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

Abstract

Neural implicit k-space representations have shown promising results for dynamic MRI at high temporal resolutions. Yet, their exclusive training in k-space limits the application of common image regularization methods to improve the final reconstruction. In this work, we introduce the concept of parallel imaging-inspired self-consistency (PISCO), which we incorporate as novel self-supervised k-space regularization enforcing a consistent neighborhood relationship. At no additional data cost, the proposed regularization significantly improves neural implicit k-space reconstructions on simulated data. Abdominal in-vivo reconstructions using PISCO result in enhanced spatio-temporal image quality compared to state-of-the-art methods. Code available at ***.git.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/3260_paper.pdf

SharedIt Link: https://rdcu.be/dV5EA

SpringerLink (DOI): https://doi.org/10.1007/978-3-031-72104-5_59

Supplementary Material: https://papers.miccai.org/miccai-2024/supp/3260_supp.zip

Link to the Code Repository

https://github.com/compai-lab/2024-miccai-spieker

Link to the Dataset(s)

N/A

BibTex

@InProceedings{Spi_SelfSupervised_MICCAI2024,
        author = { Spieker, Veronika and Eichhorn, Hannah and Stelter, Jonathan K. and Huang, Wenqi and Braren, Rickmer F. and Rueckert, Daniel and Sahli Costabal, Francisco and Hammernik, Kerstin and Prieto, Claudia and Karampinos, Dimitrios C. and Schnabel, Julia A.},
        title = { { Self-Supervised k-Space Regularization for Motion-Resolved Abdominal MRI Using Neural Implicit k-Space Representations } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
        year = {2024},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15007},
        month = {October},
        page = {614 -- 624}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposed a novel self-supervised k-space regularization enforcing a consistent neighborhood relationship for dynamic MRI reconstruction with neural implicit k-space representations (NIK). This method requires no additional calibration data and can improve the performance of NIK.

  • 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.

    1. The proposed method was easy to follow, and the benefits of self-supervised k-space regularization were well described.

    2. The paper provides experiments with simulated and MRI data, showing the potential to be used in MRI applications.

    3. The PISCO regularization seems novel as it provides an additional regularization enforcing the neighborhood relationship from a global view while still maintaining the implicit mapping from k-space coordinates to k-space data.

  • 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.

    1. The PISCO step in the pipeline uses the predicted results from the pre-trained NIK model, which means the performance of the regularization highly relies on the accuracy of the NIK model. This brings the concern of the reconstruction improvement when the NIK model is suboptimal.

    2. The PISCO method splits the data into several subsets and is designed on the assumption that each Ws optimized on the s-th subset should have the same solution. However, this seems improper as it is possible that Ws can overfit to the subset that this is optimized on. The authors did not include a detailed analysis of the results of the consistency-enforcing step. It will be convincing if the authors can provide some results, including how well will Ws be consistently optimized on different data subsets.

    3. The results lack sufficient evidence of qualitative results. For the static reconstruction, the improvement of PISCO-NIK seems marginal, and it is hard to see the improvement. Both NIK and PISCO-NIK failed to reconstruct sufficient spatial resolution (and apparently, the spatial resolution was different across the field of view). For the dynamic reconstruction, while the authors claim that PISCO-NIK provides smoother results with sharper vessel structures, PISCO-NIK blurs the image slightly and NIK can provide sharper and finer results.

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

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

    N/A

  • 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

    1. As written in the weaknesses section, the method will be more convincing if the authors provide some results, including how well Ws will be consistently optimized on different data subsets, which can directly prove the effectiveness of the self-consistency regularization.

    2. As the proposed method adds an optimization step to NIK, I am curious whether it will make the training time longer. Also, the paper does not include an ablation study like the effect of the local neighborhood kernel size, the size of the subset (it should affect the self-consistency optimization theoretically), etc.

  • 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?

    Results presented in the paper were not convincing enough to show the improved performance of the proposed method.

