List of Papers Browse by Subject Areas Author List
Abstract
Parallel imaging (PI) has demonstrated notable efficiency in accelerating magnetic resonance imaging (MRI) using deep learning techniques. However, these models often face challenges regarding their adaptability and robustness across varying data acquisition.
In this work, we introduce a novel joint estimation framework for MR image reconstruction and multi-channel sensitivity maps utilizing denoising diffusion models under blind settings, termed Blind Proximal Diffusion Model in Parallel MRI (BPDM-PMRI). BPDM-PMRI formulates the reconstruction problem as a non-convex optimization task for simultaneous estimation of MR images and sensitivity maps across multiple channels. We employ the proximal alternating linearized minimization (PALM) to iteratively update the reconstructed MR images and sensitivity maps. Distinguished from the traditional proximal operators, our diffusion-based proximal operators provide a more generalizable and stable prior characterization. Once the diffusion model is trained, it can be applied to various sampling trajectories. Comprehensive experiments conducted on publicly available MR datasets demonstrate that BPDM-PMRI outperforms existing methods in terms of denoising effectiveness and generalization capability, while keeping clinically acceptable inference times.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/2563_paper.pdf
SharedIt Link: pending
SpringerLink (DOI): pending
Supplementary Material: https://papers.miccai.org/miccai-2024/supp/2563_supp.pdf
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{Li_Blind_MICCAI2024,
author = { Li, Xing and Yang, Yan and Zheng, Hairong and Xu, Zongben},
title = { { Blind Proximal Diffusion Model for Joint Image and Sensitivity Estimation in Parallel MRI } },
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
Parallel imaging (PI) and compressive sensing (CS) are techniques to accelerate the acquisition of MRI images and improve data quality. PI uses multiple receiver coils to simultaneously capture MRI data, allowing for faster data acquisition as the signals from different coils can be combined to reconstruct the image. CS is a mathematical framework that exploits the sparsity of signals to reconstruct images, which enables efficient under-sampling of the k-space data without compromising on quality of reconstructed images.
This paper introduces a framework to jointly estimate the MRI k-space-to-image reconstruction and multi-channel sensitivity map, leveraging a denoising diffusion probabilistic model (DDPM). The proposed framework, termed Blind Proximal Diffusion Model in Parallel MRI (BPDM-PMRI), uses the proximal alternating linearized minimization (PALM) to update the reconstructed images and sensitivity maps iteratively. The proposed framework shows competitive performance against
- 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 optimization problem is intuitive, and the authors made it clear where the diffusion models are involved in this optimization, and explained why it is reasonable for using diffusion model to approximate priors of MR images and sensitivity maps. However, among all variants, the choice of DDPM is not very well justified. See Weakness 1.
- Qualitative results look good. So does the ablation study.
- 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.
- Insufficient details are given regarding the diffusion models. In the Experiment section, it seems the authors trained the diffusion models to (presumably) generate the MR images and sensitivity maps from noise, following the DDPM training and DDIM inference — is that true? If so, what is the noise profile (presumably Gaussian noise) and how do you justify it under this problem setting?
- Experiments might be slightly under-sufficient, with only 1 dataset of 55 subjects.
- More issues deferred to detailed constructive comments.
- 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?
I am not too concerned about reproducibility as long as the authors plan to release the code.
- 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
-
In the Methodology section, the authors began describing the optimization problem without defining the problem formulation and/or the variables used. It will be very helpful if such background information can be provided. Variables and sub/super-scripts can be especially confusing without definition when it involves diffusion models.
-
In Figure 1, it looks confusing because while $x_t$ is purely noise in the DDPM forward diffusion, $x_t$ is a clean image in the IUM. Besides, showing the forward diffusion process alone is not very helpful to understand the overall framework. The authors may consider redesigning the figure.
-
In the Experiment section, it seems the authors trained the diffusion models to (presumably) generate the MR images and sensitivity maps from noise, following the DDPM training and DDIM inference — is that true? If so, what is the noise profile (presumably Gaussian noise) and how do you justify it under this problem setting?
-
Typo: Page 3, section 2, paragraph 2. “The first term” instead of “The fisrt term”.
-
Typo: Page 6, section 3. Do you mean “Experiments” or “Experimental Results”?
-
Minor issue: Page 5. The acronym “SoS” does not seem to be defined?
-
Minor issue: Page 5, section 2.2. Are two pretrained DDPM models parametrized differently? If so, I believe they shall be denoted as $R_\theta$ and $R_\phi$, respectively, both in the main text and in equation 8?
-
- 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?
The authors have not very well justified the specific choice of DDPM (rather than, for example, an image-to-image Schrodinger Bridge). Besides, the empirical results are not very outstanding judging from the single dataset they evaluated on. Quite a few comments to address. I will postpone the decision until after the rebuttal.
