Paper Info Reviews Meta-Review Author Feedback Post-rebuttal Meta-Reviews

Authors

Sudhanya Chatterjee, Suresh Emmanuel Joel, Ramesh Venkatesan, Dattesh Dayanand Shanbhag

Abstract

The ability of magnetic resonance imaging (MRI) to obtain multiple imaging contrasts depicting various properties of tissue makes it an excellent tool for in-vivo differential diagnosis. Multi-contrast MRI (mcMRI) exams (e.g. T2-w, FLAIR-T2, FLAIR-T1, etc. in a single examination) is a standard MRI protocol in a clinical setting. In this work we propose a method to accelerate mcMRI exams by acquiring only the essential contrast information for the slower contrast scans and sharing the un-acquired structure information from a fast fully sampled reference scan. The resulting artifact from the proposed structure-sharing method is then removed using a deep learning (DL) network. In addition, the proposed DL framework is driven by a smart loss function where weights of the loss function components are updated at end of each epoch using a transfer function based on its convergence to the ground truth. We show high quality accelerated reconstruction (up to 5x) at clinically accepted image resolution for (long) FLAIR-T1 and FLAIR-T2 scans using the proposed method with the faster T2-w as the reference scan.

Link to paper

DOI: https://doi.org/10.1007/978-3-030-87231-1_38

SharedIt: https://rdcu.be/cyhVl

Link to the code repository

N/A

Link to the dataset(s)

N/A

Reviews

Review #1

Please describe the contribution of the paper

This paper proposes to accelerate MR image acquisition by only acquiring the center of k-space and grafting the high-frequency information of a high-resolution image (of a different contrast) acquired during the same scan. Artifacts from this process are removed by training a deep neural network trained using actual high resolution scans.
Please list the main strengths of the paper; you should write about a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

This paper is well-written and tackles the approach of super-resolution from an interesting standpoint similar to keyhole imaging, which is common in dynamic imaging.

The method uses a dynamically weighted loss function that transitions from weighting MAE heavily to weighting MSSIM heavily as the images become more similar.

This method shows promise and would be a worthy contribution with the proper experiments.
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.

Unfortunately, this paper completely ignores modern parallel imaging techniques. Current MR imaging regularly accelerates scans by 2-4x through the use of parallel imaging from multiple coils. While the data is not cited and the acquisition is not described, it is highly unlikely that these images are not at least 2x accelerated in-plane through the use of SENSE or GRAPPA. This missing analysis and information is absolutely critical to the paper.

There is no description of how the data were acquired beyond the resolution and FOV of the images.

The comparison to random undersampling of the reconstructed image is simply not sufficient to validate this method with all of the acquisition information missing.
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

This paper will be difficult to reproduce without code provided by the authors, but can be tested on publicly available data.
Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://miccai2021.org/en/REVIEWER-GUIDELINES.html

Honestly, this paper requires a definite reworking with regards to modern parallel imaging techniques. This topic has a lot of promise building off of keyhole imaging, but without this comparison, there is no way to tell if the method is worthwhile.

In order to properly validate this method, the authors should collect raw imaging data from the scanner (after basic corrections but before parallel imaging reconstruction) and begin there. This way the full k-space data is available from every coil. Then the authors can compare against standard parallel imaging (skipping every n lines), compressed sensing methods (pseudo-random sampling of PE lines) and the proposed method, all while retaining the original multi-coil data. This would also give the authors the opportunity to extend their method to take in information from multiple coils.

The authors should describe in more detail their dataset (imaging parameters, resolution for all images) and their deep learning structure (implementation, training methods, etc.). As it stands right now (without any code provided), there is little hope of properly reproducing this work.

A small comment that normally FLAIR images are referred to as T1-FLAIR and T2-FLAIR rather than FLAIR-T1 and FLAIR-T2.
Please state your overall opinion of the paper

strong reject (2)
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

Without comparisons to common, modern, parallel imaging techniques, this paper cannot be accepted.
What is the ranking of this paper in your review stack?

5
Number of papers in your stack

5
Reviewer confidence

Very confident

Review #2

Please describe the contribution of the paper

The paper proposes a method to accelerate multi-contrast MRI exams by: (1) acquiring a full-length reference scan (2) accelerating the remaining scans in the exam by acquiring low-frequency k-space only, and (3) transferring high-frequency content onto scans in (2) from scan in (1), and finally - improving image quality via deep learning.
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.
- Novel approach: leveraging high-frequency content from a full-length reference scan to accelerate other scans and improving resulting image quality via deep learning is innovative. Presented results are convincing about the technique’s applicability and extensibility
- Residual learning: experience in deep learning is demonstrated by the apt usage of residual block
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.

