List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Yutong Xie, Quanzheng Li
Abstract
We propose a novel and unified method, measurement-conditioned denoising diffusion probabilistic model (MC-DDPM), for under-sampled medical image reconstruction based on DDPM. Different from previous works, MC-DDPM is defined in measurement domain (e.g. k-space in MRI reconstruction) and conditioned on under-sampling mask. We apply this method to accelerate MRI reconstruction and the experimental results show excellent performance, outperforming full supervision baseline and the state-of-the-art score-based reconstruction method. Due to its generative nature, MC-DDPM can also quantify the uncertainty of reconstruction. Our code is available on github.
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16446-0_62
SharedIt: https://rdcu.be/cVRT6
Link to the code repository
https://github.com/Theodore-PKU/MC-DDPM
Link to the dataset(s)
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a denoising diffusion probabilistic model based approach for under-sampled medical image reconstruction. A measurement-conditioned DDPM method is proposed in measurement domain. Promising results on MRI reconstruction are demonstrated in the experiments.
- 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 is well organized. The presentation is clear. The diffusion and sampling process are defined in the measurement domain and conditioned on under-sampling mask, which makes the proposed method different from the existing DDPM methods. Experimental results show that the proposed method is superior than compared baselines.
- 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.
This paper restricts the under-sample matrix M to be a diagonal matrix with element to be either 1 or 0. It limits the impact of the proposed method. It is worthy to study how to apply the proposed method to other under-sample patterns, such as random or spiral. The assumption that the noise in equation (5) is set as zero needs supporting evidence/justification. Comparison with DDPM on the image domain is expected to justify the superiority to do DDPM in the measurement domain.
- Please rate the clarity and organization of this paper
Very Good
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
It is expected the results can be reproduced with the given information.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2022/en/REVIEWER-GUIDELINES.html
See weaknesses.
typos: Section 2.2: reconstruct x from y as (accurate as) possible Therefore, the task of under-sampled medical image reconstruction (is) to reconstruct the posterior distribution.
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
5
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The proposed measurement conditioned DDPM for accelerated MRI reconstruction is novel. While there are some limitations and weaknesses of the proposed method, the proposed method demonstrates promising results on the test dataset.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
2
- Reviewer confidence
Somewhat Confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
Not Answered
- [Post rebuttal] Please justify your decision
Not Answered
Review #3
- Please describe the contribution of the paper
This paper presents a novel and unified mathematical framework-MCDDPM, for medical image reconstruction using under-sampled reconstruction. It is very meaningful.
- 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.
As one of the highlights, the diffusion and sampling process are defined in measurement domain rather than image domain.
- 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.
In the discussion part, the weakness of the proposed method is somewhat less.
- Please rate the clarity and organization of this paper
Excellent
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
very good.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2022/en/REVIEWER-GUIDELINES.html
- The order of references in the text is right? Does it match the template? please check it carefully.
- In page 2, paragraph 2: In this paper, We design our method. Maybe “we” is correct.
- In the discussion part, the weakness of the proposed method is somewhat less. Please give more discussion on the proposed method, if possible. This manuscript is well constructed, so I recommend it for publication.
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
6
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This manuscript is well constructed and the method are well verified by adequate results.
- Number of papers in your stack
3
- What is the ranking of this paper in your review stack?
2
- Reviewer confidence
Confident but not absolutely certain
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
5
- [Post rebuttal] Please justify your decision
I find other weakness of this manuscript by reading other reviewers’ comments, which makes me even more clear on this paper. Authors have post their rebuttal with respect to all reviewers’ concern, I guess the next version will be improved somewhat. But one more thing, I am interested in it most, I thought they will say something about the improvement of discussion, but they didn’t. Actually, discussion is vital to their paper. So, I change my overall opinion on this paper, as I am not sure they will revise the discussion part.
Review #4
- Please describe the contribution of the paper
-This paper applies DDPM to undersampled MRI reconstruction. -The condition is on under-sampling, which inherent the data consistency in the recon pipeline. -The proposed method allows uncertainty quantification.
- 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 application of DDPM to accelerated MRI reconstruction is novel. -The condition is on the measurement.
- 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 not clear. Apart from the typos, the paper introduces CT, rather than MRI in section 2.2. I understand the authors want to introduce a general framework, but it makes me feel confused. -The authors only use magnitude image for DDPM, however, MRI is complex-valued in nature, and the phase information has clinical significance. How to apply to complex-valued MRI? -The comparison with U-Net is not sufficient. More comparison with state-of-the-art models, like unrolled networks, is needed.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
The dataset is publicly available fastMRI single-coil knee image. The detail configuration is described. However, it is not clear how to split the training/validation/testing dataset. Is this the same with the challenge?
