List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Zhenwei Wang, Yifan Lv, Mengshen He, Enjie Ge, Ning Qiang, Bao Ge
Abstract
Fiber tract segmentation is a prerequisite for the tract-based statistical analysis and plays a crucial role in understanding brain structure and function. The previous researches mainly consist of two steps: defining and computing the similarity features of fibers, and then adopting machine learning algorithm for clustering or classification. Among them, how to define similarity is the basic premise and assumption of the whole method, and determines its potential reliability and application. The similarity features defined by previous studies ranged from geometric to anatomical, and then to functional characteristics, accordingly, the resulting fiber tracts seem more and more meaningful, while their reliability declined. Therefore, here we still adopt geometric feature for fiber tract segmentation, and put forward a novel descriptor (FiberGeoMap) for representing fiber’s geometric feature, which can depict effectively the shape and position of fiber, and can be inputted into our revised Transformer encoder network, called as FiberGeoMap Learner, which can well fully leverage the fiber’s features. Experimental results showed that the FiberGeoMap combined with FiberGeoMap Learner can effectively express fiber’s geometric features, and can differentiate the various fiber tracts, furthermore, the common fiber tracts among individuals can be identified by this method, thus avoiding additional image registration. The comparative experiments demonstrated that the proposed method had better performance than the existing methods. The code is openly available at https://github.com/Garand0o0/FiberTractSegmentation.
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16431-6_14
SharedIt: https://rdcu.be/cVD4V
Link to the code repository
https://github.com/Garand0o0/FiberTractSegmentation
Link to the dataset(s)
N/A
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a fiber classification method to identify fiber bundles from the whole brain tractography. The proposed method uses a new descriptor to represent the shape and position information of the fiber bundles, which are combined and then fed to the network for training a transformer-based deep learning model. Experiments have been performed to investigate the effectiveness of FiberGeoMap. In addition, the proposed method has been applied to autism 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 two main advantages of this paper are as follows:
- A new method of fiber information description is proposed. This method, which is based on spherical coordinates, computes local and global information separately and combines two kinds of information together as a whole sent to the subsequent deep learning model.
- The transformer module is used to resolve the fiber segmentation problem with FiberGeoMap as input for the transformer.
- 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.
Experiments seem unfair, such as inconsistent training and testing data. The baseline methods are proposed in 2018, not the latest ones. A qualitative analysis should be also provided.
- 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
Good.
- 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 experimental part, the proposed method completes the segmentation of 103 fiber bundles in total, resulting in an average dice score of 0.93. Some of the fiber bundles segmented by the proposed method are not included in the other two comparison methods, therefore it is unreasonable to directly compare the average dice of the other two methods with the average dice of 103 fiber bundles. The evaluation should be improved by comparing the dice scores of fiber bundles belonging to the results of all methods.
-
Are the models trained and tested on the same sets of HCP data? Taking one of the baseline methods, TractSeg, as an example, TractSeg uses 105 subjects for training and testing, and used five-fold cross-validation, which is different from the setting of the proposed method, i.e., 205 subjects. The paper must include more experimental details to ensure fair comparisons of different methods.
-
The two comparison methods in the experimental part were both proposed in 2018. There are new methods proposed in this field in recent years, such as DeepWMA (proposed in 2020). These new methods should be considered in the evaluation.
-
Why is there only quantitative analysis in the experimental part? It will be more convincing if the qualitative analysis can be included in the experimental part.
-
The resolution of Fig. 6 should be increased. Currently, if you zoom in to see the performance of individual fiber bundles, they are blurry and hard to see.
-
Evaluation with only one dataset is not very sufficient to demonstrate the effectiveness of the proposed method. More datasets should be considered.
-
- 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?
Please refer to the comments in box 5.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
2
- 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 reading the authors’s rebuttal and other reviewers’ comments, I would like to raise my rating.