- 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
Weak Accept — could be accepted, dependent on rebuttal (4)
- [Post rebuttal] Please justify your decision
Although I am still finding the formulation counter-intuitive, with the DDPM forward process and image-to-image Schrodinger bridge reverse process for IMU, the authors have addressed my other concerns and I would like to increase the rating.
Rank in stack: 1/3.
Review #2
- Please describe the contribution of the paper
The authors developed a diffusion deep learning model to reconstruct MRI images of knees from parallel imaging (reduced k-space by including sensitivity maps) and compressive sensing (reconstruction of sparse k-space sampling) without calibration. They incorporates denoising diffusion models for both images and sensitivity maps. They showed accuracy advantages.
- 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.
- comprehensive testing was done against many traditional and deep learning alternatives
- some ablation studies was done to show the importance of both image and sensitivity map updates to validate the approach.
- 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 enhancement achieved seems modest for most of the settings.
- From the resulting images, it is not clear if the differences provided by the proposed methods can affect clinician’s work. There seem to be rather minor differences between proposed method and those of existing methods of Deep-SLR and DDPM. Will a doctor care which they use?
- 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?
no
- 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 proposed method seems to have novelty, and seems to work well, matching good existing techniques in terms of accuracy. The study is robustly conducted. However, the improvement seems to be quite modest and limited to specific settings. It further seems that a clinician may not care much about the minor improvements provided, as the resulting images in Fig 2 did not seem to differ much from technqiues such as Deep-SLR and DDPM. Will the proposed technique have a clinical impact?
- 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 work seems to have novelty and is well done, but the improvements seems modest
- Reviewer confidence
Not confident (1)
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
Weak Accept — could be accepted, dependent on rebuttal (4)
- [Post rebuttal] Please justify your decision
My concern was that the improvements were modest, and clinicians may not care whether or not the currently proposed method is used. The authors response did not address this much. However it is the case that the proposed method is robust and works well, and has the advantage of being unsupervised. So my response remains the same as weak accept.
Review #3
- Please describe the contribution of the paper
This paper proposed a diffusion model for MR image reconstruction. The method was evaluated on public dataset, outperformed previous 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 claimed that this is the first diffusion based model for blind CS-PI, which seems novel.
- 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
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 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?
No.
- 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
Can the authors explain the task of image and sensitivity estimation and Parallel MRI for layman?
- 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?
No.
- Reviewer confidence
Not confident (1)
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
Weak Accept — could be accepted, dependent on rebuttal (4)
- [Post rebuttal] Please justify your decision
Given my limited familiarity with this topic, it appears that the authors have addressed all the comments raised by the reviewers. I’m inclined to trust their expertise in this matter and accept the paper for publication.
Review #4
- Please describe the contribution of the paper
In this paper, the authors proposed a joint approximation approach of diffusion models for blind parallel reconstruction. The method is trained and tested using the NYU dataset (n = 55).
- 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.This joint optimization method is novel, and the authors used the PALM method in the optimization. 2.The authors compared their model with eight traditional and two deep-learning-based methods.
- 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.
My concerns are mainly about the experiments. The current experiments are not convincing enough to support the advantages of the new method.
- 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?
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
My concerns are mainly about the experiments. The current experiments are not convincing enough to support the advantages of the new method. 1.For the ablation studies in Table 2, the authors claimed that the IUM is more important. However, they did not give enough details of the inputs for all ablation studies. Did they use the under-sampled MR images x_t in the inputs of SMUM for the ablation study “w/o IUM”? If not, it cannot prove that the IUM is important. 2.For the acceleration factors R_A, authors only used R_A = 4, 6 and 8. Why they do not use R_A = 2? Since the results of their method when R_A = 4 are worse than Deep-SLR, I want to see the comparison between their method and the Deep-SLR when R_A is 2. 3.The authors only did experiments on one dataset, so it would be better if they could test their method in more datasets. 4.The font size of Table 2 is too small.
Moreover, the method part is hard to read. The authors need to make it more clear. Moreover, When training the method, are the IUM and SMUM trained only once, or do they need to be trained iteratively?
- 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?
This work have some weakness, but the authors have done enough work to clarify their work. They need to answer my concerns.
- 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
Weak Accept — could be accepted, dependent on rebuttal (4)
- [Post rebuttal] Please justify your decision
The authors answered all my questions, and I think the answers are convincing. There is only one concern, when R_A = 4, the Deep-SLR has achieved better results than the proposed method. When R_A = 2, will Deep-SLR achieve even better results?
