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

Authors

Jinwei Zhang, Hang Zhang, Chao Li, Pascal Spincemaille, Mert R. Sabuncu, Thanh D. Nguyen, Yi Wang

Abstract

Quantitative imaging in MRI usually involves acquisition and reconstruction of a series of images at multi-echo time points, which possibly requires more scan time and specific reconstruction technique compared to conventional qualitative imaging. In this work, we focus on optimizing the acquisition and reconstruction process of multi-echo gradient echo pulse sequence for quantitative susceptibility mapping as one important quantitative imaging method in MRI. A multi-echo sampling pattern optimization block extended from LOUPE-ST is proposed to optimize the k-space sampling patterns along echoes. Besides, a recurrent temporal feature fusion block is proposed and inserted into a backbone deep ADMM network to capture the signal evolution along echo time during reconstruction. Experiments show that both blocks help improve multi-echo image reconstruction performance.

Link to paper

DOI: https://doi.org/10.1007/978-3-030-87231-1_23

SharedIt: https://rdcu.be/cyhU7

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 work proposes multiple extensions to a learning pipeline optimizing both the sampling pattern and the regularizer for a multi-echo gradient echo sequence. The sampling pattern optimization method is extended to the multi-echo scenario and a unrolled iterative reconstruction based on ADMM is extended with temporal processing to take the signal evolution into account. The authors evaluate their method by extending a related reconstruction approach to this setting and providing quantitative and qualitative comparisons.
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.
- Extensive and solid evaluation
- Seemingly close to clinical applicability
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 presentation does not highlight the introduced innovations and enough and makes it hard to separate them from the state-of-the-art
- Variable-density patterns are disadvantageous in clinical settings because they do not form continuous trajectories
- Individual contributions seem like small adaptations of previous methods to a technologically more mature level
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

Reproducibilty should be excellent since the authors will publish the code and dataset.
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 present a huge amount of complex material. A later journal contribution will allow for better introduction of all components. For this conference contribution the abstract and conclusion could be modified to highlight the particular contributions of this work. From my understanding those consists mostly of extending LOUPE-ST to the multi-echo setting and introducing a deep learning model which can capture temporal dependencies in the unrolled-iterative ADMM reconstruction. It is hard to understand from the section about extending LOUPE-ST why this extension is not straightforward. The authors should explain the associated problem and solution in more detail or cut down on this part. Deep-ADMM seems to be presented as a new variant of unrolled-iterative reconstruction. However, when searching online, multiple authors already proposed unrolling ADMM updates (e.g. https://ieeexplore.ieee.org/document/8550778 ). The authors should highlight the differences of their method to previous methods. I want to emphasize that I appreciate that the authors thoroughly compare their method against a modified MoDL baseline and carry out hypothesis testing to validate the benefits of the proposed method. However, if I interpret the results correctly the sampling pattern optimization seemed to improve results significantly while using the Deep-ADMM model did not significantly outperform the MoDL method. The authors should discuss this and provide insight why they choose to present deep-ADMM nevertheless. A technological concern from my side is the optimization of a variable density pattern. As frequency-encoded readouts in MRI are extremely fast compared to phase encoding steps, continuous trajectories are vastly preferable. In practice undersampled cartesian and spiral trajectories are therefore prevalent. The authors should at least discuss this topic and how this could be addressed in practical applications.
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 proposes extensions of current methods on a relevant MRI acquisition technique. The evaluation is thorough and the results potentially interesting for the MRI community.
What is the ranking of this paper in your review stack?

1
Number of papers in your stack

5
Reviewer confidence

Somewhat confident

Review #2

Please describe the contribution of the paper

This study proposes a deep learning method to optimize the acquisition and reconstruction of a multi-echo gradient echo MRI sequence for QSM mapping. The authors adapt the previous LOUPE-ST method to the multi-echo acquisition for temporal sampling optimization, and design a temporal feature fusion block merged into a deep ADMM unrolled reconstruction network to capture the signal evolution of multiple echoes. The proposed method shows superior performance compared with model-based deep learning and LLR reconstruction 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.

A sampling pattern learning block to learn the reconstruction and optimal sampling masks simultaneously. Temporal feature fusion block to capture the dynamic signal evolution of the multi-echo images. Thorough validation and ablation study to confirm the contribution of each component.
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.

Not validated in prospectively undersampled data. The QSM quantification accuracy is not evaluated.
Please rate the clarity and organization of this paper

Good
Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance

The code and datasets are not provided. The reproducibility is moderate.
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. In deep ADMM, the output of each iteration contributed equally to the final reconstruction loss. Why not use a weighted combination of them, considering later iterations should contribute more?
2. In page 6, the authors describe the training loss as “by minimizing a channel-wise structural similarity index measure (SSIM) loss”. Actually, SSIM should be the large the better. Please clarify.
3. In page 3, ‘j-th coil’ should be ‘j-th echo’.
4. The authors are encouraged to discuss how the proposed method can be adopted in real MRI acquisition.
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 idea of simultaneous sampling pattern optimization and reconstrction is not new. But the novety here is to extend the previous method to multi-echo MRI.
What is the ranking of this paper in your review stack?

