List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Jiahao Huang, Angelica I. Aviles-Rivero, Carola-Bibiane Schönlieb, Guang Yang
Abstract
Deep learning has shown the capability to substantially accelerate MRI reconstruction while acquiring fewer measurements. Recently, diffusion models have gained burgeoning interests as a novel type of deep learning-based generative methods. These methods aim to sample data points that belong to a target distribution from a Gaussian distribution, which has been successfully extended to MRI reconstruction. In this work, we proposed a Cold Diffusion-based MRI reconstruction method called CDiffMR. Different from conventional diffusion models, the degradation operation of our CDiffMR is based on k-space undersampling instead of adding Gaussian noise, and the restoration network is trained to harness a de-aliaseing function.We also design starting point and data consistency conditioning strategies to guide and accelerate the reverse process. More intriguingly, the pre-trained CDiffMR model can be reused for reconstruction tasks with different undersampling rates. We demonstrated, through extensive numerical and visual experiments, that the proposed CDiffMR can achieve comparable or even superior reconstruction results than state-of-the-art models. Compared to the diffusion model-based counterpart, CDiffMR reaches readily competing results using only 1.6 ∼ 3.4% for inference time. The code is publicly available at https://github.com/ayanglab/CDiffMR.
Link to paper
DOI: https://doi.org/10.1007/978-3-031-43999-5_1
SharedIt: https://rdcu.be/dnwjb
Link to the code repository
https://github.com/ayanglab/CDiffMR
Link to the dataset(s)
N/A
Reviews
Review #1
- Please describe the contribution of the paper
This paper talks about the MRI reconstruction from undersampled acquisitions. A diffusion-based model, named Cold Diffusion MRI Reconstruction (CDiffMR), is proposed that 1) replaces the standard Gaussian noise degradation operator with a K-space undersampling degradation (KSUD) module and 2) applied two k-space conditioning strategies to guide the reverse process. The goal is to improve the reconstruction of undersampled MR images with better accuracy and time efficiency.
- 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 KSUD module enables CDiffMR to learn an explicit relationship between the input undersampled images and the target fully-sampled images.
- The starting point conditioning (SPC) and data consistency conditioning (DCC) strategies allow the trained CDiffMR to be reused for reconstruction tasks within a reasonable undersampling rate.
- The comparison experiment shows the improved performance of reconstruction accuracy and reduced inference time. The ablation study proves the benefit of the two conditioning strategies and KSUD.
- 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 description of the contribution model components could be better.
- The author could provide more details on the implementation of CDiffMR.
- The ablation study on KSUD could be more elaborated.
- 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
The authors have plans to make the code and dataset public, which would benefit the community in recognizing this problem. The evaluation metric is clearly described so that the result would be convincing.
- 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/2023/en/REVIEWER-GUIDELINES.html
- The contribution of reusing the pre-trained CDiffMR model with any undersampling rates in Section 1 is confusing. In fact, the undersampling rate should be within the range of the one used in pre-training the model according to the ablation study in Section 4. Could the author elaborate on that?
- An overview figure describing where the KSUD is applied would be helpful to understand the flow of the CDiffMR model.
- There should be an ablation study on CDiffMR using standard degradation with Gaussian noise.
- A visualization of ablation studies, especially for Fig. 4B, would be helpful to understand the improvement of SPC and DCC.
- 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 work uses a diffusion-based model with KSUD and two conditioning strategies in the restoration process to improve the reconstruction of MR images. Sufficient experiments have been done to support the benefit.
- 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
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #3
- Please describe the contribution of the paper
The main contribution of this work is the introduction of a Cold Diffusion-based MRI (CDiffMR) reconstruction method, which uses k -space under-sampling instead of adding Gaussian noise in the degradation step for diffusion model-based image reconstruction. Compared to the conventional diffusion model-based counterpart, CDiffMR could largely improve the inference speed.
- 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 main strength of the paper is the introduction of a Cold Diffusion-based MRI (CDiffMR) reconstruction method. It uses k -space under-sampling instead of adding Gaussian noise in the degradation step for diffusion model-based image reconstruction. Moreover, like any other diffusion-based approaches, the pre-trained CDiffMR model is very general and can be reused for reconstruction tasks with different undersampling rates.
- 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.
No
- 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
The authors have chosen to share the code, pretrained models and data upon acceptance.
