Two-Step Feature Selection for Identifying Developmental Differences in Resting fMRI Intrinsic Connectivity Networks

Round 1

Reviewer 1 Report

see attached file

Comments for author File: Comments.pdf

Author Response

Reviewer 1: In general, the manuscript could use careful editing to clarify several statements and to clean up minor grammatical and language issues. As currently written, the potential significance of this work to the neuroimaging community is not strongly communicated.

 

Introduction

 (1) This manuscript introduces a new method for analysis of rs-fMRI data based on a two-step process of feature selection. However, the introduction fails to develop the rationale for the choice of method, providing instead a simple list of methods available for ICN identification, dimensionality reduction and feature selection. While this section does not require a critical evaluation of each possible method, it should provide the reader with the reasons for the choice of (a) a two-step process, and (b) the specific methods chosen for that process. The final paragraph does not clearly define the purpose of the study, instead rambling from the results of other studies that are best suited to the discussion to methods that are not mentioned elsewhere in the introduction such as graph analysis, and a comparison between feature-selection methods. If these latter two methods are intended as validation steps, it is not clear from this paragraph.

Reply: Sincerely here, we quite appreciate your affirmation of our paper. Based on your invaluable suggestions, we have added more details about the shortage of the existing feature selection methods and the advantages of the two-step feature selection method proposed here.

Since feature selection methods can not only remove the redundant information or invalid features but also keep the key information of the original data set, the feature selection methods are used widely to functional connectivity network study. Based on the combination of feature selection search and the construction of the classification model, filter and wrapper are two typical types of feature selection methods. Additionally, ensemble is another popular kind of feature selection method. Filter methods mainly depend on the attribute of features, and the evaluation criteria is independent from classifiers. Wrapper methods embed the feature selection search with the classification performance of the classifier in order to improve the accuracy of the classifier, but the drawback of this method is the higher risk of the overfitting and computationally intensive when the classifier has a high computational cost. Ensemble methods are based on different sampling strategies to extract multiple sample sets, and then, a specific feature selection algorithm is used to obtain multiple sets of feature subsets.

For the proposed two-step feature selection method, it couples two effective feature selection methods to assemble an optimal feature set for building the prediction model. This method takes full advantages of the support vector machine recursive feature elimination, which can efficiently reduce noise and irrelevant features in classification task. In addition, by combining the F-score with correlation score, only the discriminative features can be kept. Our results suggest that the two-step feature selection method performs the best in comparing with other six significant filter, wrapper and ensemble feature selection methods.

All the details above have been highlighted in the revised paper.

 

Specific Issues and suggestions:

(1) This manuscript focuses on a single neuroimaging modality and is not arguing that it is uniquely able to provide information on mesoscale network connectivity. Therefore, I suggest that there is no need to list other methods in the initial paragraph. Rather, focus on why this specific modality was chosen

Reply: Thanks for your invaluable suggestions. In order to focus on the functional magnetic resonance imaging (fMRI) data and make the theme of the introduction clearer, we have expurgated the other neuroimaging modality in the part of introduction.

 

(2) Paragraph 3 is again a simple list of methods that are not used here. Either re-write this section to indicate why the method you chose is superior to each other method listed, focus on some of the more common methods and provide a reason why yours is superior, or simply provide the rationale for your choice of method

Reply: It is so kind of you to give us such insightful comments, we have adopted your suggestion to rewrite this paragraph. Because this paragraph should focus on the feature selection methods and more detail about other methods such as independent component analysis (ICA), independent vector analysis (IVA), principle component analysis (PCA), non-negative matrix factorization (NMF) and a sparse representation and dictionary learning (SRDL) are extraneous with the main part, we have abridged the other methods and added the detail of the feature selection methods and the advantages of our method in the introduction.

 

(3) Paragraph 4 moves to the choice of study population, but doesn’t really indicate why the authors chose to begin with this population since their method is not unique to childhood. There is also no real justification for the use of the parcellation method chosen relative to other parcellation methods (e.g. high dimensional ICA, HCP parcellation).

