List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Prabhjot Kaur, Atul Singh Minhas, Chirag Kamal Ahuja, Anil Kumar Sao
Abstract
Limited accessibility to high field MRI scanners (7T,
11T) has motivated the development of post-processing methods to im-
prove low field images. Existing post-processing methods have
shown the feasibility to improve 3T images to produce 7T-like images [3]. It has been observed that improving lower field (LF, ≤ 1.5T) images
comes with additional challenges due to poor image quality such as the
function mapping 1.5T and higher field (HF, 3T) images is more com-
plex than the function relating 3T and 7T images. Except for [10], no
method has been addressed to improve ≤1.5T MRI images. Further, most
of the existing methods, including [10] require example images, and
also often rely on pixel to pixel correspondences between LF and HF im-
ages which are usually inaccurate for ≤1.5T images. The focus of this
paper is to address the unsupervised framework for quality improvement
of 1.5T images and avoid the expensive requirements of example images
and associated image registration. The LF and HF images are assumed
to be related by a linear transformation (LT).The unknown HF image
and unknown LT are estimated in alternate minimization framework.
Further, a physics based constraint is proposed that provides an addi-
tional non-linear function relating LF and HF images in order to achieve
the desired high contrast in estimated HF image. This constraint exploits
the fact that the T1 relaxation time of tissues increases with increase in
field strength, and if it is incorporated in the LF acquisition the HF con-
trast can be simulated. The experimental results demonstrate that the
proposed approach provides processed 1.5T images with improved image quality, and is comparably better than
the existing methods addressing similar problems. The improvement in
image quality is also shown to provide better tissue segmentation and
volume quantification as compared to scanner acquired 1.5T images. Its application on improving 0.25T images has proved its advantages.
Link to paper
DOI: https://doi.org/10.1007/978-3-031-43999-5_13
SharedIt: https://rdcu.be/dnwwr
Link to the code repository
https://drive.google.com/drive/folders/1WbzkBJS1BWAje8aF0ty2SWYTQ9i0B7Yr?usp=sharing
Link to the dataset(s)
N/A
Reviews
Review #1
- Please describe the contribution of the paper
The authors present a novel physics-based method for estimating 3T MR images from 1.5T images. The method addresses two challenges that previous methods have struggled with when applied to improve low-field MR image qualities: paired images and lower-field image improvement. To address these issues, this work proposes an unsupervised framework and introduces an additional physics-based constraint, a non-linear function linked with LF and HF images. The proposed method was tested on a custom dataset containing five healthy subjects of about 25 years from three different scans and compared to some existing contrast synthesis and HF image estimation methods using seven metrics. The experiment results demonstrate the proposed method’s effectiveness and robustness through ablation studies.
- 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.
- Advanced physics-based signal scaling is introduced to improve the LF image quality.
- In-depth analyses are provided on the effectiveness and robustness of physics-based regularization and multiple relaxation times computation of tissue voxels using a series of ablation studies.
- Subjective analysis by clinical experts has shown the superiority of the proposed 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 authors claim that different architectures of deep learning can achieve better performances than previous methods, although no existing methods address lower field image improvement. The supervised methods that the authors compared to (ScSR, CCA, MIMECS, ED) are almost from 2018 before. I strongly recommend that the authors review advanced works, such as some new supervised image reconstruction methods. In addition, it is necessary to include an evaluation and comparison between the performance of the proposed method and more advanced methods. I am curious about the performances of these methods in lower field image reconstruction.
- It is recommended that the authors provide more detailed rationales for AM framework design and its necessity. In Section 3.2, the authors claim that once the data fidelity and the physics based regularizer are combined in AM framework, the corresponding image is sharper and improved in contrast. However, in Fig. 1. (a), PSNR in (iii) is 0.01 higher than in (iv), so I am confused that the AM framework appears to have no benefits. In addition, I recommend the authors provide the experiment of (i) + NLM.
- 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
It is 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
1． Section 3.3, The paper claims that unsupervised nature can improve the robustness to inaccuracies among pixels than methods based on paired images, which seems weird or not well reasoned. In my view, data with lots of wrong labels have brought an unreasonable prior assumption, which will cause terrible results. The author addresses that the proposed approach is an unsupervised method. So I think the authors should compare the proposed methods with some novel unsupervised deep learning methods (e.g. GAN, diffusion model). 2． Section 3, lambda_2 is chosen as 0.4, what is the rationale behind the selection? Please provide more details on the hyperparameter tuning process, including either empirical evidence or theoretically/domain-justifiable strategies. 3． Section 3.1, can you provide more detail on the size of MR images and the inference speed?
