.

4) Regarding R1 suggestion about the work from Nilsson et al. (2015): This interesting work is somewhat in the same vein as what is done in Tortoise, except high b-value volumes are predicted using CHARM model. Such synthetic data generation may indeed be promising future avenues to improve Eddeep generalizability.




Meta-Review

Meta-review not available, early accepted paper.



back to top