Review #2
- Please describe the contribution of the paper
The authors propose a tractography segmentation method. This is a nicely designed method and shows a good result.
- 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 is a nicely designed method and shows a good result.
- 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.
some major concern about the paper.
-
The experimental evaluation seems not fair to either TractSeg or WMA. The author combined the two atlases together. The two methods have different definitions even for the same tract. This is a known issue due to the lack of concuss. Please refers to https://doi.org/10.1016/j.neuroimage.2021.118502. So when comparing the tracts that are overlapped in these two atlases, there should be bias introduced. Also the two methods are performed with different tractography algorithms.
-
Second, the computation of the evolutions is not clearly described. Dice is for volumetric overlap, while prevision and recall are class prediction. Did the authors convert the streamlines to masks somehow?
-
The application on autism seems not necessary and redundant. Is “the proportion of fibers” a fair measure for “abnormities”? Do we expect ASD individuals have such abnormities? As this is a technical paper, I would suggest adding experiments on additional datasets from multiple acquisitors as performed in TractSeg and WMA studies.
-
- 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
good
- 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
The authors propose a tractography segmentation method. This is a nicely designed method and shows a good result. Below are some major concern about the paper.
-
The experimental evaluation seems not fair to either TractSeg or WMA. The author combined the two atlases together. The two methods have different definitions even for the same tract. This is a known issue due to the lack of concuss. Please refers to https://doi.org/10.1016/j.neuroimage.2021.118502. So when comparing the tracts that are overlapped in these two atlases, there should be bias introduced. Also the two methods are performed with different tractography algorithms.
-
Second, the computation of the evolutions is not clearly described. Dice is for volumetric overlap, while prevision and recall are class prediction. Did the authors convert the streamlines to masks somehow?
-
The application on autism seems not necessary and redundant. Is “the proportion of fibers” a fair measure for “abnormities”? Do we expect ASD individuals have such abnormities? As this is a technical paper, I would suggest adding experiments on additional datasets from multiple acquisitors as performed in TractSeg and WMA studies.
-
- 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?
This is a nicely designed method and shows a good result.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
3
- Reviewer confidence
Very confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
4
- [Post rebuttal] Please justify your decision
After reviewing the authors’ rebuttal, I am not convinced by their answers to my previous questions. About the comparison, the authors seem to claim that combing the two atlases brings unfairness to their proposed method, the other way around that the reviewers considered. If so, why not testing the proposed method on each atlas individually. Anyway, combining the two atlases allows the authors accessing to more data with a larger number of tracts. Not sure if there are potential biases. In addition, regarding my concern about the tract definition. Let’s take a simple situation, you have a new dataset that you want to identify the left AF. TractSeg gives AF_1, WMA gives AF_2 , and the proposed method gives AF_3. Then, how do you measure the accuracy of AF_3, and in a real application, how would a user to choose across these three tracts? From the neuroanatomy perspective, I don’t think it is reasonable to combine two definitions together when there is a disagreement. I understand the authors tried to use as much data as possible. But I am not sure if that is a good way. For other questions, the authors respond that one of the reasons not comparing DeepWMA is its output combines left and right together, while it should be a very easy task to separate if needed. And my concern about the application to study autism stands, still not convinced if group differences should be analyzed (maybe for a journal where you have spaces to expand your results and discussion); rather, it might be a more valuable experiment to show the method can be applied to a variety of data, but this is minor. Based on these, I am sorry that I need to lower my score for this submission, but I still believe the technical innovation of the paper is promising.
Review #3
- Please describe the contribution of the paper
A novel fiber tract segmentation approach is proposed. First a novel descriptor, called FiberGeoMap, encodes the geometric features of the fiber by transforming to spherical coordinates and binning to predefined bins. Second, a computational framework, based on Transformers and multi-head self attention, is proposed for segmentation. Extensive evaluation using 103 fiber tracts plus no-tract categories on 205 HCP subjects are shown. Comparison with state of the art approaches shows significant improvement in dice scores over Tract-Seg and WMA. In addition, the application of proposed method in a clinical setting shows reduced fiber density (ratios) in Autism versus controls.
