List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Jaejin Cho, Yohan Jun, Xiaoqing Wang, Caique Kobayashi, Berkin Bilgic
Abstract
Diffusion MRI is commonly performed using echo-planar imaging (EPI) due to its rapid acquisition time. However, the resolution of diffusion-weighted images is often limited by magnetic field inhomogeneity-related artifacts and blurring induced by T2- and T2* relaxation effects. To address these limitations, multi-shot EPI (msEPI) combined with parallel imaging techniques is frequently employed. Nevertheless, reconstructing msEPI can be challenging due to phase variation between multiple shots. In this study, we introduce a novel msEPI reconstruction approach called zero-MIRID (zero-shot self-supervised learning of Multi-shot Image Reconstruction for Improved Diffusion MRI). This method jointly reconstructs msEPI data by incorporating deep learning-based image regularization techniques. The network incorporates CNN denoisers in both k- and image-spaces, while leveraging virtual coils to enhance image reconstruction conditioning. By employing a self-supervised learning technique and dividing sampled data into three groups, the proposed approach achieves superior results compared to the state-of-the-art parallel imaging method, as demonstrated in an in-vivo experiment.
Link to paper
DOI: https://doi.org/10.1007/978-3-031-43907-0_44
SharedIt: https://rdcu.be/dnwdq
Link to the code repository
https://github.com/jaejin-cho/miccai2023
Link to the dataset(s)
https://www.dropbox.com/s/2rteu3vmtbj15kx/example.mat?dl=0
Reviews
Review #2
- Please describe the contribution of the paper
The authors proposed a method to reconstruct diffusion images acquired with multishot schemes. The authors use a zero-shot deep learning architecture for this. The topic is relevant as multishot sequences are less prone to artifacts than EPI ones.
- 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 paper focuses on dMRI. Previous related methods have been applied to other sequences (e.g., ref 19). Although the experiments can be improved, the method is sound. The authors share code, which is greatly appreciated by the research community.
- 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 main architecture of Fig 1. processes in parallel the image and the k-spaces. Convolutions in the k-space are equivalent to multiplications in the image space. Thus, the whole processing can be performed in the image space with an improvement in efficiency. There is no motivation for processing the k-space or explanations why this is important.
- The use of the virtual coil is not motivated.
- In the experiments, it says that an acceleration of R=5 was used. The acceleration method is missing. It is not clear if partial Fourier is included in the R=5.
- The number of iterations of the DC layer is fixed to 10. A sensitivity analysis or a motivation of the choice of this hyperparameter is missing.
- It is not clear how the 2nd crossing fiber images are constructed.
- Since the authors have a sliver standard (5-shot EPI data), it might be possible to perform quantitative evaluation of the methods. Those are missing.
There are some comments related to MR physics:
- In the abstract, it is mentioned that multishot sequences are affected by T2 and T2* voxel blurring. Almost all dMRI sequences are SE, which means that there are no T2* effects.
- diffusion MRI usually disregards the phase of the signal. Thus, even if there are shot-to-shot phase differences, they might become irrelevant for reconstruction
- The voxel size is 1x1x4. This setting might be used for DWI, but isotropic voxel sizes are recommended for HARDI, which is the main application mentioned by the authors.
- 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
They provide code, so it is highly reproducible.
- 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
I think it would be good to have a medical physicist in the team to test the data in more relevant images. Improving the experimental section is crucial (see my comments of the “drawbacks” section).
- 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 paper has important drawbacks, but in general, the method is sound. For this, I am leaning to accept it.
- 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 #3
- Please describe the contribution of the paper
This paper proposes a new method for multi-shot diffusion-weighted MRI with zero-shot self-supervised learning reconstruction. The method uses zero-shot self-supervised deep learning reconstruction with a virtual coil and results show improved performance compared to related 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 method is novel and the results seem promising.
- 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 cleear if the comparisons are fair.
- 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
Acceptible
- 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
Title: Improved Multi-Shot Diffusion-Weighted MRI with Zero-Shot Self-Supervised Learning Reconstruction
This paper proposes a new method for multi-shot diffusion-weighted MRI with zero-shot self-supervised learning reconstruction. The method uses zero-shot self-supervised deep learning reconstruction with a virtual coil and results show improved performance compared to related methods.
