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

Authors

Yuhang Sun, Dongming Wei, Zhiming Cui, Yujia Zhou, Caiwen Jiang, Jiameng Liu, Qianjin Feng, Dinggang Shen

Abstract

Motion correction is a fundamental preprocessing step for liver dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI), which can be used for the diagnosis of benign and malignant tumors. Previous studies have difficulty in aligning small structures, e.g., tumors and vessels, due to the remarkable intensity changes over different images. Except for measuring physiologic parameters in DCE-MRI, the time-intensity curves (TICs) can also be used to constrain the alignment of small anatomical structures such as small tumors and vessels. In this work, we propose a coarse-to-fine motion correction scheme with smoothness constraint of TICs to correct the motion in liver DCE-MRI. Specifically, the proposed motion correction scheme consists of two major stages. First, different time point images are registered to the selected fixed image pairwisely via a fully convolutional network (FCN), which outputs their corresponding coarse displacement vector fields (DVFs). Second, all of the coarse DVFs are further refined jointly under the group similarity of the warped time points and the fixed image, together with the smoothness constraint of TICs at a fine level. To our knowledge, our work is the first in constraining the motion correction using TICs for better alignment of small structures. Experimental results on liver DCE-MRI demonstrate that our proposed method can obtain a more accurate alignment of small structures (e.g., tumors and vessels) than state-of-the-art methods.

Link to paper

DOI: https://doi.org/10.1007/978-3-030-87234-2_69

SharedIt: https://rdcu.be/cyl9f

Link to the code repository

N/A

Link to the dataset(s)

N/A

Reviews

Review #1

Please describe the contribution of the paper

The manuscript uses a method which combines pairwise and groupwise image registration methods for motion correction in liver DCE-MRI. Experiments include comparisons with other methods for pair-wise registration.
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 simplicity of the presented method is appealing - an extension of the well-known voxelmorph method for groupwise image registration.
- The method performs well on the given problem
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 novelty of the presented method, and the application, is limited
- the experiments lack comparisons with other groupwise registration methods (there are several to choose from - elastix for example). Ideally, a comparison with images collected by gating would be included, but that is probably a bit too high expectations for a conference paper (even in MICCAI)
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

Seems OK, but I advice the authors to share the 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://miccai2021.org/en/REVIEWER-GUIDELINES.html

Fig 1: Please avoid using abbreviations in captions (to ensure that they can be read stand-alone) Fig 1: Which interpolation method is used? There seems to be an over-shoot issue. Page 4, last paragraph: How is the mean image computed? Is it computed on a residual deformation field (without an affine contribution for example)? Please clarify and motivate your choice. Eq 3, and the paragraph below: The formal use of the time parameter needs to be improved. Please specify the values t takes in eq. 3, instead of a buit hand-waiving in the paragraph below. Page 5, last paragraph: How are the hyperparameters set? I assume that they are set based on the data used for the evaluation. How does this dependency affect the result, and how the method will generalize to other data? Please add in the discussion. page 6, first sentence: “for having” shouls be rephrased. page 6, first paragraph: The dataset is partitioned in data for training and for testing. In addition, three-fold cross-validation is performed. How is the test data used? Which data is used to generate the values given in the results-section (test or cross-validation)? page 8, middle paragraph: Explain TRE, or give a reference. page 8, middle paragraph: 0.82% -> 0.82 percentage points
Please state your overall opinion of the paper

borderline accept (6)
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
- The simplicity of the presented method is appealing - an extension of the well-known voxelmorph method for groupwise image registration.
- The method performs well on the given problem
- The novelty of the presented method, and the application, is limited
- the experiments lack comparisons with other groupwise registration methods (there are several to choose from - elastix for example). Ideally, a comparison with images collected by gating would be included, but that is probably a bit too high expectations for a conference paper (even in MICCAI)
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

Review #2

Please describe the contribution of the paper

The authors proposed a two-step (coarse-to-fine) method for motion correction for liver DCE-MRI. The first step utilizes a CNN to generate coarse DVFs and the corresponding warped images in a pair-wise manner. The loss function is regualrized by a smoothness term of the DVFs. The second step takes as input a group of warped images from the previous step and produces the fine DVFs and the final warped images. The loss function is regularized by the smoothness of the time-intensity curves (TICs) and the smoothness of the DVFs.
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 paper is well written and easy to follow.
2. The main contribution of the paper - utilizing the smoothness of the TICs as a constraint in the loss function, has been validated.
3. The alignment of subtle structures appears to be improved thanks to the proposed TIC based constraint term in the loss function.
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.

