List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Juyeon Heo, Pingfan Song, Weiyang Liu, Adrian Weller
Abstract
Magnetic Resonance Fingerprinting (MRF) is a promising approach for fast Quantitative Magnetic Resonance Imaging (QMRI). However, existing MRF methods suffer from slow imaging speeds and poor generalization performance on radio frequency pulse sequences gen- erated in various scenarios. To address these issues, we propose a novel MRI physics-informed regularization for MRF. The proposed approach adopts a supervised encoder-decoder framework, where the encoder per- forms the main task, i.e. predicting the target tissue properties from input magnetic responses, and the decoder servers as a regularization via reconstructing the inputs from the estimated tissue properties us- ing a Bloch-equation based MRF physics model. The physics-based de- coder improves the generalization performance and uniform stability by a considerable margin in practical out-of-distribution settings. Exten- sive experiments verified the effectiveness of the proposed approach and achieved state-of-the-art performance on tissue property estimation.
Link to paper
DOI: https://doi.org/10.1007/978-3-031-43895-0_42
SharedIt: https://rdcu.be/dnwyV
Link to the code repository
https://github.com/rmrisforbidden/CauMedical.git
Link to the dataset(s)
https://github.com/rmrisforbidden/CauMedical.git
Reviews
Review #2
- Please describe the contribution of the paper
The paper proposes a novel approach to Magnetic Resonance Fingerprinting (MRF) that addresses issues related to slow imaging speeds and poor generalization performance on radio frequency pulse sequences generated in various scenarios. The proposed approach adopts a supervised encoder-decoder framework, where the encoder performs the main task of predicting the target tissue properties from input magnetic responses, and the decoder serves as a regularization via reconstructing the inputs from the estimated tissue properties using a Bloch-equation based MRF physics model. The paper presents extensive experiments that verify the effectiveness of the proposed approach and achieve state-of-the-art performance on tissue property estimation.
- 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 approach introduces a physics-based decoder that acts as a strong regularizer, providing informative feedback and contributing to the training of a better encoder. This leads to improved generalization performance and uniform stability in practical out-of-distribution settings.
- 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.
In general, this manuscript is interesting. However, it also has some issues, which are commented on below.
- The paper does not provide a detailed explanation of the choice of the three-layer fully connected neural network used in the encoder, which may need an ablation study.
- In Fig.3, are the T1 and T2 gold standard image compare to some other images as the prediction error or the original image? If they are the original image, then it should use gray images rather than the color images.
- in Table2, what’s the unit of the numbers?
- In FIg.4, even though the BolchNet error is significantly lower than previous methods, the significance of the improvement is still unclear. Will such method can be directly used to replace the origianl method even through the predicted value looks still have lots of error ?
- 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
The architecture is clear and the formulation is well defined.
- 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
N/A
- 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 decision is based on the lack of the clarify of the paper.
- Reviewer confidence
Somewhat confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
5
- [Post rebuttal] Please justify your decision
After read through author’s response and other reviewer’s comments, I’m inclined to accept this paper due to it’s novelty and experiments details.
Review #4
- Please describe the contribution of the paper
The paper approaches the problem of Magnetic Resonance Fingerprinting (MRF) by the proposed physics-informed regularization. The key contribution of proposed method is the Bloch decode, which utilizes the Bloch equations-based MRI physics model to reconstruct the input responses from the estimated tissue properties and consequently regularizes the training of the encoder to produce generalizable tissue properties. Author claims that the proposed BlochNet can achieve better generalization performance across different data.
- 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 topic is interesting and clinically significant.
-
The paper is well organized and easy to follow.
-
The idea of leveraging Bloch equations as regularization seems effective and interesting.
-
- 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.
-
There is no mention of ablation strategy that would have guided the selection of many hyperparameters. Yet it seems that no proper tuning opportunity has been given to the proposed method. It appears that the selection was purely heuristic.
-
Some implementation details are missing. (e.g., batch size, learning rate, the maximum epochs, etc.)
-
- 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
Authors mentioned that the code will be made available upon acceptance.
- 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
-
Authors may compare the relevant resource use of different methods. Ideally, just performance gain is not enough for real-world ML. To show the practicability of the proposed method in the real-world setting, authors may provide more details about resource usage.