- 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?
This paper is well written, easy to read and technically sound. A lot of experiments including qualitative and quantitative evaluations in image quality and brain tissues segmentation are conducted to show the effectiveness of proposed method. However, the dataset is small, hope authors can test on more cases acquired from more sites to prove the generalizability of proposed method.
- 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
The paper with title: Estimation of 3T MR images from 1.5T images regularized with Physics based Constraint presents an alternative minimization inverse problem framework for MRI super resolution. The authors present thorough analysis and visualization, very solid work.
- 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 present a novel optimization based approach for MRI super-resolution. The authors include a thorough studies, from image visualization, quantitative metric to user studies, very solid work.
-
The idea of Physics base regularizer is novel.
-
I appreciate the ablation studies for better understanding of the Physics based Regularizer term.
-
- 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.
-
Would love to learn about the optimization algorithm to save equation 1.
-
On physics-based regularizer, I appreciate this section, however, besides the dominant T1 relaxation difference, there are also other effects can contribute, for example diffusion effect, susceptibility effect, I wonder could you elaborate on how those other attributes impact the results?
-
- 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 claimed to release the source code.
- 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
Very solid work, one potential direction - incorporate physics-based regularizer to DL-based unrolled networks - would expect good results.
- 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?
Solid work, thorough analysis, good reasoning
- 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 #4
- Please describe the contribution of the paper
This paper presents a method for estimation of high-field-strength MRI from low-field-strength ones, using physics based regularization and ADMM-based optimization. The paper aimed at improving MRI acquired with low fields (e.g.,1.5T and below). Experimental results demonstrated better perceptual quality of reconstructed images and higher tissue segmentation accuracy of the proposed method.
- 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 targeted problem is relatively novel. Although reconstruction of 7T from 3T has been studied extensively, fewer works are devoted to the reconstruction of lower field images (e.g. reconstruction of 3T from 1.5T).
-
The advantages and disadvantages of the proposed method are clearly presented in the experiments section, showing worse referenced-based accuracy but improved perceptual quality and segmentation accuracy.
-
- 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.
- Writing and language need significant improvement, with numerous sentences with grammatical errors. In particular, the method part has important points that are unclear. See below for details.
- 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
Authors claimed to release code on GitHub.
- 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 address the following questions and issues.
-
In method, how is T1 relaxation time estimated? Is T1 assumed to be constant within all pixels with the same tissue type? How strong is this assumption?
-
How is c in eq (1) estimated? What is the relation between c and r?
-
What is the relation between sl and y? Please elaborate on “the acquired signal for y”.
-
What is the rationale to model transformation from HF to LF by a convolution in eq. 1?
-
The following sentences are either grammatically wrong, unclear, or hard to understand for the reviewer:
It has been observed that improving lower field (LF, ≤1.5T) images comes with additional challenges due to poor image quality such as the function mapping 1.5T and higher field (HF, 3T) images is more complex than the function relating 3T and 7T images [10].
In [10], the mapping between example 3T and simulated 0.36T images using CNNs to estimate 3T from scanner-acquired 0.36T images is learned,
The matrix c ∈ Rm×n represent the pixel wise scale when multiplied with y generates the image with contrast similar to HF image.
The factor r for the given voxel for SE sequence by assuming long TE, and T2 to be same across different FS (assumption derived from literature [12]) be computed as: …
In real practice there exist many voxels with more than one tissue kind present in them, and is the reason for disconitinuities present in Fig. S1.
Though this work performs well with probability maps but we use pixel intensity to denote the probability of tissue to avoid the additional and expensive tissue segmentation step as follows: …
“The quantitative measure used to evaluate performance of different methods is dice ratio, and is reported in Table S2, and its comparison is mentioned in Table S3.” - What does “its comparison” refer to?
Here, the knowledge of acquisition physics to simulate HF image is exploited, and used it in a novel way to regularize the estimation of HF image.
Lower the FS image is, harder is to get accurate image registration, and thus proposed method proves to be a better choice.
-
In Fig. 1, at which stage is the regularizer on h added?
-
In Tables 1,2, and 3, what does “HM” mean?
-
- 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 problem addressed and the method proposed seem interesting. However, the manuscript is unclear in writing, with numerous grammatical errors and many important points not clearly explained.
- 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 proposed work is novel and interesting, providing a new physics-based and optimisation-based approach for estimating HF MRI from LF MRI. However, there are some major concerns over the clarifications of the method and results. Please clarify the AM framework design and its necessity, and its interpretation of performance in Fig.1. Please also provide more details and clarifications for eq. 1 and choice of hyperparameters as questioned by reviewers in the rebuttal.
