List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Yuning Gu, Yongsheng Pan, Zhenghan Fang, Jingyang Zhang, Peng Xue, Mianxin Liu, Yuran Zhu, Lei Ma, Charlie Androjna, Xin Yu, Dinggang Shen
Abstract
In human MRI studies, magnetic resonance fingerprinting (MRF) allows simultaneous T1 and T2 mapping in 10 seconds using 48-fold undersampled data. However, when “reverse translated” to preclinical research involving small laboratory animals, the undersampling capacity of the MRF method decreases to 8 fold because of the low SNR associated with high spatial resolution. In this study, we aim to develop a deep-learning based method to reliably quantify T1 and T2 in the mouse brain from highly undersampled MRF data, and to demonstrate its efficacy in tracking T1 and T2 variations induced by MR tracers. The proposed method employs U-Net as the backbone for spatially constrained T1 and T2 mapping. Several strategies to improve the robustness of mapping results are evaluated, including feature extraction with sliding window averaging, implementing physics-guided training objectives, and implementing data-consistency constraint to iteratively refine the inferred maps by a cascade of U-Nets. The quantification network is trained using mouse-brain MRF datasets acquired before and after Manganese (Mn2+) enhancement. Experimental results show that robust T1 and T2 mapping can be achieved from MRF data acquired in 30 s (4-fold further acceleration), by using a simple combination of sliding window averaging for feature extraction and U-Net for parametric quantification. Meanwhile, the T1 variations induced by Mn2+ in mouse brain are faithfully detected. Code is available at https://github.com/guyn-idealab/Mouse-MRF-DL/.
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16446-0_41
SharedIt: https://rdcu.be/cVRTA
Link to the code repository
https://github.com/guyn-idealab/Mouse-MRF-DL/
Link to the dataset(s)
Available on request for research purposes only
Reviews
Review #2
- Please describe the contribution of the paper
A deep learning based MRF parameter mapping method is proposed for preclinical MRF data, achieving 4-fold further acceleration compared to the baseline MRF 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 proposed deep learning based MRF method uses a sliding-window averaging to extract spatial features and combine the use of map loss and recon loss.
- The results show a 4-fold acceleration and faithful T1 mapping both before and after Mn2+ injection.
- 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.
As only T1 map is needed to detect Mn2+ tracer effect, a T1 mapping only method with much shorter acquisition time can be used instead of using MRF and complicate network and map both T1 and T2 information. In other words, simultaneous T1 and T2 mapping is not mandatory here.
- 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
A detailed description of the model and the data used is illustrated.
- 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/2022/en/REVIEWER-GUIDELINES.html
A deep learning based MRF parameter mapping method is proposed which uses a sliding-window averaging to extract spatial features and combine the use of map loss and recon loss. With the proposed method, a 4-fold further acceleration is achieved compared with the baseline MRF method. Due to CSF flow and large T1 and T2 property, T2 estimation is always not as faithful as parenchymal tissue under MRF parameter mapping framework. Besides, CSF normally is not of interests. Thus I suggest to further analyze on different brain tissue(such as WM, GM and CSF), which should better show how different methods or different acceleration factor data perform.
- 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
7
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
A deep learning based MRF parameter mapping method is proposed for preclinical MRF data, and it uses a sliding-window averaging to extract spatial features and combines the use of map loss and recon loss. It achieves a 4-fold further acceleration compared to the baseline MRF method, which should also be applicable to human MRF for acceleration.
- Number of papers in your stack
5
- What is the ranking of this paper in your review stack?
2
- 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 presents a deep learning method to compute T1 and T2 map by MR fingerprinting in Mice that have been injected with a Mn2+ tracer. The method mostly replicates blocks that have been applied in Human before, but trained on Mouse data. An ablation study was performed.
- 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.
- Significance: There is a lack of computational methods development for small animals and this paper fills a gap. It convincingly shows that DL-based MRF can be applied to Mouse data and saves acquisition time by allowing a higher undersampling factor.
- Validation: The impact of the different components of the network is correctly evaluated in an ablation study.
- Writing: The paper is well written and very easy to follow.
- 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 proposed method is not hugely novel as it mostly builds on architectures that have been proposed for MRF in Humans before. (Method novelty is not necessarily a claim by the authors here, and they correctly cite previous works).
- Most components have no discernable impact on the reconstruction score. The use of a U-Net over a template-matching approach, and the use of the moving window over a feature extractor (i.e., a single conv layer) are the only choices that seem to clearly improve things.
- In general, an issue with single dataset studies like these is that it is unknown if the proposed method requires retraining on each new dataset, meaning time and compute, or if the learnt features are general enough to be applied to any new dataset (with the same sequence). The impact of the paper would be much stronger if generalizability was investigated. Another impactful question is: can you design/train the network in such a way that any number of spiral shots or any number of time frames can be used as input, so that the most is made of the available data.
- 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 checklist states that the dataset is downloadable, but I did not see an (anonymized) link in the manuscript
- Same for code and models
- The exact architecture of the UNet (number of features per layer) is not provided. Not a problem if code is released.
- Hyper-parameter strategy is not mentioned, but maybe no hyper-parameter search was performed (which would be fine)
- Time/cost/memory not reported
- No statistical significance tests
- 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/2022/en/REVIEWER-GUIDELINES.html
- In general, I think it is a good scientific paper, with sound experiments. My main concern is the low impact it may have, due to limited novelty of the methods and narrow application. It may not be the best-suited conference for this paper.