-
In term of the experiment results, the proposed method shows an obvious improvement on Anatomical data (Table 1), but it seems looks like improvement becomes smaller on Phantom data (Table 1). In general, it would be nice to see the standard deviations of performance metrics across the test subjects in additional to their mean, and ideally when performance differences from baselines are claimed they should be backed up by a significance test. Authors may conduct statistical significance test.
-
- 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 rating reflects innovation of the approach MRF and presented good performance. The merits outweigh the weaknesses.
- 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
5
- [Post rebuttal] Please justify your decision
The authors’ rebuttal address my questions I would like to keep my rating.
Review #5
- Please describe the contribution of the paper
The proposed for estimating quantitative tissue properties from magnetic responses to pseudo-random pulse sequences model consists of two modules (a) the encoder is a fully-connected neural network that takes as input multiple flattened MR images (each acquired with different acquisition parameters), and outputs the T1 and T2 quantitative images, (b) The decoder is a “Bloch equation solver” that takes as inputs the predicted T1 and T2 maps, and the acq. parameters corresponding to the encoder’s inputs, and outputs the reconstructed MRIs by solving the resulting Bloch equations. The encoder serves as the MRF module and is trained using L2 losses on the predicted quantitative maps, as well as on the recon. images from the Bloch eqn solver. The latter loss term acts as regularization, enhancing generalization performance. Results demonstrate that the proposed regularization improves predictions slightly for in-distribution, and substantially for out-of-distribution pulse sequences.
- 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.
-
Incorporating domain knowledge is often crucial for obtaining good generalization performance from neural networks, especially in settings of small annotated datasets, as is often the case in medical image analysis. In this context, using a Bloch equation solver to regularize the MRF training is an interesting idea.
-
The performance of the proposed method on unseen test pulse sequences is substantially better than other methods.
-
Writing and organization in the paper are very good. Thanks!
-
- 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.
-
Implementation novelty of ‘efficient Bloch equation solver’ may be overstated. There appear to be other GPU implementations on the web.
-
Comparison with an important related work (Scholand et al ISMRM 2020) is missing (see full citation in point 1 of detailed comments below).
-
Some details of the experimental setup are missing: (a) training / test splits, (b) values of hyper-parameters L, p, (c) no statistical significance tests of results are carried out.
-
- 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 code will be made available upon acceptance.
- 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
-
Highly related work that is not cited or compared with: Scholand, Nick, et al. “Generic quantitative MRI using model-based reconstruction with the Bloch equations.” Proc. Intl. Soc. Mag. Reson. Med. Vol. 28, 2020 Scholand, Nick, et al. “Quantitative MRI by nonlinear inversion of the Bloch equations.” Magnetic Resonance in Medicine, 2023. This seems to be effectively using a different regularization than the proposed approach. It would be useful to see the comparative performance of the two regularizations. -
Regarding the EPG solver implementation: (a) Details of the implementation are missing - what type of optimization is used (e.g., gradient descent?), what are the “three bloch stages”? What does the ‘torch jit package’ do? Section 3.2 promises that the Appendix contains more information about the EPG solver. Which information is this referring to? (b) Please explain the procedure of [27], at least at a high level. Is the main contribution in this regard that the authors implement [27]’s procedure on a GPU instead of a CPU? Please state the differences wrt [27] clearly. (c) Further, it appears that an auto-differentiable, GPU implementation of EPG is available publicly: https://somnathrakshit.github.io/projects/project-mri-sim-py-epg/3754.html, https://github.com/utcsilab/mri-sim-py/tree/master/epg. Please cite / acknowledge this. Does the proposed implementation offer any further advantages than this implementation? If there are no significant differences, please tone down novelty claims in this regard.
-
It is said that the length of response sequences, L, is greater than p. What is p? What are the values of L and p in the experiments?
-
Please mention the evaluation metric in the captions of tables 1 and 2.
-
The authors might want to avoid using “anti-causal mechanism”, a jargon that bears little connection to the content of the paper.
-
Please improve clarity of the explanation of the Bloch equations in the appendix. In particular, it is hard to follow the paragraph “The Bloch equations represent a non-linear mapping from…”. This space can be used to better describe how exactly one would solve this equation to get the reconstructions $\hat{X}_n$ from predicted tissue properties $\hat{\Theta}_n$ and acquisition parameters $\phi$?
- Please improve the clarity of figure 1 in the appendix: in the current version, it is hard to read the axis labels and ticks.
-
- 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?
