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

Authors

Zhuobin Huang, Hongmin Cai, Tingting Dan, Yi Lin, Paul Laurienti, Guorong Wu

Abstract

Functional resonance magnetic imaging (fMRI) technology has been widely used in understanding cognition and behavior by characterizing the functional interaction between distant brain regions. Since the network topology of functional connectivity (FC) often dynamically shifts along with the change of brain states, it is challenging to identify the change point (transition between tasks) of functional connectivity without requiring prior knowledge of experiment settings. Although striking efforts have been made to detect changes on BOLD (blood-oxygen-level-dependent) signals, little attention has been paid to characterize the trajectory of whole-brain functional connectivity, which is more closely correlated to brain state change. Since FC is essentially a symmetric positive definite (SPD) correlation matrix, we present a change point detection network (CPD-Net) tailored to (1) learn the low-dimensional geometric feature representations of whole-brain functional connectivity on the Riemannian manifold of SPD matrices, and (2) automatically detect the brain state changes on the unseen functional neuroimages. It is worth noting that our CPD-Net is a manifold-based neural network to the extent that we leverage the alignment between the known functional tasks and the stratification underlying the learned low-dimensional FC feature representation on the Riemannian manifold of SPD matrices to steer the learning of geometric patterns from functional brain networks. We have evaluated the accuracy and replicability of our CPD-Net on task-based fMRI data from HCP (human connectome project) database, where our manifold-based CPD-Net achieves more accurate and consistent results than current learning-based CPD methods.

Link to paper

DOI: https://doi.org/10.1007/978-3-030-87234-2_51

SharedIt: https://rdcu.be/cyl8N

Link to the code repository

N/A

Link to the dataset(s)

N/A

Reviews

Review #1

Please describe the contribution of the paper

They devised a deep manifold-based network (CPD-Net) to detect change point of brain state.
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 first attempt to detect the change of brain states using the manifold-based neural network.
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.

It has no obvious weaknesses. But the originality is not very high, because the problem is not new, and the results and conclusion is not very interesting, they only use a revised deep neural network, which has not been used to resolve the problem.
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

They evaluate the replicability of the brain state change detections between test and retest data where the task schedules are different. It seems reasonable.
Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://miccai2021.org/en/REVIEWER-GUIDELINES.html

The authors described much for the methods, for example, the section2.3 (Back-propagation to solve the network minimization) seems verbose and not very necessary. They should explain more about why they do that.
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 paper has no fatal weakness and no great advantages.
What is the ranking of this paper in your review stack?

2
Number of papers in your stack

3
Reviewer confidence

Confident but not absolutely certain

Review #2

Please describe the contribution of the paper

This article uses the Riemannian Manifold theory to reduce the dynamic FC matrix and uses the mean-shift method to clustering task states, and then evaluates the model’s accuracy and replicability.
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 article uses the Riemannian Manifold theory to reduce the dynamic FC matrix and uses the mean-shift method to clustering task states, and then evaluates the model’s accuracy and replicability.
Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
1. The training set and testing set have been determined before the experiment, is there different result when using random and repeated training set and testing set?
2. This study discusses the impact of using SPD-DNN to reduce input dimension or not, should other dimensionality reduction methods be considered?
3. This model uses deep learning, so what is the configuration of the computer in this experiment and how long did the training take?
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 dataset HCP is open
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

See Q4
Please state your overall opinion of the paper

accept (8)
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

the paper introduces a new method to mease brain state
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

In this paper, the authors proposed a change point detection framework for dynamic functional connectivity (dFC) data. The framework is composed of 1) a manifold-based neural network for nonlinear dimensionality reduction and 2) a mean-shift recurrent neural network for task stratification. By combining the two components, the proposed framework can better stratify the dFCs so that the stratification matches the underlying task designs.
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 provides a novel framework that seems to be effective in both of the two components. The first dimensionality reduction component considers the manifold structure of SPD matrices so that it should perform better than traditional dimensionality reduction methods. The second MS-RNN component clusters the time points that belong to the same task closer, which also seems to help the clustering by pushing different clusters away from each other.
2. The visualization is pretty clear and shows the effect of each component.
Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
1. Both the SPG-DNN and MS-RNN is proposed by the previous research. So there isn’t much novelty in terms of the methodology.
2. Detecting 2back vs 0back in the working memory task is well-studied already and it is not a novel application.
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

Open source data is used. The methods are adopted from other papers. So overall the experiments can be replicated given that the hyperparameters are reported.
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. As there are as many as 7 tasks available in the HCP dataset, it will be better if the author can detect change points in the other scans to show that the trained model is indeed generalizable.
2. It will also be more interesting if the author can extend the framework to resting-state data where people are more curious about the change point there considering it is unconstrained.
3. From the Figure 2, it looks like MS-RNN smooths the data points within the same cluster. Can the authors discuss if simpler methods like “cluster+smoothing” can achieve similar performance? If not, why?
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 components of the framework are not novel and the application is also not very interesting. However, the combination is pretty interesting and the improvement is convincing compared with other methods. The paper can be a lot better if the author provide more holistic study on different task/resting-state scans that are available in the HCP dataset.
What is the ranking of this paper in your review stack?