- One of the stated objectives of the authors is to track T1 variations induced by Mn2+, and the training set includes images acquired before and after Mn2+ injection. While the authors do show T1 differences between time points, an even stronger experiment would be to train on pre-injection images and test on post-injection images (and vice versa) to see if the inclusion of both conditions in the training set reduces a bias in the fitting process; or conversely if training on control cases only generalizes to Mn2+ cases.
- It could be clarified how the data was undersampled for training: were the same two spiral arms always used? Or were two spiral arms out of 48 randomly selected?
- It is a bit surprising that the moving window does that much better than the feature extraction layer. It seems to me that the FE layer has the flexibility to learn a simple moving window. Does it mean that the FE is just more difficult to train? (t would be useful to state the kernel size – input and output features, width and height – of the FE layer)
- It’s not perfectly clear to me if C-2 contains exactly one consistency loop? (i.e. C-1 would not use the consistency module at all)
- Similarly, does the fully sampled component (FS) correspond to lambda != 0, that is the loss include loss_recon?
- There are no statistical tests, and it seems to me that no true difference exists between SW+U/SW+U+FS/C-2.
- 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 recommendation is mostly driven by the lack of technical novelty and the narrow application. The science is sound but lacks a broad impact.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
3
- 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
The manuscripts presents a deep learning-based method to generate T1 and T2 maps from undersampled MRF data acquired in mouse brain pre and post Mn2+ administration. Specifically, the developed pipeline involves sliding-window cascaded modules, a U-Net to infer T1 and T2 maps, a network to generate back MRF from the inferred T1 and T2 maps, and a data-consistency module to suppress estimated errors. The performance was evaluated by means of MAPE, PSNR and SSIM and proven to be higher with the dedicated components. The approach to solve the limitations from MRF is novel and well addressed.
- 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.
- Overall, it is a well written manuscript with supporting validation.
- The application of the modules with sliding-window and U-Net for MRF reconstruction is appealing.
- The proposed framework has been well explained.
- The method achieved an on par precision compared to the baseline method.
- The ablation study is sound and validates the improvement of each component in the proposed pipeline.
- The proposed work tackles the clinical need for robust MRF reconstruction from undersampled data.
- 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.
- Data augmentation during training: an up-down flipping may barely augment usable data, other simple data augmentation techniques could be more useful.
- Missing discussion on related works [3,4] and explanation of the lower performance of the cascaded network.
- Please rate the clarity and organization of this paper
Excellent
- 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 checklist mentions the code will not be available in a repository as it is still under development, which limits 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://conferences.miccai.org/2022/en/REVIEWER-GUIDELINES.html
- In the abstract, I would suggest mentioning only the approach that yielded better results, to avoid confusion.
- The implementation details should be described, i.e., training time, test time, etc.
- Discussing the related works [3, 4, 5] would help to highlight the value of the contribution.
- SW+U+FS was expected to yield better results, with the current tone of the manuscript, but it did not. Rephrasing or further discussion would clarify this.
- Data augmentation approach could have been more exploited. The works [3, 4, 5] have not employed data augmentation.
- 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
7
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The presented work aims to solve the existing problem of MRF acquisitions with a well explained framework achieving promising results.
- Number of papers in your stack
5
- What is the ranking of this paper in your review stack?
1
- 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
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 deep-learning-based MRF parameter mapping method for preclinical MRF data. The main novelty is the incorporation of spatial correlations into the network to improve prediction accuracy. Overall the majority of the reviewers were very positive about the paper and indicate only minor issues, such as - why not use only t1-mapping rather than MRF for this specific application, while one reviewer commented on limited novelty and narrow application. Based on the overall scores and reviewers’ comments the paper is suitable for presentation at MICCAI.
- 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). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
2
Author Feedback
We thank the reviewers and the meta-reviewer for their overall positive comments and constructive suggestions. We believe that all reviewers understand our method correctly and their suggestions are to-the-point. We agree with reviewer 2 that the novelty is not a highlight of this deep learning method and our main contribution is to facilitate the application of MRF method to the understudied preclinical MRI research. (1) Justification on motivation (meta, R1) We demonstrated the efficacy of proposed MRF method, although preliminarily, to track only the T1 variation brought forth by MR tracer. This work serves as our first step towards the ultimate goal of simultaneously imaging two MR tracers by simultaneously tracking the T1 and T2 variation enabled by MRF [1], which will facilitate studying the glymphatic system [2] of mouse brain by two MR tracers. We will gradually improve our method so that it will reliably estimate T2 reduction in mouse brain caused by a T2-tracer (e.g., 17O-labeled water) as T2 variation can be significantly smaller compared to T1, and simultaneously induced T1- and T2-tracers. (2) Generalizability of proposed method (meta, R2) We agree that a deep-learning framework generalized enough to handle data with different undersampling ratio, or T1 and T2 from a randomly large range, can be intriguing. However, current framework design and training procedure limited the capability of the method to have a parametric quantification module that can adapt to the variation of undersampling ratio, and T1-T2 range. We will think about the issue of generalizability and try to improve our method in the future. (3) We will provide more implementation details of proposed method, clarify and further explain the results of our ablation study (actually reviewers’ interpretation of our results are correct). A github link to current network implementation will be provided in the updated paper. However, the MRF dataset is private to an institutional preclinical MR imaging center. The data can be available on request, but only for research purpose.
[1] Anderson CE et al., Dynamic, Simultaneous Concentration Mapping of Multiple MRI Contrast Agents with Dual Contrast - Magnetic Resonance Fingerprinting. Sci Rep 9:1–11 (2019). [2] J. J. Iliff et al., Brain-wide pathway for waste clearance captured by contrast-enhanced MRI. J. Clin. Invest., vol. 123, no. 3, pp. 1299–1309, (2013).