Reply: Thank you very much for your comments. Yes, the proposed two-step feature selection method is not only applicable to the population data here. We only used this data to show the good performance of the method. In our current research, we utilize it to some other dataset, and it can still get good results.

In addition, this manuscript focuses on finding out the ICNs from the high dimension but small sample size data. However, the other machine learning methods, such as high dimensional ICA, have no function on removing the redundant information or invalid features while keeping the key information of such kind of data. That is why we consider other feature selection methods as the comparing methods.

 

(4) Lines 79-80: I would quibble with the idea that ICNs are “a great challenge to identify” as there are many methods available to do so, and numerous publications. Instead, this section should focus on the potential benefits of machine learning algorithms to assist or improve on current methods for ICN identification

Reply: We do really appreciate your insightful comments. We have adopted your suggestion to focus on the advantages of the machine learning methods. Because the FC data usually has high dimension but small sample size, i.e., n<<p, where n is the sample size and p is the number of features, it is a great challenge to identify ICNs from FC data. For such data, it is too difficult to discover the potential information contained in the data from a limited number of observations, which form a cognitive concept of the data or complete identification task. Thus, we propose the two-step feature selection method to identify ICNs from FC data, it can take full advantages of the support vector machine recursive feature elimination (SVM-RFE), which can efficiently reduce noise and irrelevant features in classification task. In addition, by combining the F-score with correlation score, only the discriminative features can be kept.

 

(5) Line 40-41: What is meant by the phrase “prediction of individuals”?

Reply: Thanks for your helpful suggestion and we have adopted this suggestion to correct the sentence of “prediction of individuals” into “diagnosis of disease” to make it clear.

 

(6) Lines 92-93: This sentence concerns “the most significant differences” and then proceeds to encompass all possible differences, given the final “among others”

Reply: Thanks for your great suggestion and we have adopted this suggestion to correct the sentence of “the most significant differences” into “the significant differences” to make the sentence precise.

 

Methods

Given that this method is being proposed for high-dimensioned data and few datasets, some indication of minimal numbers of datasets in validation and test sets should be provided. Acknowledging that methods for maximizing the use of small data sets (i.e. Leave-one-out approaches) can be leveraged, it would still be helpful to note any lower bounds.

Reply: Thank you very much for your suggestion. For this kind of small data set, the proportion of training set and test set in this paper is 70% and 30% respectively, which is based on some previous experience (please refer Mitchell T M . Machine learning[M]. McGraw-Hill, 2003). In this way, a better training process can be obtained. Thank you for your recommendation of other methods, we will try them in our following experiment.

 

Specific Issues

(1) Lines 105-106: “accuracy of prediction and identification” of what?

Reply: We do really appreciate your insightful comments. This sentence is indeed not complete enough. In fact, it should be the prediction accuracy and the identification of the sparse ICN changes. It has been revised in the revision.

 

(2) Line 125: “calculated by (??)” is missing information

Reply: Thank you for your careful discovery, this is due to the negligence of the personal editor, has been revised.

 

(3) Line 129: Does the phrase “which may however have some discriminative ability” indicates that this feature should be retained, and this is the point of the next several sentences.This point could be more clearly made in the initial sentence.

Reply: Thank you very much for your question. In fact, our algorithm does achieve the retention of those potential discriminative features, making up for the shortcomings of the original SVM-RFE algorithm.

 

(4) Lines 128-144: The section on classification accuracy is unclear. What criteria are being used for accuracy? Does this assume a binary classification? How large must the training and validation sets be to provide optimal results, given that this method is stated to work on small population data sets?

Reply: Thank you very much for your question. The classification accuracy is used to justify if the feature with the smallest score calculated by (3) should be deleted. Specifically, if one feature with small weight is picked, while the corresponding classification accuracy on the validation set decreases after its removal, then this feature should be reconsidered. Additionally, multiple subsampling strategies on the training data are utilized and a series of feature selection processes are performed to avoid mistakenly removing such features. Yes, it is a binary classification since there are two groups. For such small data sets, the training and validation sets used here are 75% and 25% of the sample ratio, respectively.