- 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 novel geometric descriptor for fiber tracts capturing global and local information.
- A novel Transformer and self attention based segmentation framework.
- Extensive validation, including ablation experiments.
- Evaluation using 205 HCP subjects and comparison with state of the art approaches, where proposed method significantly outperforms state of the art methods, namely, Tract-Seg and WMA.
- Application of the proposed tract segmentation method to clinical setting (Autism versus controls).
- The concepts, results, are well illustrated.
- Code and video for tracts shared via GitHub.
- 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.
See some minor issues in detailed comments.
- 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 code and video for predicted fiber tracts openly available via GitHub.
- 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
Congratulations on this great work. It was easy to read and follow. Please consider some minor suggestions, mainly pertaining to some complex sentences. Minor:
- Please check for complex sentences, and break into multiple sentences to improve readability. See examples below.
- Complex sentence in the Abstract: “Experimental results showed that the FiberGeoMap combined with FiberGeoMap Learner can effectively express fiber’s geometric features, and differentiate the 103 various fiber tracts, furthermore, the common fiber tracts across individuals can be identified by this method, thus avoiding additional image registration in preprocessing.”
- Not clear, what is implied. Please check the following sentence on page 2:”therefore, these fibers need to be clustered or segmented into a relatively small number of fiber tracts, the fibers within each tract are similar and each fiber tract should have the relatively independent meaning, namely fiber tract segmentation or fiber clustering.”
- Complex sentence on page 2: “Previous studies have focused on three cate-gories of features, which are geometrical [4,5], anatomical [6,7] and functional [8,9] features in chronological order, and seem more and more reasonable and in line with the requirement of fiber clustering, but the uncertainty also increased in sequence, for example, anatomical feature based method depend on anatomical segmentation and registration, while the anatomical atlases are various, registration techniques are also not mature.”
- Complex sentence on page 2: “Accordingly, for FiberGeoMap, we proposed a revised Transformer network, called as FiberGeoMap Learner, which can efficiently explore the FiberGe-oMap features, and then we trained the model with the all fibers from 205 HCP sub-jects [12], the experimental results showed that our method can obtain the accurate and corresponding fiber tract segmentation across individuals.”
- Please define the names of the tracts before the first use of acronyms on page 4, and Figure 3. “3 fiber tracts (STR_right, STR_left, and MCP) from….”
- Typo on page 6: “The hyper-parameters of mainly include:”, remove of or modify sentence.
- Complex sentence on page 6: “In order to quantitatively demonstrate the results, we used dice score [14], accuracy, precision and recall as the evaluation criterion, and computed these values between each predicted fiber tract and the corresponding fiber tract atlas from each subject in the test set, and averaged the values among fiber tracts and individuals”
- Complex sentence on page 8: “Considering an extreme situation, for a fiber streamline, our method can only classify the all voxels on the fiber streamline as one fiber tract, but TractSeg may classify voxel #1 as tract #1, and voxel #2 as tract #2, and so on, but actually these voxels should belong to the same class, this apparently did not conform to the common sense in neuroscience and may resulted in the higher dice score than ours.”
- Complex sentence in Conclusion on page 8: “In this paper, we proposed an effective fiber tract segmentation method, which can identify the accurate and common fiber tracts across individuals based on a novel representation of fiber’s geometrical feature, called as FiberGeoMap, and we also tailored Transformer neural network to meet the input FiberGeoMap”
- 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
8
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper proposes novel fiber tract segmentation approach including a novel geometric descriptor for fiber tracts and use of Transformers and self-attention for segmentation. The extensive evaluation shows significant improvement over state-of-the-art approaches. In addition, clinical application is demonstrated by comparing fiber density in autism versus controls.