None
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
Paper is not straightforward to reproduce:
- What are the acquisition parameters of the acquired data?
- What are the other implementation details of the DL network?
- “filter size, filter count and growth factor of 3, 16 and 12 respectively” is ambiguous
- What was the split between healthy vs pathological data?
Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://miccai2021.org/en/REVIEWER-GUIDELINES.html
The authors are encouraged to convert this work into a journal submission, with the following additions:
- Improve all figures for increased legibility
- Acquiring data with other undersampling masks (radial, spiral)
- Demonstrating generalization performance (or lack thereof) across at least one other anatomy and/or contrast
- Radiologist survey to assess diagnostic value
Please state your overall opinion of the paper

strong accept (9)
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
- Novel approach
- Relevance of topic
- Good overall study design
What is the ranking of this paper in your review stack?

1
Number of papers in your stack

4
Reviewer confidence

Confident but not absolutely certain

Review #3

Please describe the contribution of the paper

This paper proposes a new method to accelerate the data acquisition of FLAIR-T1 and FLAIR-T2 MRI. The method acquires only low-frequency k-space data in FLAIR MRI and combines it with the high-frequency k-space data in T2-W MRI which provides structural information. Deep learning is used to remove artifacts in the image reconstructed from combined k-space data. Experiments show a superior performance of the proposed method for reconstructing undersampled brain FLAIR-T1 and FLAIR-T2 images.
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. The method is novel. Grafting high-frequency k-space data of the reference image into the low-frequency data of target image is novel. Previous deep learning approaches mainly performed zero-filling on the k-space data. The dynamically weighted loss is also novel. Previous approaches mainly used a fixed weight or MAE-only loss.
2. There is also novelty in application. Previous studies used deep learning to accelerate T2-W MRI and used T1-W MRI as reference, while this study aims to accelerate FLAIR-T1 and FLAIR-T2 MRI and uses T2-W as reference.
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.
1. The data collection process isn’t clear enough. The authors didn’t provide enough information about the MRI scanner, the acquisition protocols, the patient cohort (e.g., what diseases are covered). It’s also unclear whether the dataset only includes brain images.
2. The performance measures are not reported with confidence intervals or standard deviations in Table 1.
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 data and code are not available to public. Details about the data acquisition process are missing (see major weakness 1). The description of algorithm is clear and quite detailed. For better reproducibility, the authors can add more details about model training (e.g., number of epochs, how is the network initialized, what optimizer is used, etc.)
Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://miccai2021.org/en/REVIEWER-GUIDELINES.html
1. Text in Figures 2, 3 and 4 is too small and not readable. Please enlarge the text. Consider moving some figures to supplementary materials (e.g., Fig 3 or 4) if space is limited.
2. Add standard deviations or confidence intervals in Table 1.
3. Add details about data acquisition process (scanner, protocol, etc.) and more information about the dataset (what organs and diseases are included?)
4. Add discussion on limitations of the method.
5. Add analysis about the visual results in Fig 5 and 6. Does the proposed method yield better visual results?
Please state your overall opinion of the paper

Probably accept (7)
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

The paper has good novelty in methodology. The idea of k-space grafting and using deep learning to remove grafting artifacts is interesting and can be applied to other problems in the field. The paper also studies the new application of accelerating FLAIR MRI using T2-W as reference.
What is the ranking of this paper in your review stack?

2
Number of papers in your stack

4
Reviewer confidence

Confident but not absolutely certain

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 reviewers found this paper to be well-written with an interesting and promising premise. However, two key weaknesses were noted. First, the paper largely ignored currently available parallel imaging techniques for acceleration. Given that the level of acceleration of the proposed approach (up to 5x) is similar to current techniques, it is not clear much of an advantage is provided. Second, there is limited information about the MRI acquisition used in the evaluation of the method. Please address these issues in the rebuttal.
What is the ranking of this paper in your stack? Use a number between 1 (best paper in your stack) and n (worst paper in your stack of n papers).

4

Author Feedback

We would like to thank all reviewers for their comments. Please find our response below.

R#1 on parallel imaging (PI) techniques: An important goal of this paper was to propose a method to accelerate long, SNR starved but clinically critical inversion recovery sequences in mcMRI exams (as FLAIR sequences). From an acquisition perspective, acceleration can be obtained in multiple ways (as discussed in [1,5] referred to in “Introduction”): 1) hardware based acceleration, as PI, where acceleration is limited due to g-factor which penalizes SNR (more so relevant for SNR starved FLAIR imaging sequences) 2) readout based methods which rely on signal encoding and pulse sequence design (PSD) such as simultaneous multi-slice (SMS) readouts, 3) sampling and reconstruction based accelerations, such as pseudo-random undersampling+CS/DL-recon. Our method falls in the third category; hence comparisons were done with pseudo-random sampling schemes and reconstruction [7,13,4,8] in our study. Our method proposed a novel sampling and reconstruction scheme which provides scan time reduction by exploring keyholing for mcMRI examinations and DL reconstruction. Any PI/SMS techniques shall run on top of such a sampling and reconstruction-based acceleration methods to further accelerate image acquisition. There is plenty of literature exploring this complementarity (e.g. [Saranathan et al.,2012,JMRI],[Mende et al.,2007,JMRI],[Hadizadeh et al.,2011,Eur.J.Radiology], [Muckley et al.,2020,arXiV] etc.). To summarize, the proposed novel approach to accelerate mcMRI examinations is complementary to PI techniques (hence should not been seen as a replacement or contender for PI/SMS which are based on Tx/Rx design and PSD principles). To the specifics of this study, all contrast data was acquired using a multi-channel Rx coil and our method enables acceleration over any PI based acceleration already present. Joint optimization to decide realizable acceleration limits using proposed mcMRI-keyholing approach and PI/SMS simultaneously can be a prospective study to this work.