Author Feedback
We thank all reviewers for their valuable comments and appreciate them for affirming the novelty and contributions of our paper. We carefully addressed the issues raised by the reviewers and itemize our responses to major points as follows: 1)R1 concerned about the modest performance improvement. Our method is an unsupervised approach. While the performance improvement over supervised methods like Deep-SLR is modest, our unsupervised method has strong generalization capability, enabling reconstruction under any sampling pattern and rate, which has a clinical impact and supervised methods cannot achieve. 2)R3 concerned about the the task of image and sensitivity estimation and Parallel MRI. Parallel MRI is a commonly used imaging technique in clinical practice. It involves the simultaneous acquisition of data using multiple receiver coils to accelerate reconstruction. Sensitivity estimation refers to the quantification of the receiver coil sensitivities. By estimating the sensitivity distribution for each coil, it becomes possible to accurately disentangle signals from different locations, thereby improving image quality. 3)R4 confused about the lack of problem formulation in the method section and the confusion about sub/super-scripts after incorporating the diffusion model. The content at the beginning of Chapter 2 is the problem formulation. If possible, we will rename that section as “Problem Formulation”. In addtion, before incorporating the diffusion model, the iteration number was represented by a superscript in the iterative algorithm. After adding the diffusion model, the iteration number is represented by a subscript to maintain consistency with the diffusion formula. 4)R4 confused about the representation of x_t in Fig.1 and redesigned it. In Fig.1, at the DDPM forward stage, x_t represents the noisy image. In the reverse diffusion, the input to the IMU is the undersampled image (x_t), and to maintain consistency with the diffusion formula, it is defined as x_t. In future, we will consider to redesign Fig 1. 5)R4 concerned about the diffusion models to generate the images and sensitivity maps following the DDPM training and DDIM inference, the noise profile and justification under this problem setting. The diffusion model is trained by progressively adding noise to fully-sampled images and their corresponding estimated sensitivity maps. During inference, undersampled images (R_A= 4,6,8) and initial sensitivity estimates (estimated from ACS sampling line) are used as inputs for iterative MR image generation. It reduces the number of sampling steps and has no need to sample from noise by incorporating the diffusion model, significantly accelerating the process. Thus, the noise profile remains Gaussian. As indicated by Eq. 6 in the manuscript, both sub-problems assume Gaussian noise conditions and correspond to Gaussian denoising formulations. The optimization algorithm for each step is consistent with the sampling process of DDPM and DDIM. 6)R4 and R6 suggested to modify the spelling and formatting issues and improve the readality of method section. We will make the corresponding modifications according to your suggestions and reorganize the method formulas in future. 7)R6 confused about the IUM importance in ablation studies and IUM and SMUM process. The under-sampled MR images x_t are input to the SMUM, which performs gradient descent without proximal operator. During training, the diffusion models for image and sensitivity are trained only once. IUM and SMUM are inference stages, the inverse diffusion sampling is combined with the iterative algorithm. 8)R6 concerned about the lack of experiments at R_A=2 and the need for more validation. Under R_A=2, traditional methods such as GRAPPA have achieved satisfactory reconstruction performance. To validate the superiority of deep learning methods, we conducted evaluations under R_A=4,6,8. In addtion, we will consider to validate our method on more datasets in future work.
Meta-Review
Meta-review #1
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
The reviewers have gave positive comments on the manuscript after rebuttals. One reviewer expressed concerns about the modest improvements and questioned the clinical relevance of the proposed method. The authors’ response did not fully address these concerns, yet the method is recognized for its robustness and unsupervised nature, resulting in a weak accept. Another reviewer, with limited familiarity with the topic, felt the authors satisfactorily addressed all reviewer comments and trusted the authors’ expertise, leading to an acceptance recommendation. A third reviewer found the formulation counter-intuitive but acknowledged that their other concerns were addressed and decided to increase their rating. Lastly, a reviewer raised a specific concern regarding the comparative performance of the proposed method and Deep-SLR, questioning whether Deep-SLR might achieve even better results under certain conditions. Overall, the work is acceptable.
- 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).
The reviewers have gave positive comments on the manuscript after rebuttals. One reviewer expressed concerns about the modest improvements and questioned the clinical relevance of the proposed method. The authors’ response did not fully address these concerns, yet the method is recognized for its robustness and unsupervised nature, resulting in a weak accept. Another reviewer, with limited familiarity with the topic, felt the authors satisfactorily addressed all reviewer comments and trusted the authors’ expertise, leading to an acceptance recommendation. A third reviewer found the formulation counter-intuitive but acknowledged that their other concerns were addressed and decided to increase their rating. Lastly, a reviewer raised a specific concern regarding the comparative performance of the proposed method and Deep-SLR, questioning whether Deep-SLR might achieve even better results under certain conditions. Overall, the work is acceptable.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
- 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).
N/A