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

Abstract

Recently, there have been significant advancements in the development of portable low-field (LF) magnetic resonance imaging (MRI) systems. These systems aim to provide low-cost, unshielded, and bedside diagnostic solutions. MRI experiences a diminished signal-to-noise ratio (SNR) at reduced field strengths, which results in severe signal deterioration and poor reconstruction. Therefore, reconstructing a high-field-equivalent image from a low-field MRI is a complex challenge due to the ill-posed nature of the task. In this paper, we introduce diffusion model driven neural representation. We decompose the low-field MRI enhancement problem into a data consistency subproblem and a prior subproblem and solve them in an iterative framework. The diffusion model provides high-quality high-field (HF) MR images prior, while the implicit neural representation ensures data consistency. Experimental results on simulated LF data and clinical LF data indicate that our proposed method capable of achieving zero-shot LF MRI enhancement, showing some potential for clinical applications.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/3647_paper.pdf

SharedIt Link: https://rdcu.be/dV5FI

SpringerLink (DOI): https://doi.org/10.1007/978-3-031-72104-5_74

Supplementary Material: https://papers.miccai.org/miccai-2024/supp/3647_supp.pdf

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{Lin_Zeroshot_MICCAI2024,
        author = { Lin, Xiyue and Du, Chenhe and Wu, Qing and Tian, Xuanyu and Yu, Jingyi and Zhang, Yuyao and Wei, Hongjiang},
        title = { { Zero-shot Low-field MRI Enhancement via Denoising Diffusion Driven Neural Representation } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
        year = {2024},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15007},
        month = {October},
        page = {775 -- 785}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The authors propose a method to recover a HF MRI from a LF MRI. Using a standard optimization approach, a regularized inverse problem is split into two sub-problems, one corresponding to data fidelity and another a prior de-noising problem. A diffusion-based algorithm is proposed that integrates the diffusion step with a data fidelity correction, the solution to the latter is parameterized by an INR and estimated incorporating prior regularization from the diffusion model. Experiments are included showing quantitative and qualitative improvements over relevant alternative methods.

  • 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 authors tackle an important and potentially high-impact problem. The method is well-justified and clearly formulated. The approach to incorporate a data-driven prior for INR regularization seems promising. The results are impressive compared to seemingly reasonable alternatives.

  • 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.

    From my understanding, the method requires re-fitting an INR T separate times for inference. Unless I missed it, there is no mention of the computational time required to run the method. INRs can take a long time to fit for large-scale 3D images, though their employment of the hash-grid set-up may remedy this issue to some extent. I found the dataset description unclear; specifically, what is the “simulated low field experiment data”?

  • 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 does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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 authors appear to describe only the real data and not the simulated dataset. The latter’s definition should be included or more clearly indicated where this information can be found.
    • Some discussion and/or evaluation of the computational time of the method should be included.
    • Does the data need to be registered to the same space that the prior diffusion model was trained in? If so, how was this done/is it a significant limitation?
  • 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 Accept — could be accepted, dependent on rebuttal (4)

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

    Well formulated approach to a high impact problem and good results. This reviewer found the manner of integrating prior information into the INR estimation promising. Some additional details should be included to clarify a few remaining points (see above).

  • Reviewer confidence

    Somewhat confident (2)

  • [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

    The paper presents an adaption of diffusion models as a generative prior to the task of translating from low-field MRI to high-field MRI. They showcase their method on an in-house dataset and outperform a variety of competing methods. They showcase that the resulting reconstruction has similar properties to the high-field MRI using automatic segmentation methods.

  • 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.

    It is an interesting method for the MICCAI community with wide-reaching potential if low-field MRI becomes comparable to high-field MRI. I think this method can be applied to other non-paired reconstruction problems as well. The results of the method are superior to a wide range of other methods. In principle, the paper is well written and the key idea used here is easy to extract from the paper.

  • 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 paper is impossible to reproduce.