- 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/2023/en/REVIEWER-GUIDELINES.html
The work introduces a Cold Diffusion-based MRI (CDiffMR) reconstruction method, which uses k -space under-sampling instead of adding Gaussian noise in the degradation step for diffusion model-based image reconstruction. The authors also compares Linear and Log k -space undersampling strategies in the degradation step using the FastMRI dataset. The main benefit of such a CDiffMR method is that the reconstruction process can be largely accelerated: Compared to the diffusion model-based counterpart, CDiffMR reaches readily competing results using only 1.6 ∼ 3.4% for inference time.
This work is very interesting. The manuscript is very well-written and the evaluations are well-designed. I have no further comments.
- 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
7
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Apart from the above strengthens, the manuscript is very well-written and the evaluations are well-designed.
- Reviewer confidence
Very confident
- [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 #4
- Please describe the contribution of the paper
The presented manuscript investigates a novel deep learning-based method called CDiffMR for accelerating MRI reconstruction using fewer measurements. The authors utilize a special kind of cold diffusion model, which differ from conventional diffusion models by using k-space undersampling instead of Gaussian noise. The method incorporates de-aliasing functions and unique conditioning strategies to guide and speed up the reverse process. The authors further use the pre-trained model for different undersampling rates, showing that CDiffMR achieves comparable or superior results to state-of-the-art models while significantly reducing inference time.
- 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 idea of the novel K-Space Undersampling Degradation (KSUD) is neat and valuable in principle. This approach may be particularly advantageous for reconstructing undersampled MRI data, especially in situations with limited data like dynamic MRI (even though not investigated in this work).
- 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.
Invariance to Forward Operator: One of the main advantages of having a generative model is that in principle it should be invariant to the forward operator. The authors never show that in any of the experiments a different forward operator, respectively subsampling scheme (Spiral, Radial) than that in training was used. This is a major drawback of this manuscript and should be considered for improving this work.
Benchmark: Although the authors use data from a public benchmark, a submission to the testing system of the fastmri challenge was omitted. This would have also been interesting regarding the performance for lower subsampling rates, what leads me to my next point.
Usability: This work is only evaluated at extremely high subsampling rates that are not commonly used in clinical practice (at least 16X is rather high). Why did the authors choose 8X and 16X subsampling?
Generative abilities: Furthermore, as this work builds upon a generative framework, it should be able to generate samples or at least show a certain diversity. Why has this not been shown or mentioned?
Unfortunately, in a nutshell the authors failed to show the benefits of a generative prior.
- Please rate the clarity and organization of this paper
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
The requirements resulting from the answers to the checklist are met in the manuscript.
- 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/2023/en/REVIEWER-GUIDELINES.html
For improving this work, I would recommend: Testing the scheme on different sampling trajectories in order to show that the proposed scheme really holds what the authors claim. As well as submitting the results to the fastmri benchmark in order to see practical performance and a fair comparison.
- 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
4
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The main factor that lead to my decision are the lack of proper evaluation in terms of invariance to forward operator as well as the the lack of evaluation on the benchmark and omitting the benefits of a generative prior.
- Reviewer confidence
Very confident
- [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
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 paper presents a Diffusion model for MRI Reconstruction from undersampled data. Main technical novelties include: 1) replacing the conventional Gaussian noise degradation with an undersampling degradation and 2) introducing k-space conditioning strategies to guide the reverse process. Results suggest improvements in reconstruction accuracy and time. Two reviewers were very positive with minor comments related to some missing details in the ablation studies. The third reviewer raised major concerns regarding validation and the undersampling factors that were used. In addition, authors should demonstrate clinical impact using the clinical annotations provided by the fast-mri+ dataset and compare with e2e varnet which provides the SOTA result on this dataset. Although the overall rank of the paper is high, the concerns of the third reviewer are important and I therefore expect the authors to address them in the rebuttal.
Author Feedback
Our proposed Cold Diffusion Model (CDiffMR) explicitly learns k-space undersampling, instead of using Gaussian noise for MRI reconstruction, with an improved reconstruction quality and reduced time (R1,R3,R4). CDiffMR is general and can be reused for reconstruction tasks across a reasonable range of undersampling rates (R1,R3). We greatly appreciate all the suggestions from reviewers.