1
Number of papers in your stack

5
Reviewer confidence

Confident but not absolutely certain

Primary Meta-Review

Please provide your assessment of this work, taking into account all reviews. Summarize the key strengths and weaknesses of the paper and justify your recommendation. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

As mentioned by the reviewers that the originality is limited in both the problem and approach. The authors should clarify their technique contribution and novelty in the rebuttal. In addition, the author should also clarify if their method can be extended to resting-state data and how to evaluate it.
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).

9

Author Feedback

We thank the reviewers for their constructive feedback. We address the critiques below and will incorporate all feedback in the final version.

Major concerns: Q1: The novelty is not very high in terms of neural network components (R1, R7). A: Brain stage change detection is one of the popular topics in neuroscience. However, very few methods paid attention to the intrinsic data geometry of the function network, which is eventually represented by a symmetric and positive-definite (SPD) matrix. Thus, our major contribution is that we present a manifold-based deep learning approach to characterize the dynamic functional fluctuations. Compared to the conventional methods, which often break down the network topological information by vectorizing the brain network into a data array, the learned geometric patterns by our CPD-Net allow us to track the trajectory of the whole-brain functional network more accurately along with the brain stage changes. Furthermore, our method probably provides the basis for exploring the geometry insight of the whole functional brain network. The existing SPD-DNN is used as a dimension reduction component in our CPD-Net.

Q2: Lack of the intuition behind network design (R1) A: In general, we consider each functional brain network as a data instance in the Riemannian manifold of SPD matrix. Furthermore, we cast the change detection problem as a clustering process on the Riemannian manifold. In a nutshell, we can adapt the iterative mean-shift procedure from Euclidean space to Riemannian manifold using the well-studied Lie-group algebra. Since the dimensionality of the brain network is relatively high, it is of high demand to reduce the dimensionality. In this context, we design our CPD-Net with two components, where we concatenate SPD-DNN with an RNN implementation of mean-shift. By doing so, our CPD-Net is able to associate the functional brain networks with the pre-scheduled cognitive tasks on the Riemannian manifold. We provide the detail of gradient calculation in the back-propagation step in Section 2.3. We will make it clear in the final version.

Q3: Application to resting state (R7) A: We completely agree with this reviewer that the overarching focus is to identify functional dynamics to the resting state. Our proposed CPD-Net offers a new way to achieve this goal using the manifold-based deep learning technique. In this work, we mainly focus on validating the change detection performance on task-fMRI data, which has the ground truth to do the quantitative evaluation. In this context, our trained deep model with the optimal hyper-parameters based on the task-fMRI data can test on resting-state fMRI data. As no ground truth is available, two-sample covariance matrix testing can be used to examine if functional connectivity patterns of two consecutive data segments split by the detected change points are significantly different, the differences can be used as the surrogate evaluation criterion.

Minor concerns: (R2) Regarding the distribution of the dataset, we split training (5 folds), validation (1 fold), and testing (4 folds) set in a 10-fold manner. We repeat the 10-fold validation process three times and find the state change detection results are very consistent (74.55%, 74.54%, 74.58%).

(R2) Regarding the possibility of using other dimension reduction methods, our CPD-Net is general to work with other manifold-based approaches to reduce the dimensionality.

(R2) Regarding the spec of PC, we use an Intel (R) Core (TM) i7-8700 CPU @ 3.20GHz PC without graphic card. The training time is about 5h. The running time of the testing set is 5mins (about 1s/subject).

We thank for the suggestion of using “clustering+smoothing” from R7. Since the data distribution is complex and non-linear, we have the concern that such a two-step approach might have limited power in disentangling the latent clusters. This is also the reason we use an iterative mean-shift process. But it is worthy of comparison.

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.

Though I am convinced by the authors regarding their method’s novelty, I also agree with the reviewers’ concerns about extension to resting state fMRI data. For task data, we have ground truth in some extent, but we also have other methods can fullfil the similar goal. If the proposed method can not prove the potential for rs fMRI, the usefulness is significantly compromised, since it is difficult to judget what is the value of the proposed method over other approaches.
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).

18

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 were responsive to the reviews and the meta-reviewer critiques. While the novelty is modest, the paper presents a new way of analyzing fMRI dynamics that can potentially translate across datasets.
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

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 work proposed a low-dimension Riemannian manifold theory-based symmetric positive definite matrices representation approach together with the mean-shift method to clustering task states. The work studied the brain state changes with fMRI data. The work was evaluated on the HCP dataset for its accuracy and replicability.

The strength of this work is with the novel application of manifold-based neural network to brain functional MRI data analysis. The work is solid and was evaluated in publicly available dataset.

The weakness is lack intuition explanation with the work. The authors did a good job in the rebuttal to clarify some concerns from the reviewers and AC.

An “Accept” recommendation is made to recognize its novel application and may bring new ideas to brain functional MRI data analysis research.
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

back to top

Detecting Brain State Changes by Geometric Deep Learning of Functional Dynamics on Riemannian Manifold