2
Number of papers in your stack

5
Reviewer confidence

Very confident

Review #3

Please describe the contribution of the paper

This paper proposes methods for k-p space sampling pattern optimization and image reconstruction in QSM with MEGRE sequence. It extends the LOUPE-ST method to multi-echo imaging for sampling pattern optimization, and adds a novel temporal feature fusion block into the deep ADMM model to better capture temporal correlation during reconstruction. Experiments showed superior results from the proposed method on a human brain dataset.
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. Using recurrent networks in the deep ADMM method to capture temporal dynamics is quite innovative. Experiments showed significant improvement in image reconstruction performance with this new method. The idea is also generalizable to other multi-echo imaging tasks besides QSM.
2. The authors clearly explained the data collection, processing and division methods in Section 3.1, providing details on the scanner, imaging protocols, post-processing algorithms, and data size (number of subjects/slices).
3. The proposed method is compared with a sufficient number of competing methods and shows consistently superior performance in PSNR, SSIM and visual results. The ablation study is comprehensive and demonstrates the importance of each novel component in the proposed method.
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 dataset size is relatively small, consisting of 7 subjects. No cross validation is performed to better evaluate the method’s generalization ability.
2. The novelty in the sampling pattern optimization block is not clear to me. How did the authors extend LOUPE-ST to multi-echo scenario? What’s the difference between the proposed method and LOUPE-ST? How is the single-echo SPO method implemented?
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 indicate in the checklist that they will release data and code if paper is accepted. There are also abundant details on data collection and algorithm implementation in the paper (Sections 3.1 and 3.2). Overall the paper has very good reproducibility,
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. Please add more details on the proposed sampling pattern optimization block. Explain its difference from LOUPE-ST and the baseline single-echo SPO.
2. Some text in Figure 1 is too small. Please make the text more readable.
3. Consider adding cross validation.
4. The under-sampling is performed retrospectively in this paper. How would the method perform on new data acquired with the learned optimal sampling pattern? It’s interesting to see results on prospectively under-sampled data.
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 very good novelty and the ideas can be generalized to other multi-echo imaging tasks beyond QSM. The experiments are comprehensive and show solid improvement achieved by the proposed method.
What is the ranking of this paper in your review stack?

2
Number of papers in your stack

5
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.
This paper describes a deep learning method to optimize both acquisition and reconstruction of multi-echo gradient echo MRI for QSM mapping. A previous method (LOUPE-ST) is extended to be applied to multi-echo data. Training/Validation/Evaluation is performed on 3/1/3 subjects.

Strength:
- This is a very interesting & relevant problem, as QSM can suffer from long acquisition times and accelerating the scan will enable eg higher resolution.
- The paper is dense in details and thus allow the interested reader to follow the method in detail - which is great. (see the flip side of that in the weaknesses)
- The evaluation of all the new components is performed very thorough and is good to follow.
- The evaluation includes comparison with a range of methods and evaluation techniques.
- The authors indicate that they will make data and code available - which is great and will together with the thorough methodological description be of great use.
Weaknesses / areas for clarification:
- The optimization and choice of Uj, the sampling pattern is not very well described at the moment. While it is referred to as a “variable density sampling pattern with a fixed undersampling ratio”, it is not so clear how these are updated and optimized. - Explicitly questioned in the reviews were also the practicability of scanning such a pattern (as in how will k-space be effectively sampled, which gradient trajectories allow to achieve the final optimized pattern?). This relates to another reviewer comment, that no prospectively undersampled data is included. Seeing this optimized pattern in action and that the proposed algorithm (a) proposes a practical pattern and (b) works on prospectively undersampled data would be greatly useful.
- I completely agree with the reviewer’s comments that - given the density of the material - every effort should be made to increase overview and understanding for the reader. Esp. Figure 1 is key and needs urgent improvements such as larger text font, clearer flow of data, important test in bold, less white eg in (c) but increasing the blue/grey/sand boxes. Similarily - and in line with my previous comment - a more schematic illustration of the final sampling pattern (or even the evolution over iterations to the optimal one) would be of great use. -(Eg added to Figure 2). I am aware this is in the appendix, but the quality provided there does not allow to really assess the differences obtained over time. One idea might be to encode the echos with different colors and show the multi-echo pattern in one k-space?
- The chosen color map / scale in Fig. 2 is not helpful - would be good to rethink!
- Might be good to add a sentence to how this could improve other multi-echo acquisitions.
- There are quite a lot of “a” and “the” and prepositions missing which makes the paper in parts somewhat hard to read - this could be easily improved!
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).

2

Author Feedback

N/A

back to top

Temporal Feature Fusion with Sampling Pattern Optimization for Multi-echo Gradient Echo Acquisition and Image Reconstruction