 

(5) Figure 1: How are test data actually used? This figure implies that these data simply feed to a ROC analysis. Do these data feed into the algorithm at step 1, step 2 or at initiation?

Reply: Thank you very much for your question. In fact, the test data is only used for the detection the performance of the proposed two-step feature selection model, i.e., in the subsection 3.5., comparison with other feature selection methods. It has never been fed into the algorithm at step 1, step 2 or at initiation. The entire feature selection part is completed by the training set, which has been divided into two parts, one for feature selection, and another for validation.

 

(6) Results

The introduction to this section contains largely redundant information, with required details introduced in subsequent paragraphs. All that is need here is the initial sentence and the final sentence. As noted, the use of a graph theoretical approach and the comparison to other feature selection methods was not justified in the introduction.

Reply: Thank you for your suggestion. As for the redundancy of this paragraph, we have deleted the redundant content belonging to subsequent sub-paragraphs according to your comments, and maintained the simplicity of this part.

 

(7) Line 270: it is unclear to me how the feature vector V is actually defined: are feature vectors between dFC matrices concatenated in the feature selection, or is each vector analyzed independently? Assuming concatenation, are feature vectors from all subjects combined to form a large feature matrix or, again, is each analyzed independently?

Reply: Thank you very much for your comments. The feature vector V of each subject is a description of the average dFC between each of two ROIs. Thus, each subject has a unique feature vector V, and is analyzed independently.

 

(8) Line 277: In dividing this data set into training and test sets, is there equal representation of the two age groups of interest in both sets? How was this ensured? If it was not done,why not?

Reply: Thank you very much for your question. For the dataset we used here, although the samples of the two age groups are not balanced, but the difference of the size of the two groups is small; the numbers of children are 193, and the numbers of adults are 204. Thus, we use the same proportion of training data as well as testing data for the two groups; the consistency is achieved by stratifying sampling to scale the original data by 70% for training and 30% for testing.

 

(9) Line 288: How was assignment to RSNs actually done?

Reply: In order to facilitate the analysis on these 264 ROIs which are identified by Power et al. [1], the functional roles to brain regions are assigned from the work of Smith et al. [2], based on which, different resting state networks (RSNs) are defined [3]. 

The paper is as follows:

[1] J. D. Power et al., “Functional network organization of the human brain,” Neuron, vol. 72, no. 4, pp. 665–678, 2011. 

[2] S. M. Smith et al., “Correspondence of the brain’s functional architecture during activation and rest,” Proc. Nat. Acad. Sci. USA, vol. 106, no. 31, pp. 13040–13045, 2009. 

[3] S. M. Smith et al., “Network modelling methods for fMRI,” NeuroImage,

vol. 54, no. 2, pp. 875–891, 2011.

 

(10) Line 298: “COTCN contains …. etc” is completely uninformative. What, specifically, are the regions considered to constitute this network? You are free to call your networks whatever you define, but I point out that regions generally considered part of some of these RSNs are not included here.

Reply: Thank you very much for your question. In fact, the definition of RSNs  is based on the current generic definition template, and the correspondence between the brain regions and the RSNs are also defined in literature.

 

(11) Line 303: more information on those ROIs labeled “uncertain” is needed. Why, if they are not attributable to a RSN, are they included?

Reply: According to the corresponding relationship between ROIs and RSNs, those ROIs with which the functional roles are uncertain, are uniformly merged into the uncertain RSN. Because the functions of this network is still controversial, thus, in this paper, the functional connections within it dose not been further discussed.  

 

(12) Figure 5: This figure needs to be improved. The numbers along the outer perimeter of the figure are unintelligible and, since no legend or label indicates their meaning, uninformative. They would be removed, making the color coding for the RSNs clearer.The legend for this figure should indicate the color codes and what the legend constitutes. The color scheme for the legend makes differentiating each RSN very difficult.