R#1,R#2,R#3 on MRI data acquisition details: Three sets of private dataset were used for the study. Abbreviations: Slice thickness (ST), spacing b/w slices (SS), in-plane resolution (PR). FLAIR T2: ST-5mm, SS-6mm, PR- range of 0.43mm to 0.91mm isotropic. FLAIR T1: ST-5mm, SS-6mm, PR-in range of 0.43mm to 0.56mm isotropic. T2FSE: ST-5mm, SS-6mm, PR- in range of 0.43mm to 0.45mm isotropic. Data split w.r.t. field strength: FLAIR T1: Train #1.5T/#3.0T:67/163, Test #1.5T/#3.0T:5/20. FLAIR T2: Train #1.5T/#3.0T:176/41, Test #1.5T/#3.0T:18/3. All images were resampled to matrix size of 256x256 for DL training. Only brain MRI data was used.

R#2,R#3 on DL specs: Specified DL model features: filter size, growth rate, filter count correspond to Densenet model definition described in Huang et. al 2017, reference [6]. All models were trained using Adam optimizer for 300 epochs (loss curves shown in supplementary material). Training was performed in Python using TensorFlow v1.14.0.

R#3 on confidence interval or standard deviation (S.D.) for performance metrics: S.D. on performance metrics for all models are available and shall be included upon request. The proposed method performed better in this regard as well. Owing to the character limitation, we provide an example here for coeff of variation of SSIM for FLAIR T1 reconstruction (mean SSIM provided in Table-1 of paper): Proposed 2x/3x/4x/5x-0.018/0.037/0.061/0.082. Random 2x/3x/4x/5x-0.141/0.119/0.125/0.118.

#R2 on alternate sampling schemes: All images were acquired using Cartesian scheme in this study and the results hold for cartesian acquisition. However, this concept of mcMRI keyholing can be extended to other sampling schemes such as radial and spiral as well where only contrast sections (e.g. points around center of the spokes in case of radial) are acquired for accelerated scans and peripheral k-space is shared from a reference scan

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 sufficiently addressed the main concern from Reviewer 1. The acquisition details that were lacking in the original submission should be included the revision. Given that other than these criticisms, reviewers were generally positive about the approach and results, I favor acceptance.
After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.

Accept
What is the rank of this paper among all your rebuttal papers? Use a number between 1 (best paper in your stack) and n (worst paper in your stack of n papers).

5

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 describes an approach acceleration multi-contrast MRI (here T1 and T2 FLAIR) by grafting and deep learning of superfluous information. This is a really well written paper as pointed out by the reviewers. Major strengths include:
- Interesting and promising premise, real world application.
- well written, well described, in general a paper of high quality.
The weaknesses identified included:
- Existing parallel imaging techniques were ignored
- Limited information about the acquisition.
Both were addressed in the rebuttal, the paper is very clear with lovely schemata helping the reader understand this interesting and novel method.
After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.

Accept
What is the rank of this paper among all your rebuttal papers? Use a number between 1 (best paper in your stack) and n (worst paper in your stack of n papers).

1

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.

This work presents a novel acceleration method in the acquisition of multi-contrast MRI (FLAIR-T1 and T2). The proposed idea is to acquire only low-frequency k-space data one contrast and combine it together with the high-frequency k-space data of the other contrast. A deep learning -based strategy is also presented for artefact removal in the image reconstructed from multi-contrast k-space data. The less experienced reviewers are more favorable to this paper while the more confident one has a strong argument against this work as it does not compare with parallel imaging. I think though that the authors in the rebuttal well-position the type of acceleration technique they propose and why their selected comparison approaches are justified. I agree with the reviewer that PI should also be explored but for a conference paper, in my opinion, the presented work is providing already a relevant contribution. The combination and comparison with PI should though be included in a future extension/journal of this work. I think the authors also clarified the missing details requested by the met-reviewer. Disclaimer: I am not expert in MR physics.
After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.

Accept
What is the rank of this paper among all your rebuttal papers? Use a number between 1 (best paper in your stack) and n (worst paper in your stack of n papers).

8

back to top

Acceleration by deep-learnt sharing of superfluous information in multi-contrast MRI