Author Feedback
This rebuttal addresses the clarifications of the AM framework used: its necessity, its impact in results in Fig.1, solving eq.(1), choice of hyperparameters of the equation.
Necessity of AM framework:Since proposed approach is unsupervised approach, both h and x are unknown. If solved for both h and x simultaneously, the optimization problem becomes non-convex which is difficult to solve. Instead this optimization problem is convex if it is solved for one of the h and x variables at a time, and is easier to solve. Hence, we chose the AM framework.
Impact of AM Framework on results: Combining data fidelity term with Physics Based Regularizer(PBR) had less effect on PSNR wrt using only PBR(iii)-Fig.1 In literature PSNR (a widely-used average measure of similarity with a reference image) is proved to give similar value even for blurred images: page7line 2-7. We caution against relying solely on PSNR and emphasize the need for considering other metrics. In our study, we observed that the difference of other quality metrics for images in (iii) and (iv) is significant in terms of image sharpness (3.19), PIQE (4.7), and SSIM (0.0762). Although we only mentioned PSNR for simplicity following the literature convention, these metrics can be included in Fig.1 to provide a comprehensive evaluation. Qualitative analysis clearly demonstrates that in order to obtain both sharp details and high contrast we need to combine (iii) and (ii) in AM framework, outperforming both DF and PBR (iii) and (ii) respectively, thus justifying the benefits of our proposed AM design. R1: (i) + NLM has been observed to provide blurred image details than (i) with similar contrast as in (i) - can be added in supp material. LF images can surely be simulated by simulating LF contrast and synthetic noise[16].
How to solve for eq.(1): First, randomly initialize the variable h and estimate x by keeping h fixed. The regularized estimation of x is followed from [doi: 10.1109/msp.2019.2949470]. For implementation, we utilized the cvx toolbox in Matlab.
Selection of hyper-parameters in AM: Lambda_2 is chosen empirically to maximize PSNR in a leave-one-out cross-validation fashion. Importantly, varying lambda_2 in [0.01:0.01:5] had minimal impact on PSNR (0.05dB). The primary purpose of lambda_2 is to prevent divergence of h (as observed when lambda_2=0) by constraining its energy.
Why convolution in eq.(1)? The rationale behind assuming linear mapping (LM) between LF and HF images is to capture general changes in images when FS is changed. The LM is represented with a convolution kernel (h) instead of a matrix because it reduces the number of parameters to be estimated. The non-linear mapping is being captured by PBR.
Relation between c and r Interesting question, the formulation of estimation of c and r is the same in this work. The variable r is the ratio of acquired signals for y and x, and is computed independently for different T1 relaxation times. The variable c is re-derived, similar to r i.e., is the pixel wise ratio of y and pseudo estimate of x to be inserted in eq.(1). The notation of variables is taken from literature[16].
Assumptions Are we assuming the T1 relaxation times for the same tissue pixels as constant? There are two ways to define pixel’s belonging to a tissue type : (i) binary belonging (WM/GM masks) or (ii) soft belonging (WM/ GM probability maps), and we indirectly used (ii) way. The pixels for which probability to belong to WM/GM is 1 (indirect: say WM(/GM)=top 5%ile(/bottom 20%ile) of pixel intensity) are assumed to have a constant T1 relaxation time. We agree that it may not be strong assumption because T1 relaxation may vary (GM often vary in different regions) but there are only a few pixels with probability=1. For remaining pixels, T1 relaxation time is estimated as a linear combination of T1 of WM and GM for these voxels instead of assuming it constant( advantages of using (ii) over (i) are illustrated in Fig.S1)
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.
The rebuttal is satisfactory and has clarified the major concerns raised in the reviews. The method is sound and interesting, with sufficiently well validated experiments. The clarifications are suggested to be included in the manuscript before publication.
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.
After rebuttal, the reviewers haven’t changed their scoring. However, based on my review of the rebuttal and the paper, the work can be accepted and the responses explained reviewers’ concerns.
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 thoroughly reviewing the authors’ feedback and the final decisions of the reviewers, it is evident that the majority of the reviewers are inclined to accept the paper. They recognize the significant impact of the work for the community.
While one reviewer recommends rejection, they still acknowledge the interesting nature of the proposed method.
After careful consideration of the reviewers’ feedback and opinions, the Meta Reviewer leans towards accepting the paper. The recognition of the paper’s impact by the majority of reviewers, coupled with the acknowledgment of its interesting methodology by the dissenting reviewer, suggests that the paper holds value and should be accepted for publication.