Reply: It is very kind of you to give us such valuable advice. In the original Figure 5, the numbers along the outer perimeter of the figure represented the 264 ROIs. These ROIs correspond to 12 RSNs. In order to make the correspondence between the ROIs and the RSNs clear, the numbers are still retained. While, we have made some improvement on it. We have adopted your advice to change the color scheme and added the correspondence between the 264 ROIs and the 12 RSNs. In this way, it is easier for readers to understand this figure.

 

(13) Line 306: If you are going to call “uncertain ROIs” RSN 13, include it in the RSN figure so that the locations of these ROIs is clear to the reader.

Reply: Thank you very much for your suggestion. Based on it, we have improved Figure 4 by adding the uncertain RSN in it.

 

(14) Figure 6: What is the purpose of the different colors of connectivities between RSNs? The figure legend is not informative.

Reply: Thank you for your suggestion. In order to facilitate the reader's understanding, we have replaced the original one with two subgraphs, and marked the enhanced as well as weakened connectivities between RSNs in different colors.

 

(15) Figure 7: Here the connection strengths are not, as in Figure 6, between RSNs but between ROIs and that is not clearly stated in either the text or the figure legend.

Reply: Thank you for your comments. The connection strength here is between the ROIs, and we have added the statement in the figure legend.

 

(16) Lines 316-332: This section should be clarified as to WHAT changes in the named RSNs are“remarkable” before simply listing those RSNs in which differences could be noted between groups. This is particularly true as the ‘child’ group is rather broad and encompasses cognitive development in multiple domains

Reply: Thank you for your good suggestion. We have emphasized the changes of the functional connectivity in the revised paper. Additionally, in order to explain that the results can reflect the specific functional differences during brain development, we also gave some explanation and existing revelation of the functions for the RSNs.

 

(17) Lines 333-349: This section belongs in the discussion – which doesn’t actually exist.

Reply: Thank you very much for your suggestion. In order to make the structure of the article clear, this paragraph has been put into the discussion section.

 

(18) Section 3.4: In this section, graph methods were applied to “two brain networks” based on“selected ICNS”. The process by which these networks were constructed, and the final matrix constructed are not noted. Please provide details on not only the fundamental methods but also the statistical evaluation of these data.

Reply: Thanks for your great suggestion. The definition of the brain network is the physical or structural connections linking neuronal units, and we use the graph theory to depict the brain network, where the brain is divided into 264 regions (ROIs) which are called as nodes, and the edge is the pairwise relationship between nodes. The size of the connection matrix is 264 times 264, where the diagonal element is 1, and the element of (i,j) in the matrix is the correlation between node i and node j. In order to make the connection matrix more sparse, based on the selected ROIs, we retain the elements of the corresponding position of the selected ROIs in the connection matrix, and let the rest elements be the zero. Finally, we use the obtained connection matrix and the formula in Table 2 to calculate the four parameters of the network.

 

(19) Conclusions

There should be a discussion section included, with two specific components: one focused on methods development issues (performance data, generalizability, etc) and the other on the developmental results (how these results fit into the literature, what new information is gained by this method over, for example, the graph methods used).

This is not a study of “normative development” as a rather broad developmental period is encompassed and there was no effort to subdivide either data set to actually study “development”. It was instead, a classification problem, and should be presented as such. In that case, how the proposed classification method outperforms other classification methods should be a focus of the discussion/conclusion.

Reply: Thank you for your invaluable suggestion. We have adopted your suggestion and added a discussion section before the conclusion section, which consists of two paragraphs. One paragraph describes the advantages of the two-step feature selection proposed in this paper, and the other on the developmental results of the method.

The purpose of this paper is to identify the essential differences of brain functional connectivity as one grows and find out the nature of RSNs change. The classification is used to verify the effectiveness of the proposed two-step feature selection method, and the classification performance of of different feature selection methods which were used for comparision is shown in subsection 3.5.