The paper compares against LORAKS, but it is not clear which version of LORAKS is being used. Is LORAKS implemented with virtual coils as in Ref. {A]? Or is it implemented using the phase information from Refs. [8] or [9]? The comparison might be unfair if LORAKS is implemented with just the support information from Refs. [8] and [9], and it is important to be specific about how LORAKS was used.
[A] T. H. Kim, J. P. Haldar. LORAKS Software Version 2.0: Faster Implementation and Enhanced Capabilities. USC-SIPI-443, May 2018.
Since the authors are using virtual coils, it is also important to acknowledge that existing methods like LORAKS already use methods like virtual coils [A] or symmetry constraints [8,9,A] for exactly the same purpose.
The paper is not citing the LORAKS work on EPI [B]-[C], and it is not clear whether the LORAKS implementation is based on the state-of-the-art method for EPI [C]. The LORAKS reconstruction results in this paper look much worse than the RAC-LORAKS results in Ref. [C] for similar acceleration which may indicate that the paper is comparing against one of the outdated methods. If so, then this is an important caveat for the paper to mention.
[B] R. A. Lobos, T. H. Kim, W. S. Hoge, J. P. Haldar. Navigator-free EPI Ghost Correction with Structured Low-Rank Matrix Models: New Theory and Methods. IEEE Transactions on Medical Imaging 37:2390-2402, 2018.
[C] R. A. Lobos, W. S. Hoge, A. Javed, C. Liao, K. Setsompop, K. S. Nayak, J. P. Haldar. Robust Autocalibrated Structured Low-Rank EPI Ghost Correction. Magnetic Resonance in Medicine 85:3404-3419, 2021.
- 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 method is novel, and the results are promising. It is not clear if the comparisons are fair, but this can be okay as long as the authors revise to make any limitations of their comparison clear.
- 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
This paper works on the multi-shot diffusion-weighted MRI, using zero-shot self-supervised learning. The proposed approach is based on the learning denoisers in both k-space and image-space, in the formulation of reconstruction model for each diffusion direction. The zero-shot SSL is implemented inspired by [19] based on three splits of sampling masks for network training. Experiments show better results than the compared methods of SENSE, and LORAKS.
- 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 task of multi-shot diffusion-weighted MRI, based on zero-shot SSL, is an important task. The proposed approach is based on the learning denoisers on k-space and image-space.
- 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 major limitations of this work include the unclear writings and insufficient experimental evaluations.
(1) Eqn.(3) shows the reconstruction model for n-th diffusion direction. In the optimization problem, it involves two learnable networks N_i, N_k. How to train N_i, N_k and optimize x based on this optimization problem?
(2) The zero-shot SSL is not clearly presented. How to define the zero-shot SSL for MRI, and why this approach can achieve zero-shot SSL?
(3) How about the computational time for processing/reconstruction on each subject, including the training time?
(4) The results only show limited examples in Fig. 3 and 4, how about the quantitative comparisons on the reconstruction results using different methods?
(5) In the visual comparisons, it is better to magnify the patches to show the differences of results produced by different methods.
- Please rate the clarity and organization of this paper
Poor
- 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 paper provided a link to the code and data. However, the implementations details in the paper are not sufficiently clear, e. g., network hyper-parameters (filter sizes, etc.). The formulation of DC layer should be also given for this task.
- 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
Please see the strengths and weakness. The authors are suggested to improve the clarity of the writings, for example, the optimization in Eqn. (3) and its relation to the training of denoisers and the image reconstruction. It is also better to conduct ablation analysis on the image and k-space denoisers.
- 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?
My major concerns are on the unclear writing, insufficient experiments, comparisons and ablations studies.
- 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
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.
The paper presents a method for multi-shot diffusion-weighted MRI with zero-shot self-supervised learning reconstruction. The topic is important as multishot sequences are less prone to artifacts than EPI ones. The method is interesting and the results seem promising. In the rebuttal, the authors are expected to: 1) Clarify the comparison methods utilized, demonstrate whether the comparisons were made in a fair manner, and include relevant LORAKS works. 2) Improve the paper writting. Clarify the method details and experimental settings. 3) Provide quantitative results of different methods for comparison. 4) Add computational complexities of different methods.
Author Feedback
CNN in k-space (R2, R4): Nonlinear activations in the k-space network introduce distinct characteristics from the image domain network. In practical applications, Ref [2,6] showed that the CNN denoiser is effective in different ways in the image and k-space domains. Specifically, the k-space CNN efficiently eliminates folding artifacts by exploiting annihilation filters, while the image-domain CNN is more effective in denoising. Ref [2,6] also demonstrated the merits of using both domains, resulting in an improvement in performance based on PSNR and SSIM.