Experiments on Lower Undersampling Rate (R4) CDiffMR shows its advantage at a higher undersampling rate since it explicitly learns the k-space undersampling process. X4 results demonstrate a similar trend to X8 results but were not included in the main paper due to page limitations. As requested, we now provide results with a X4 (Method: SSIM↑, LPIPS↓): D5C5: 0.804, 0.167; DAGAN: 0.791, 0.212; SwinMR: 0.810, 0.160; DiffuseRecon: 0.813, 0.180; CDiffMR-LogSR: 0.812, 0.157; CDiffMR-LinSR: 0.814, 0.160.
Experiments on Undersampling Schemes (R4) Our paper aims to maintain consistency of the degradation operator between the diffusion model and the inverse problem, and innovatively incorporate the k-space undersampling as a prior into our proposed CDiffMR, allowing it to learn the specific k-space undersampling process (R1, R3). Additionally, our method can also be applicable to reconstruction with different undersampling schemes. We have added tests using new Radial and Spiral KSUD, and provide the new results of reconstruction using Radial10% and Spiral10% masks. CDiffMR-C: Cartesian KSUD, presented in the main paper. CDiffMR-R/S: Radial/Spiral KSUD. Radial10% (Method:SSIM↑,LPIPS↓): D5C5:0.753,0.236; DAGAN:0.740,0.253; SwinMR:0.741,0.236; DiffuseRecon:0.750,0.280; CDiffMR-C:0.762,0.259; CDiffMR-R:0.765,0.227. Spiral10% (Method: SSIM↑, LPIPS↓): D5C5:0.749,0.254; DAGAN:0.725,0.260; SwinMR:0.726,0.259; DiffuseRecon:0.749,0.301; CDiffMR-C:0.749,0.301; CDiffMR-S:0.763,0.246. Our pre-trained CDiffMR-C model performs well for radial/spiral reconstruction tasks. Notably, CDiffMR-R/S, which utilises the undersampling scheme prior, achieves even better results, confirming the effectiveness of the KSUD.
FastMRI Benchmark (R4) We appreciate the suggestion regarding the FastMRI benchmark. However, we want to emphasise that our work offers unique insights into the field of MRI. Besides, we hope that R4 appreciates that the value of our technique, like other existing techniques, is not solely dependent on participating in the FastMRI challenge. In addition, due to different experimental settings, even if we submit to the FastMRI benchmark now, the comparison will be unfair. We will plan elaborative implementation on the FastMRI benchmark to ensure a fair comparison. This decision is primarily based on two points: 1) our contribution of unique insights through extensive experiments, and 2) the limited rebuttal time and policy set by MICCAI. We firmly believe that our approach holds its own merit and stands out independently. Our codes are anonymised and available at https://anonymous.4open.science/r/CDiffMR-53F4.
Ablation Studies on Gaussian Noise (R1) CDiffMR-GN can provide visually satisfying reconstruction results with rich details, but it is not compatible with our conditioning strategies. Consequently, CDiffMR-GN generates many “fake” details that do not exist in the ground truth images, resulting in extremely poor PSNR/SSIM scores but reasonable LPIPS. The quantitative results for CDiffMR-GN (AF:SSIM↑,LPIPS↓): x4:0.741,0.168; x8:0.650,0.243; x16:0.584,0.336.
Generative Abilities (R4) This paper focuses on MRI reconstruction rather than synthesis tasks. We have designed two conditioning strategies specifically to suppress any hallucination artefacts. This is necessary because generating “fake” details in MRI reconstruction tasks can reduce data fidelity and result in misleading information, which causes risks in clinical applications. CDiffMR can generate new samples by removing conditioning, but it would not be relevant to the task at hand.
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.
Authors provided some more details in the rebutal that should be included in the final version of the paper. Yet some concerns remain. Specifically those related ot objective evaluation using the fastMRI benchmark. Majority of the reviewers were positive about the paper. Therefore even though some concerns remain, I recommend to accept the paper for MICCAI.
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.
This paper combines MR physics with diffusion model for fast MR recon. The idea is interesting to the community. Concerns were raised regarding experiment and validation, which should be critical to the actual value of this work. The authors tried to explain their situations in terms of experimental design, which were mostly reasonable.
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.
The paper received diverse ratings initially and was brought the rebuttal phase. R4 recommended reject with several concerns, but did not respond to the authors’ rebuttal. I looked at the paper and the rebuttal. The authors provided a comprehensive rebuttal to address the reviewers’ concerns. The questions, in particular raised by R4, have been clearified in the rebuttal. The paper is currently in a good shape to publish in MICCAI.