  • 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 #2

  • Please describe the contribution of the paper

    This work introduces a k-space-based regularization approach (PISCO) that operates independently of any additional training or calibration data

  • 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.
    1. The PISCO method is a novel self-supervised k-space regularization technique in this field. It is very flexible, being effective with or without calibration data.
    2. The method is validated both through simulated data and in-vivo abdominal MRI, providing robust evidence of its effectiveness.
  • 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.
    1. Figures 3 and 4 are too small, but I still sense that PISCO causes some extra blurring? Discussion is needed.
  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

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

    N/A

  • 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
    1. Enlarge the figures.
    2. More introduction on previous related work, e.g. other regularization techniques. Make comparisons if possible.
  • 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 is novel, effective and well preseted with sufficient evidence.

  • 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 #3

  • Please describe the contribution of the paper

    The paper describes a self-supervised calibration procedure for regularizing neural implicit models for MRI reconstruction. Demonstrated motion-resolved abdominal MRI at high quality based on the introduced method.

  • 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 written very clearly, the motivation for and technical details of the proposed method are explained well. The idea of using self-consistent estimation as inspired by GRAPPA/SPIRiT etc traditional k-space reconstructions in MRI for regularizing NIK 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.

    Comparisons against deep learning baselines beyond NIK are limited.

  • 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?

    N/A

  • 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

    The paper reads very well overall, but some further clarification might be possible:

    • There may already be another MRI reconstruction method abbreviated as PISCO by Haldar et al. I think this is fine, but I just wanted to let the authors know.

    • An argument is given in the Intro that it might be hard to adopt image-domain losses into k-space. I am not entirely sure about this argument, image and k-space domains could be visited back and forth via a Fourier transformation in order to express regularizers in any one domain or both. See for instance doi: 10.1109/TMI.2022.3220757 mixing k-space and image-domain losses.

    • NIK is related to DIP-style methods that are rely on model adaptation on test data, but maybe with the difference that DIP is commonly run in image domain. Some discussions might be useful.

    • Images are a bit dark in general (except zoom-in windows).

    • Some ablations might be useful.

  • 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

    Strong Accept — must be accepted due to excellence (6)

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

    Addressed an important imaging problem of broad interest to the community, well executed study design and clearly written paper with a sufficient amount of technical details.

  • 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



Author Feedback

We thank the reviewers for the positive feedback on our manuscript and their recognition of our work as novel (R1), effective (R3) and sound (R4). We are grateful for their recognition of the flexibility of the method and the “well executed study design” (R4) using simulations and in-vivo data.

Our manuscript introduces the novel concept of PISCO, which leverages self-consistency in k-space by solving multiple subsets for weight estimates. As pointed out by reviewers, the effectiveness of the method would be further supported by analysis of the actual weight estimates (R1) as well as ablations (R1/R4). We agree with these important points and analyzed the estimated weights throughout the learning process, observing convergence of the weights from multiple subsets. Also, as stated in the manuscript, we included Thikonov regularization and overdetermination of the weight subsets to enhance robustness against outliers and prevent overfitting. Yet, in this work, our main focus was to provide a thorough technical explanation to ensure a clear understanding of the method. Unfortunately, due to space constraints, these additional findings and ablations could not be included in the final manuscript but will be valuable inclusions for future extensions of this work.

With regards to the improved performance of PISCO (R1), we demonstrate significant quantitative improvements. In in-vivo data, we show noise reduction in our results, aligning with our goal to enhance NIK’s representation capability in lower-sampled outer k-space. We observe smoother vessel structures and anatomically more plausible dynamic motion using PISCO, consistent with the anatomical expectations set by XD-GRASP50 (high temporal resolution but strong undersampling artefacts). While this was the first introduction of this work, we expect further improvement when exploring further design choices.

Once again, we sincerely thank the reviewers for their insightful feedback and valuable suggestions. We will incorporate these comments, including points regarding formatting (R3/R4), into our final manuscript to the best of our abilities.



Meta-Review

Meta-review not available, early accepted paper.



back to top