Abstract

Integrating dual-domain (i.e. frequency domain and spatial domain) information for magnetic resonance imaging (MRI) reconstruction from undersampled measurements greatly improves imaging efficiency. However, it is still a challenging task using the denoising diffusion probabilistic models (DDPM)-based method, due to the lack of an effective fusion module to integrate dual-domain information, and there is no work exploring the effect that comes from denoising diffusion strategy on dual-domain. In this study, we propose a novel center-to-edge DDPM (C2E-DDPM) for fully-sampled MRI reconstruction from undersampled measurements (i.e. undersampled k-space and undersampled MR image) by improving the learning ability in the frequency domain and cross-domain information attention. Different from previous work, C2E-DDPM provides a C2E denoising diffusion strategy for facilitating frequency domain learning and designs an attention-guided cross-domain junction for integrating dual-domain information. Experiments indicated that our proposed C2E-DDPM achieves state-of-the-art performances in the dataset fastMRI (i.e. The scores of PSNR/SSIM of 33.26/88.43 for 4x acceleration and 31.67/81.94 for 8x acceleration).

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/0748_paper.pdf

SharedIt Link: pending

SpringerLink (DOI): pending

Supplementary Material: N/A

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{Zha_CentertoEdge_MICCAI2024,
        author = { Zhao, Jianfeng and Li, Shuo},
        title = { { Center-to-Edge Denoising Diffusion Probabilistic Models with Cross-domain Attention for Undersampled MRI Reconstruction } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
        year = {2024},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15007},
        month = {October},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This work concentrates on effectively integrating dual-domain information and generative models for MRI reconstruction, which is a positive attempt for medical imaging community.

  • 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 paper gives a novel DDPM model that not relies on the noise adding and removing to process k-sapce data, and shows improved results. It is important to promote the fast MR imaging.

  • 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.

    It is not very clearly in section 2.1. Authors should further improved it.

  • 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?

    If possible, please provide the code to better implement the method.

  • 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) Some works have successfully combined the dual-domain information and diffusion model, such as [1]. So, authors should carefully check their literature review.

    (2) It is confusing in Eq. (6) that there is no undersampled MR image. Why the fully sampled MR image entirely involve the sampling process? Please explain this clearly.

    (3) This work combines the dual-domain information with the parallel way. Have the authors tried the alternative strategy like KIKI-Net, i.e., k-space data, image data, k-space, image data? Which one is better?

    (4) Why do authors only use the C2E strategy in the k-space domain? What about the image domain? Please explain it.

    (5) Ablation study shows that both C2E and Ag-Cd-J can boost the final results. However, it is not clear that the specific effects of different components on the reconstructions. Please charify it.

    (6) It is hard for readers to reproduce the proposed method. Please give more details.

    [1] Dual Domain Diffusion Guidance for 3D CBCT Metal Artifact Reduction

  • 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?

    It is an interesting way that use the C2E strategy to process MRI reconstruction. This enriches the application of generative models in medical imaging.

  • 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

    The paper introduces a ‘diffusion probabilistic model’ that reconstructs, separately, the undersampled k-space as well as in the image domain. The undersampled k-space or the ‘zero filled reconstruction image’ is used as condition for the diffusion-denoising process. With an ‘attention-guided cross domain junction’ the inverse Fourier transformed reconstructed k-space and reconstructed image are fused into a final output image. The method is trained and evaluated on the fast-MRI single-channel knee dataset and achieves the best PSNR and SSIM compared to relevant reference 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.

    Proper introduction of a novel method. High quality results on a relevant public dataset.

  • 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 work is presented for single channel images. Extension to multi-channel data may be non-trivial; it is not discussed.

  • 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?

    The main aspects of the method seem to be properly introduced and referenced, some details are missing. If the code is published they would be available.

  • 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

    Figure 1: layout would be improved by aliging the conditional images to the correct positions w.r.t. the K_t and S_t images. (instead of the dashed red arrows) Eq 8: what is the size of W_Q, W_K, W_V? as given it seems they are #components * #voxels, which would be very large to learn and/or low dimensional. If it is patch based, that should be described and different symbols should be used. Reference 19 is generic and does not specify a precise tokenization that would be directly applicable to the current paper. Related to this: what is ‘fully connected’ in the final layers? The structure and variable names suggest the entire images, but that would be 320^4 weights/channel, which is far larger than the training dataset.

    The conclusion states ‘The Ag-Cd-J module elimi- nates the error accumulation between dual domains by utilizing the multi-head attention layer to capture complementary information and establish coherent constraints between dual domains’ Although it might be learned implicitly, there is no explicit coherence of constraints established, nor is it validated that this is (strongly) learned. As such I consider this part not supported by the current work.
    The work is presented for single channel images. Extension to multi-channel data may be non-trivial (mainly due to resource constraints); and obviously is future work

  • 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 method is novel and presents a way to integrate cross domain information into the DDPM, which is an interesting line of research that is relevant for the MICCAI community. In my opinion the paper is presented well enough to be a good contribution to the discussions on this topic.

  • 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 authors propose a novel method to reconstruct undersampled MRI based on denoising diffusion probabilistic models. The novelty of the approach consists of the denoising process in frequency domain (C2E denoising) and of the cross-attention module improving the integration of spatial and frequency domain information. The proposed approach outperforms baseline and state-of-the art methods by an important margin.

  • 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.

    Both aspects of the proposed methods are novel and interesting:

    • C2E is proposed to replace global denoising diffusion strategy and considers the specificity of frequency domain. C2E denoising process is particularly impactful as it is not limited to MRI.
    • Ag-Cd-J module provides a new method to merge spatial and frequency domain for MRI improving existing approach MRI reconstruction
      In addition, ablation studies backed the benefits of each of the contributions convincingly.
  • 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 quantitative results on fastMRI do not match prior studies. For example, MC-DPPM scores reported in the paper are lower than in the original study (it is even worse than the minimum of PD and PDFS score). The author could provide the scores separated between PD and PDFS, or at least provide explanation for the difference. This strongly undermines the quantitative results and the claim that the proposed approach outperforms existing ones. Adding standard deviation would also strengthen the quantitative results.

  • 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
    • The figure explaining Ag-Cd-J module would benefit from being compared to concatenation approach (the same way C2E is compared with convention denoising). In general, the method figure is interesting but could be improved by being split into several sub figures as the innovation is relatively independent.
    • The Standard deviation on quantitative results should be added.
    • Future works could explore other denoising processes in ablation studies (based on the distribution of the energy in the spectrum, for example).
    • Fig 3 could be made denser to save space for improved method figure
  • 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 approach is novel and proposes important insights on denoising diffusion strategy for frequency domain that could apply in other application domains. The Ag-Cd-J module is also of interest for dual domain information processing, hence important for MRI application. The quantitative comparison with state-of-the-art approaches could be improved but does not question the main contribution of the paper, which is the novelty of the methodology that is backed by ablation studies.

  • 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

N/A




Meta-Review

Meta-review not available, early accepted paper.



back to top