    Furthermore, no significance testing was performed on the results, and the visual results are not very convincing, the areas are just a lot smoother. If this improves clinical usability, has yet to be determined.

    While no quantitative values on the segmentation results are given it is claimed that their method outperforms other methods, due to one slice in an entire brain scan.

  • 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 does not provide sufficient information for reproducibility.

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

    The optimization of the MLP is never specified. The diffusion function s_theta is just specified as pre-trained on 3T images. The point spread function is not specified. No hyperparameter are specified at all.

  • 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

    While the general idea of the paper is easy to follow, in detail it can be hard to grasp. E.g. the MLP uses a hash function which is never described in text, only in supplementary material.

    Having access to the point spread function and inference code would significantly improve the usefullness of the paper. Is this possible? If not adding more thorough descriptions to the paper is necessary.

    For the quantitative values, adding significance testing would strengthen the results. Adding any quantitative metric to the segmentation results would strengthen this section as well as provide some backing to the claim that your method outperforms the competitors.

    Why was FLAIR used in the LF imaging and not T1w?

    In Eq3: it should be x instead of x_HF

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

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

    I think the method is very well suited for this task and I think it is beneficial to have it published and discussed, however there are glaring issues with reproducibility and the evaluation could be improved.

  • 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

    This article presents the diffusion-based approach, DiffDeuR, aiming to enhance the portable low-field MR image as it were acquired by a high-field MRI system. DiffDeuR creatively split the original image enhancement task into the data-prior and the data-fidelity sub-problems, respectively employing the score-based diffusion model and implicit neural representation. Validation was done through both synthetic and clinical data, confirming the method’s effectiveness.

  • 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.

    This approach effectively incorporates the score-based denoising diffusion model with INR, constructing a novel application for MR image enhancement via a subroutine in INR that maps the spatial coordinates of an MR image into the intensity at the corresponding image pixel.

  • 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. In the method section, the overall application process has not been clearly explained. What is the advantage and why are the diffusion model and the INR framework combined?
    2. INR maps the spatial coordinates of an MR image into the intensity at the corresponding image pixel. However, the motivation of using the new map may be less mentioned in the main text.
    3. Details on the used real LF and HF data for model training and evaluation are somewhat oblique, including quantity, data dimension, and data-preprocessing. Besides, that LF MRI uses “2D FLAIR protocol to obtain T1-weighted scans” in Section 3.1 does not make sense.
    4. The degradation model for HF MR image used in Section 2.1 may be unrealistic, as it is required to align the contrast for brain tissues to LF MR, especially for the ultra-low-field; see Marques et al (2019) and Wu et al (2016). Also, in Section 3.1, the use of FLAIR LF MRI to generate T1-weighted HF MRI may require the adaption of the brain tissue contrasts.
    5. The derivation of diffusion model in Section 2.2 seems not self-contained. Suggest perform sanity check once again.

    Reference: Marques, José P., Frank FJ Simonis, and Andrew G. Webb. “Low‐field MRI: An MR physics perspective.” Journal of magnetic resonance imaging 49.6 (2019): 1528-1542. Wu, Ziyue, Weiyi Chen, and Krishna S. Nayak. “Minimum field strength simulator for proton density weighted MRI.” PloS one 11.5 (2016): e0154711.

  • 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.

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

    The article successfully combines the strengths of diffusion models and INR methods to address the LF MRI enhancement problem, achieving promising results on both simulation and clinical data. However, the writing in the methodology section lacks clarity, making it difficult to understand the algorithmic structure in a straightforward manner. Additionally, there is room for further improvement in the experimental design.