Pluses (good idea to incorporate domain knowledge for regularizing data-driven learning + clear writing) outweigh minuses (missing comparison with Scholand’s work), in my opinion. Having said that, I think it will be important to include missing citations, and tone down novelty claims regarding the EPG solver implementation, and better explain the EPG solver itself.
- Reviewer confidence
Somewhat confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
7
- [Post rebuttal] Please justify your decision
My two main questions (position of the proposed method with respect to Scholand’s work, and the novelty of the EPG implementation) have been answered in the rebuttal. I am happy to improve my rating of the paper. I encourage author’s to include the discussion in the paper.
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 review comments are mixed. The authors may answer the reviewers’ questions during the rebuttal.
Author Feedback
We sincerely thank the reviewers for the constructive comments. We take all comments seriously.
*Experiments details and ablation study (R. #2 6-1,3 #4 6-1,2 9-1 #5 6-3 9-3,4) We utilized synthetic data for training, and Phantom/Anatomical for evaluation. RF pulses matched in Table 1 but differed in Table 2. Our setup included Adam optimizer (lr=1e-3), 3-layer encoder-decoder with varying hidden units, and maximum epochs of 100 with early stopping based on the validation set on GTX 1080 Ti. p denotes the number of tissue properties, in our case p is 2 and L is 1000. As mentioned in section 4.2, the evaluation metric is the mean squared error in log-scale. We clarify that a series of ablation studies were conducted via comparison with other methods that use different types of decoders, please refer to Table 1 and Table 2. For example, comparing BlochNet (using a physics-based decoder) with FC (using no decoder) and FC-FC (using a learned decoder) is an ablation study on the effect of the encoder and the effect of MRI physics. The pioneering work by Cohen et al proposed to use a three-layer fully-connected neural network (FC). Since we wanted to compare directly against this important work, we chose the same architecture and parameters for the encoder. We will add the full details to the revision and release our code.
- More citations (R. #5 6-2 9-1) Both the suggested literature(ISMRM 2020 and its extended full version MRM 2023) and our work propose to involve Bloch equations into the quantitative MRI reconstruction problem. The main difference is that the suggested literature estimates tissue parameters directly from k-space data, following a one-step process; while our approach estimates tissue parameters from corresponding magnetic responses derived from k-space data, following a typical two-step process. It would be interesting and meaningful to compare these two regimes. Additionally, our method aims for generalizable and robust MRF to out-of-distribution(OOD) settings by exploiting a physics-based decoder, while prior works focus on a fast and generic model-based QMRI. Due to the conference guidelines prohibiting the incorporation of new results, we will explore this comparative analysis in the revision.
*Significance tests (R. #4 9-2 #5 6-3) We have included standard deviation on 10 trials. We also present statistical significance tests on 10 trials for pairwise group comparisons using Tukey HSD test after the Normality Test and repeated ANOVA. For results in Table 1, our method is not always statistically better to every compared method. This is the expected result since estimating tissue parameters in in-distribution settings is an easy problem where all baseline methods can do well. However, for results in Table 2 (more challenging setting), our method is always statistically better (p-value < 0.001) to every compared method. This highlights the robustness and generalizability of our method in the OOD setting.
*Significance of our improvements (R. #2 6-4) Fig 4 is a very challenging OOD case, as the RF pulses used in testing are different from those used in the training stage which can lead to different magnetic responses. As far as we are concerned, no existing approaches achieved satisfactory results. While there is still ample room for improvement across all the methods, our approach took a critical step forward by incorporating physics knowledge.
*Our contribution of EPG implementation (R. #5 6-1 9-2) While there are other GPU implementations of EPGs available, our version improves efficiency by exploiting the Pytorch jit package for efficient parallelization, batch-wise computation for the 3 Bloch stages(same as [27], nutation+forced precession, rotation, and relaxation in Fourier domain), and handling complex values in Pytorch efficiently. This makes repeated EPG computations feasible during training. We plan to conduct more analysis and comparison among different implementations and tone down our claim.
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.
After rebuttal, the reviewers make agreement to accept this paper.
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 paper presents a physics informed approach for the regularization of the Magnetic Resonance Fingerprinting reconstruction problem. The paper got mixed reviews. Authors addressed some in their rebuttal. Overall interesting idea that can be relevant for the MICCAI community.
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.
Based on the reviews and the rebuttal it looks like the paper need a bit more polishing and work.