- Number of papers in your stack
4
- 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
Not Answered
- [Post rebuttal] Please justify your decision
Not Answered
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 proposes a novel method for fiber bundle classification based on the Transformer network. There is consensus about the novelty of the proposed method, but several concerns about the experimental evaluations were raised. The first is about the fairness and rigor of the comparisons. The second is the lack of comparison and validation on independent datasets.
- 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).
7
Author Feedback
We are very thankful to all the reviewers’ constructive comments and their appreciation of our work. Our responses to the reviewers’ common comments are itemed as follows. 1) Regarding the unfairness of experimental evaluation that reviewer #1 and reviewer #2 are concerned about, we describe the details of our experiments as follows. Because we segmented the all fibers into more fiber bundles (103 fiber bundles), we used the combination of TractSeg atlas and WMA atlas. Part of our training data are exactly the same as that of the TractSeg method, both are 4 / 5 of TractSeg atlas, and the remaining 1 / 5 of TractSeg atlas are used as the test set. Another part of our training data adopted WMA atlas which is also the training data of WMA method, likewise, we kept 1 / 5 of it as test data. In this way, when compared with these two methods, our method suffers unfairness instead, because our method needs to comprehensively learn the characteristics of the two atlases, while the other two methods only use their own training data. Even so, our averaged dice score on common 42 fiber tracts is 0.97, which is higher than that on 103 fiber tracts (0.93), while the dice scores of the other two methods on these 42 classes are 0.85 and 0.54, respectively. This explanation also applies to the concern of Reviewer #2 (The two methods may have different definitions even for the same tract and are performed with different fiber tracking algorithms). Note that we should not put new experimental results here considering the rebuttal rules, but we have had the dice scores on 103 classes (shown in Fig. 6), we just need to select the dice scores on the common 42 fiber tracts from 103 classes rather than providing new experiments. If the paper is accepted, we will replace these data with the averaged dice score on the common 42 classes. 2) Reviewer #1 and Reviewer #2 mentioned that additional data sets should be used for validation. We also run the proposed method on autism data, the results of fiber segmentation is shown in Fig. S1 of the supplementary file, which can be regarded as a qualitative validation. Because there are no labels for fiber tracts in autism or other datasets, quantitative validations, such as computing dice and other metrics, cannot be carried out.
Our responses to the reviewers’ specific comments are itemed as follows. For Reviewer #1: Regarding comparative experiment with DeepWMA, we actually have run the DeepWMA method, but the code released on Github did not distinguish between left and right brain fiber bundles, e.g., AF_left and AF_right were counted as one fiber bundle, so we did not compare with it. Moreover, we found that the fiber segmentation performance of DeepWMA and WMA are very close in [3]. So we only compared it with WMA method.
For Reviewer #2: As noticed by the reviewer, dice is for volumetric overlap. We computed it after converting fibers to masks. The reviewer is concerned if “the proportion of fibers” is a fair measure for abnormities. The proportion of fibers is a trial measure in this work, it is similar to fiber density in [16]. The experimental results also showed that the abnormal fiber bundles identified by this metric were partially consistent with those identified by fiber density [16].
For Reviewer #3: Many thanks to Reviewer 3 for the appreciation of our work. For complex sentences, we will break into multiple simple sentences to improve readability, if the paper is accepted.
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.
Considering the novelty of the proposed method and the responses from the authors, I think this paper can be accepted.
- 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/n (best paper in your stack) and n/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).
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.
Reviewers agreed that the technical contribution in this work is suitable for MICCAI, but even after considering the rebuttal, they disagree on whether the experimental setup is the most suitable. In my opinion, the setup is at least not fundamentally flawed, and additional or alternative evaluations could be left for future work.
- 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/n (best paper in your stack) and n/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).
8
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.
I have the same concerns as reviewer1 and reveiwer2 regarding the fair comparison, and the rebuttal to this point did not convince me.
- 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/n (best paper in your stack) and n/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).
9