  • 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. The schematic diagram in Figure 2 is not very clear.
    2. The detailed information about how to use this method is still unclear, it’s also better to use a diagram to explanation.
    3. s_θ* in Equation 5 is undefined.
    4. Algorithm1 pseudocode: The defined MLP network is F_Φ, and when used, it is M_Φ; L_DF at line 7 is undefined;
    5. How to obtain synthetic data? Is it a public data set?
    6. The amount of simulation and experimental data is unknown.
    7. The third paragraph of section 3.1 “we utilize two standard metrics”maybe three metrics?
    8. Section 3.2 mentioned “To sum up, our DiffDeuR achieves the best performance in most cases”, but from the perspective of quantitative Table1, they are all the best. “In most cases” here shows that the effect is poor in some examples? Does it indicate that the algorithm has some limitations?
    9. Why does Figure 4 only compare 2 baseline models?
  • 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 Accept — could be accepted, dependent on rebuttal (4)

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

    The clarity of presentation requires a significant improvement.

  • 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

#1:

  1. Time: Regarding the runtime, we use only a small time T in sampling to speed up the enhancement process. Specifically, we set T = 20, and with this setting, it takes about 2 minutes to enhance an LF image on a single NVIDIA RTX 3060 GPU. We will add a description of the runtime in the paper.
  2. Simulation Data: The simulation experiment involves applying our proposed degradation model (Eq. 2) to the acquired HF reference images, resulting in simulated LF data. By testing the proposed method on the simulated LF data, we can qualitatively and quantitatively assess the reliability of our work. We have provided a more detailed description of the dataset in the paper.
  3. Data registration: Data does not need to be registered to the same space. #3:
  4. Result: The proposed DiffDeuR provides remarkably clear contrast in the deep brain nuclei, as shown in Fig. 3. In contrast, the methods used for comparison, such as IREM and DPS, are limited by a smoothing effect.
  5. Parameter: We have supplemented the paper with detailed settings for each of the hyperparameters as reviewer mentioned, including the MLP optimization, the PSF function, and other hyperparameters.
  6. Hash function: We use hash encoding for encoding the coordinate grid to enhance the MLP’s ability to fit high-frequency information, details of which can be found in InstantNGP [2]. We have also added a brief description of hash encoding in the paper.
  7. Quantitative values on segmentation result: Thanks to your valuable comments, we will add quantitative metrics comparison on the segmentation results to illustrate the superiority of our method more comprehensively. #4:
  8. Method description: The details of the overall application process of the method can be found in Algorithm 1, and we have further revised the description of this section for clarification. . Regarding the combination of the diffusion model and the INR framework, our approach benefits from the strong data prior provided by the diffusion model and the physical model knowledge and flexibility provided by the INR framework.
  9. Dataset detail: We will add more detailed description for model training and evaluation dataset.
  10. Degradation model: The degradation model we used to approximate the transition from HF to LF MRI is based on the model proposed by Christopher et al. [3]. We agree with the reviewer that this model lacks considerations for tissue contrast, which may limit our work in aligning image contrast between LF and HF MRI to some extent. We will consider the proposed approach in our follow-up work to further optimize the degradation model used, we thank the reviewer for this constructive suggestion.
  11. Derivation of diffusion model: We performed a detailed revision and fixed possible derivation problems.
  12. Formula normality: Thank you for pointing out the normative problems with the formulas in the paper. We have double-checked and revised the issue.
  13. Comparison with SOTA: In fact, “in most cases” is a mischaracterization, our approach outperforms the current state-of-the-art (SOTA) in all evaluations. Regarding Fig. 4, we compare our method with two current SOTA methods, and the results clearly demonstrate the superiority performance of the proposed method. [1] Lee, Suhyeon, et al. “Improving 3D imaging with pre-trained perpendicular 2D diffusion models.” Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023. [2] Müller, Thomas, et al. “Instant neural graphics primitives with a multiresolution hash encoding.” ACM transactions on graphics (TOG) 41.4 (2022): 1-15. [3] Man, Christopher, et al. “Deep learning enabled fast 3D brain MRI at 0.055 tesla.” Science Advances 9.38 (2023): eadi9327.



Meta-Review

Meta-review not available, early accepted paper.



back to top