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2022/en/REVIEWER-GUIDELINES.html
-DDPM can be very slow, which may limits its application. The authors did not discussion/mention this point. Accelerate the DDPM needs further investigation. -More benefits of using DDPM, rather than current state-of-the-art unrolled network should be mentioned. I did not see any comparison in the experiments.
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
5
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The application of DDPM to undersampled MRI reconstruction is the novel part. However, the paper lacks the comparison with the current methods. It also not clear if DDPM is practically available. e.g. the discussion on the computational time is missing.
- Number of papers in your stack
2
- What is the ranking of this paper in your review stack?
1
- Reviewer confidence
Very confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
5
- [Post rebuttal] Please justify your decision
Overall, the authors address most of the questions. My concerns on the clarity of the paper, discussions on the DDPM practical usage have been explained, and i suggested to include/revise the in the final version.
However, my concerns on the experimental comparison with other unrolled methods are not addressed. Nevertheless, this paper has its merits for presenting in the conference.
Primary Meta-Review
- Please provide your assessment of this work, taking into account all reviews. Summarize the key strengths and weaknesses of the paper and justify your recommendation. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
This research provides a probabilistic model-based denoising diffusion strategy for under-sampled medical image reconstruction. In the measurement domain, a measurement-conditioned DDPM approach is suggested. The tests show promising results for MRI reconstruction. In this research, DDPM is used to rebuild an undersampled MRI. The condition relies on under-sampling, which is inherent in the recon pipeline’s data consistency. The suggested technique provides for the quantification of uncertainty.
First, the work is indeed an interesting topic. Experimental results could provide some insights for this field. Three reviewers have positive review comments: The diffusion and sampling process are defined in the measurement domain and conditioned on under-sampling mask, which makes the proposed method different from the existing DDPM methods (R1). The diffusion and sampling process are defined in measurement domain rather than image domain (R2). The application of DDPM to accelerated MRI reconstruction is novel (R3).
However, reviewers also have negative comments: The quantitative assessment is input dependent. Comparison with DDPM on the image domain is expected to justify the superiority to do DDPM in the measurement domain (R1). The authors only use magnitude image for DDPM (R3).
Reviewers are confident about their concerns. In the rebuttal, the authors may need to highlight:
- Dataset details.
- Some of the experiment details are not clearly shown.
- Reproducibility of the current work.
- What is the ranking of this paper in your stack? Use a number between 1 (best paper in your stack) and n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
7
Author Feedback
We thank all reviewers for their constructive comments. Q:The comparison with other methods, including DDPM on the image domain and current sota unrolled networks (R1, R3, M-R) A:There is no existing work that applied DDPM to MRI reconstruction on the image domain. However, there are some works that have applied score-based generative models to MRI reconstruction on the image domain [11,20,3],. Since score-based generative models are similar to DDPM [21], we compare our result with [3] which used the same dataset. The superiority of our method is shown in the 5th column of Table 1 in Sect 4.2. More details are provided in the 1st paragraph of Sect 4.2. Following the reviewers’ suggestion, we also compare with unrolled networks and other reconstruction methods, e.g PD-net [5], Cascaded-net [6] and KIKI-net [7], and test them on the same volumes as MC-DDPM. The result for 4x acceleration in the following table shows the improvement of our method PSNR/SSIM PD PDFS PD-net 33.09/0.832 27.74/0.586 Cascade-net 32.99/0.824 27.61/0.579 KIKI-net 31.61/0.808 27.51/0.590 MC-DDPM 36.69/0.905 33.00/0.735 More experiments will be conducted to further justify the superiority of our proposed method in the final manuscript. In addition, we want to emphasize the advantage of DDPM compared with unrolled networks: 1) reconstruction results with more details; 2) availability of uncertainty estimation. Q:Details for the dataset (R3, M-R) A:The training dataset and the validation dataset are the same as the fastMRI challenge. More details have been provided in Sect. 2 in Supplementary Materials. Q:Complex-value image, details for experiments, and reproducibility of the current work (R3, M-R) A:In our experiments, complex-value images are used both in training and testing. The real and imaginary parts are represented by 2-channel tensors. The magnitude images are computed from reconstructed complex-value results, which are used only for evaluation. We only considered the Cartesian sampling pattern and the under-sampled masks are generated based on the code provided by [23]. We have provided more details in Sect. 2 in Supplementary Materials