Even though the contribution of the proposed method as a whole has been validated (Figure 3,4, Table 1), some components of the methods, such as the two-step learning scheme, the smoothness of the DVFs as a regularization term are not fully evaluated.
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
1. The description and illustration of the proposed method is by and large clear, except for some minor details.
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
Overall I think this is a solid paper. The authors have evaluated the TICs constrained loss function (Fig.3, 4. Table 1), which I think is good. My main critic is that some of the components of the proposed methods are not fully evaluated (I understand this could be due to the page limit). For example, the authors proposed to predict the final DVFs in two steps (coarse-to-fine). How much could the two-stage framework improve the results compared to a single step method using only one CNN (to compare the warped images from the CNN in the top panel of Fig.2 with the final warped images.)? For the DVFs based smoothness regularization term in eqn 1, is it adapted from a previous publication or proposed by the authors. If the latter, it could also be evaluated in a similar way to the TICs constrained loss function.

It is a particularly nice motivation to increase the alignment accuracy of small structures such as vessels. However, the results supporting this claimed contribution (‘These results imply that our LTICs could handle subtle structure alignment better’) of the TICs based constraint is not strong enough. The authors used Fig.4 as the supporting evidence. However, in my opinion this could be more highlighted e.g., provide more results and highlight the area of interest to be examined.

some minor issues:
1. A mathematical definition or reference to the ‘L_smoothness’ in eqn 1 and 4 seems missing. Is the regularization terms adapted from a previous publication (a reference) or proposed by the author (a definition)?
2. The authors should give a (mathematical) definition to the ‘o’ symbol in eqn 5.
  ‘The ﬁnal motion correction results… are obtained by composing the DVFs of the ﬁrst stage and the second stage’ Here the ‘composing’ is rather confusing. A clear definition should be given.
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?

Overall the claimed the contribution of the proposed TIC based constraint term has been evaluated and the results appear to have validated the claimed contribution.
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

Review #3

Please describe the contribution of the paper

The paper proposes a coarse-to-fine motion correction scheme with smoothness constraint of TICs to correct the motion in liver DCE-MR.
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.

-well written. -easy to read -sufficient validation
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.

-Quantitative improvements in Table 1 seem marginal -marginal novelty
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

Not clear if the work can can be reproduced.
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. See “Analysis of dynamic MR breast images using a model of contrast enhancement” by Paul Hayton, et al.
2. Page 3, last paragraph: You mentioned that the 11th frame has the smallest intensity changes. How was this measured? Did you compare the intensities against a) every frame or just adjacent frames and b) over what domain (ROI or entire image domain)?
3. Page 4, 3rd to last paragraph: more than one images - > several/multiple images?
4. Page 4, last paragraph. Are there n=22 groupwise means images?
5. Page 5, On choosing I_f=I_{11}: Can you comment on whether choosing a specific image (affected by a very specific phase of the respiratory/cardiac motion cycle) introduces any bias to the registration results and whether or not the possible introduction of such a bias matters in this specific application.
6. Page 5, last paragraph, On normalizing intensity values: DCE analysis is highly dependent on the intensity changes within an ROI. Wouldn’t normalizing the intensity values in every frame defeat the purpose of using TICs?
7. Page 8, under Quantitative evaluation: How are the important anatomies defined/selected?
8. Comparison of proposed method (without L_{TICs}) vs pairwise voxel morph for tumor not provided
Please state your overall opinion of the paper

borderline accept (6)
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

Sufficient validation, straightforward and clear approach
What is the ranking of this paper in your review stack?

3
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 paper under consideration introduces a hybrid two-step method that combines pairwise and groupwise image registration methods for motion correction in liver DCE-MRI. A smoothness constraint is used based on the time-intensity curves (TICs) and the smoothness of the displacement vector fields (DVFs) of the coarse stage. The paper is easy to follow and is well-written. The idea is interesting for an important application and the experimental validation is sufficient. However, there some points need to be accounted for which are listed below as well as by the reviewers. The results in Table 1 show enhancement of the Dice and TRE metrics, however, without statistical analysis, it will be very hard to assess the significance. Although Table 1 document the comparison with other methods, the paper would be enriched by its comparison against other groupwise registration methods. Also, how were the fiducial markers chosen to calculate the TRE? How many are they? Etc. What about the coarse CNN registration, how are the hyperparameters are set or optimized? Empirically or experimentally? Please discuss Another point that I agree with R3 is the claim to “increase the alignment accuracy of small structures such as vessels” by the authors. The results in Fig. 4 does not a support that claim very well. This should be highlighted with more qualitative or quantitative results for the area of interest to be examined
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).