Virtual coil (R2): Ref [5] demonstrated that incorporating complex-conjugate signals from conjugate symmetric k-space regions enhances reconstruction quality in parallel MRI. In EPI, [Liao NIMG’22] shows virtual coil reduces the g-factor penalty by 20%.
Reduction factor (R2): 6/8 Partial Fourier was used in conjunction with uniform R=5 acceleration. As a result, ~15% of the k-space was covered in each shot.
Hyperparameters (R2, R4): For the 16 layer-CNN, we employed a filter size of 3x3 and 46 network depth, resulting in a total of 583,114 trainable parameters. Our network architecture is based on the MoDL structure. Ref [1], the original MoDL paper, demonstrated that the performance saturates at 8-10 iterations, resulting in a similar number of network parameters as in our study. We adopted 10 block iterations based on this rationale.
Crossing fiber (R2): We used the Bayesian Estimation of Diffusion Parameters Obtained using Sampling Techniques (BEDPOSTX) to estimate multiple fiber orientations [4,12, Hernández PLoS One’13]. RGB colors correspond to the 2nd crossing fiber in the x, y, and z directions, respectively.
Quantitative evaluation (R2, R4): Fig S1 and S2 demonstrate notable reductions in NRMSE (14%) and NMAE (9%) compared to LORAKS.
MR physics (R2): In the EPI acquisition, k-space lines are acquired during a long echo train, leading to mixed T2 and T2* voxel blurring in the phase encoding direction. Shot-to-shot phase differences makes it difficult to directly combine the complex multi-shot data. We expect the isotropic voxel size can be achieved by incorporating the g-Slider technique (Ref [16]).
LORAKS (R3): We utilized S-LORAKS, which employs phase information and k-space symmetry. Our data and approach differ slightly from those used in Ref [C]. Specifically, our data includes navigators and does not involve ACS data, whereas Ref [C] incorporates ACS regions but does not use navigator information. We will properly acknowledge the existing papers, including Ref [A-C] and [T.H. Kim, MRM’17].
Paper writing and English (R4): The paper will undergo revision with the guidance of native speakers, and grammar will be examined using appropriate software.
Optimization problem (R4): Eqn.(3) involves two trainable networks and DC. Based on Ref [2], we defined Nx = x - Dx, where D is CNN. Using alternating minimization-based solution; DC layer: x_{o+1}=(A^H A + λ_1 I + λ_2 I) (A^H b + λ_1 ζ_o + λ_2 η_o), where A is FC; Network Denoisers: η_{o+1}= V^H_C F^H D_k F V_C x_{o+1}; ζ_{o+1} = V^H_C D_i V_C x_{o+1}, where o is the optimization step (iteration) number.
Zero-shot SSL (R4): Ref [19,20] proposed zero-shot SSL for MRI, which performs scan-specific training without any external training database. To achieve this, zero-shot SSL employs physic-guided deep learning reconstruction and splits the undersampled k-space into three disjoint groups for training, calculating the training loss, and validating the network, respectively. The proposed method extends the zero-shot SSL approach for dMRI, also performing scan-specific training without any external training database.
Computational time (R4): The training time for the proposed network was 22:30 mins per diffusion direction/slice (GPU). This is expected to be reduced by transfer learning. Inference took ~1 second per direction/slice. 2-shot LORAKS took ~20 seconds per direction/slice (CPU).
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.
In the rebuttal, the authors have addressed most of the issues. The paper can be accepted after incorporating the modifications.
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.
The authors response to the reviewers’ and AC’s comments is satisfactory, especially to the more critical comments from R2, which were mostly regarding clarity of the method and experiments. I suggest we ask the authors to implement their response in the camera-ready version of the paper.
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.
After carefully evaluating the authors’ feedback and the final decisions of the reviewers, this paper has received mixed scores. One reviewer leans towards rejection, expressing concerns about the lack of clarity in the proposed technique, including the design and experimental insights.
On the other hand, two reviewers lean towards acceptance. However, they also highlight concerns about the unclear comparison fairness, lack of clarity in the design, and the need for more clarity in the experimental results.
Considering the mixed scores and the concerns raised by the reviewers, the Meta Reviewer suggests rejecting the paper. The overall evaluation of the paper’s scores supports this decision.