and our code will be released later. Q:Can MC-DDPM be applied to other sampling patterns such as random and spiral sampling? (R1) A:In our work, MC-DDPM can be directly applied to any grid-based sampling pattern, because the diagonal elements with value 1 in M are corresponding to the sampled positions. However, for non-grid-based sampling patterns, Eq. 5 in Sect. 3 does not hold. Therefore, MC-DDPM is not applicable. One can easily define the diffusion process according to the specific non-grid-based sampling method and transform the form of measurement into the form of a grid so that it can be fed into our method. Q:The reason for the assumption that noise in Eq.5 is zero (R1) A:We set the noise to be zero for simplicity. When the noise is not zero, the theory, training, and inference will be more complicated but could be extended from the current method. In the near future, we aim to apply this method to cardiac MRI imaging and will solve the zero noise case at that time. Q:The weakness of DDPM (R2, R3) A:As R3 has pointed out, the weakness is the relatively slow inference. It takes 10 seconds to generate one slice with 250 sampling steps on RTX 3090Ti. This limits the practical utilization of our method in some clinical settings. Q:The purpose of the paper (R3) A:MC-DDPM was proposed to solve the general under-sampled inverse problem which follows the form of Eq. 5 in Sect. 3. However, we introduced CT in Sect 2.2 as an example of Eq. 5, but conducted experiments on MRI, thus causing confusion. We are sorry and will revise it in the final version.
[25] Model Learning: Primal Dual Networks for Fast MR imaging. [26] A Deep Cascade of Convolutional Neural Networks for MR Image Reconstruction [27] Cross-domain convolutional neural networks for reconstruction undersampled mri
Post-rebuttal Meta-Reviews
Meta-review # 1 (Primary)
- Please provide your assessment of the paper taking all information into account, including rebuttal. Highlight the key strengths and weaknesses of the paper, clarify how you reconciled contrasting review comments and scores, indicate if concerns were successfully addressed in the rebuttal, and provide a clear justification of your decision. If you disagree with some of the (meta)reviewer statements, you can indicate so in your meta-review. Please make sure that the authors, program chairs, and the public can understand the reason for your decision.
This study develops a probabilistic model-based denoising diffusion technique for undersampled MRI reconstruction. A measurement-conditioned DDPM technique is proposed in the measurement domain. The findings of the testing are encouraging for MRI reconstruction. In this study, DDPM is employed to reconstruct an undersampled MRI. The problem is caused by under-sampling, which is inherent in the data consistency of the recon process. The suggested technique provides for the quantification of uncertainty.
The work is well written and shows interesting results. After rebuttal, all reviewers gave weak acceptance to the work. The authors promised to release codes as well. I am inclined to accept the work, but with two minor comments:
- The reproducibility is not only about open source codes, but also about multicenter/scanner studies. Data harmonization can be a solution: https://doi.org/10.1016/j.inffus.2022.01.001
- The work has an overlength Supplementary File (should be max. 2 pages) with method descriptions etc., which is a violation of the submission rules. I will leave this with the chair to decide.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
3
Meta-review #2
- Please provide your assessment of the paper taking all information into account, including rebuttal. Highlight the key strengths and weaknesses of the paper, clarify how you reconciled contrasting review comments and scores, indicate if concerns were successfully addressed in the rebuttal, and provide a clear justification of your decision. If you disagree with some of the (meta)reviewer statements, you can indicate so in your meta-review. Please make sure that the authors, program chairs, and the public can understand the reason for your decision.
Reviewers are mostly satisfied by the authors rebuttal, especially conerning the clarity in motivating the DDPM method, and the modeling aspects. There are still some concerns about the experimental details, and the discussion, which should be addressed in the manuscript. Nevertheless, the contribution of this work was found of enough interest for acceptance to the conference.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
8
Meta-review #3
- Please provide your assessment of the paper taking all information into account, including rebuttal. Highlight the key strengths and weaknesses of the paper, clarify how you reconciled contrasting review comments and scores, indicate if concerns were successfully addressed in the rebuttal, and provide a clear justification of your decision. If you disagree with some of the (meta)reviewer statements, you can indicate so in your meta-review. Please make sure that the authors, program chairs, and the public can understand the reason for your decision.
All reviewers recommended (weak) accept after evaluating the authors’ rebuttal letter. With the authors’ commitment to incorporate reviewers’ feedback in their revised manuscript, I recommend accept for this paper without further questions.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
5