5

Author Feedback

Q1: Statistical analysis of the quantitative results Thanks for the valuable comments of the AC and reviewers. In the submitted paper, we did not provide the statistical analysis of our results. We have now provided as below. (1) For the Dice of liver, the difference between our groupwise registration strategy (without smoothness of TICs constraint) and single pairwise registration strategy is significant (p=0.04). And the difference between the methods with and without L_TICs regularization is significant (p=0.04). (2) For the Dice of tumor, the difference between our method (with L_TICs) and the single pairwise registration method is significant (with p=0.04). (3) For the TRE of vessel, the difference between our groupwise registration strategy (without smoothness of TICs constraint) and single pairwise registration strategy is significant (with p=0.03). And the difference between the methods with and without L_TICs regularization is also significant (with p=0.03).

Q2: Comparison of the other groupwise registration methods We only compare (1) the pairwise registration strategy and the groupwise registration strategy and (2) methods with and without L_TICs due to space limitation. As shown in the results, our groupwise registration strategy could lead to improved performance with regard to the single pairwise registration strategy. In the reference [15] of the submitted paper, groupwise strategy has been developed in many different kinds of forms for the specific different registration tasks. And the groupwise registration in our method may be considered as a general registration step in all groupwise registration strategies. In addition, we also intend to design a task-specific groupwise registration strategy in the future as stated in the Conclusion section. As suggested, we have now performed a groupwise method of Elastix, which yields Dice of liver 92.44(0.51)%, Dice of tumor 78.54(3.07)%, and TRE of vessel 2.45(0.46)mm. These results suggest that our groupwise strategy has comparable performance with Elastix.

Q3: The selection of anatomical corresponding markers when TRE was calculated In the process of evaluation, we invited one clinical expert to define the anatomical corresponding markers of the important regions (such as the bifurcation point of vessels). For each subject to be evaluated, 2-4 markers (with obvious corresponding markers across all the time points) were selected. To calculate the TRE of one marker for each subject, the distance was calculated between all markers on the warped images (obtained by the DVFs) and the marker on the fixed image, and then the mean distance was calculated. Finally, we obtained mean and standard of the TRE of all the markers as demonstrated in Table 1.

Q4: The hyper-parameters of the coarse registration stage In the experiments of coarse registration stage, we set values of hyper-parameters experimentally. (1) The batch size of 6 is set according to the GPU memory. (2) The learning rate is set to 1e-4, which is chosen on the basis of the convergence performance during the training. (3) We trained 15000 iterations in the coarse stage according to the loss curve as well as the performance of testing subjects. (4) As for the weight of regularization in Eq. 1, we set the hyper-parameter as 1 after trying to set this parameter as 0.1, 1, 10, 100 and finally selecting the hyper-parameter value with the best visual registration results. Based on the selected value, we obtained reasonable results for the coarse registration stage.

Q5: Results of Fig.4 In Fig. 4, we thought we obtained visually-improved results in two small vessels. The first vessel was in the right of the tumor, which was better restored as it was longer and more continuous. Another vessel is located in the upper-right of the tumor, and it is also longer and more continuous. Indeed, the vessels in this patient are not obvious. We will replace Fig. 4 with more obvious visual comparison in the final paper.

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 authors’ rebuttal answered most of the concerns. The work is interesting and attractive to the MICCA readership and the authors rebuttal to make modification(s) and/or addition(s) is(are) feasible for the camera ready.
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).

3

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 authors propose to improve motion correction in DCE-MRI using two deformable registration steps, in the second of which, time-intensity curves are constrained to be smooth. As two reviewers also mention, the novelty in the proposed method is quite limited. Approaches to register time-point images with simultaneous spatial and temporal smoothness constraints have been around, although I am not sure about their application in DCE-MRI. The authors’ rebuttal does not quite address the novelty issue.
After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.

Reject
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).

17

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.

The reviewers commented on the simplicity of the presented method - an extension of the well-known voxelmorph method for groupwise image registration. The paper is also found to be clear and well written. Authors have properly addressed the remaining concerns including hyperparameter setting and benchmarking.
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).

2

back to top

Motion Correction for Liver DCE-MRI with Time-Intensity Curve Constraint