 

Thank you again for your invaluable suggestions and comments.

Reviewer 2 Report

1.

What is a lack or problem of previous studies? How authors’ method proposed here could solve the problem? This information is totally missing, so I could understand what is a strong point of this study. I agree that the data shown in Figure 8 support that the authors method have advantage over other previously suggested method, but it seems to be trivial. This my impression might be caused that there is no information what is an advantage of this method above other methods. Authors well reviewed related studies in this field in Introduction, but no information about the lack, weak-point, and shortage in these previous studies. These might be authors’’ motivation to introduce a new method.  Authors should express this in Introduction part.  Please consider it.

2.

I could not catch authors’ idea to use the dFC (dynamic functional connectivity) measure in their analysis. As authors stated, it is true that many researchers focused on the dynamic changes in functional connectivity strength. However, as shown in LINES 266-269, authors averaged dFC index between each of two ROIs, and they used the averaged index in the following analysis steps.  Why averaged? The process loses almost all dynamic aspects of functional connectivity.  Authors have to justify why they took temporal average of dFC, and have to clearly denotes what is the difference between the sFC (stational functional connectivity) and averaged-dFC, in other words, what is an advantage of using averaged dFC.

 

In Figure 7, authors showed the differences of FC between two developmental states. Is this simply taking a subtraction of Pearson’s correlation (or normalized Pearson’s correlation) between two FC values acquired in two states? Even though the subtraction between state YA (Young-adults) and C (Children)is positive, there are several types of changes in connectivity, as you know: Positive (YA) > Positive (C), Zero (YA) > Negative (C), Negative (YA) >Negative (C), and so on. Every change could be interpreted in the different ways. Positive (YA) > Positive (C) could be understood as ‘decreased positive connectivity’, Zero (YA) > Negative (C) could be ‘emergence of negative (inhibitory) connectivity’, Negative (YA) >Negative (C) is as ‘enhancement of inhibitory influence between two networks’. At least for me, it seems to be very difficult to simply conclude that ‘… several decreases in connectivity between FPTCN and SSN for young adults… (in LINES 339-340)’. Authors should declare how they calculated the differences between two FC values, what types of connectivity change was behind them, and

In addition, relating to this issue, this conclusion might be derived from the results shown in Figure 6, that shows a ‘number of different function connections (LINES 313-314)’. There are two types of difference: increase and decrease represented in red and blue color in Figure 5, respectively. How authors counted the number of these two types of differences, and summarized into one ‘number’ in Figure 6? What the color in Figure 6 represents? No information, so I could not judge whether authors conclusion is meaningful or not. I need more detailed information about this.

 

Minor

Figure 4: Ventral attention and Dorsal attention network pattern seems to be identical, so that it should be a mistake.

Figures 4 and 5: I would like to know the spatial distribution of ‘uncertain network’ shown in Figure 5.  Is it have any spatial structure, or spatially randomly distributed? Authors should add the ‘uncertain network’ maps to Figure 4, and it could help potential readers what the ‘uncertain network’ represents. Please consider. 

Figure 5: It is very difficult for me to distinguish differences of color of X mark, and it makes me hard to understand what network component has connection to other component. This figure, Figure 5, is a basis of Figure 6 that represents summarized information shown in Figure 5, so please try to improve the visibility. 

Author Response

Reviewer 2:

(1) What is a lack or problem of previous studies? How authors’ method proposed here could solve the problem? This information is totally missing, so I could understand what is a strong point of this study. I agree that the data shown in Figure 8 support that the authors method have advantage over other previously suggested method, but it seems to be trivial. This my impression might be caused that there is no information what is an advantage of this method above other methods. Authors well reviewed related studies in this field in Introduction, but no information about the lack, weak-point, and shortage in these previous studies. These might be authors’’ motivation to introduce a new method.  Authors should express this in Introduction part. Please consider it.

Reply: Thanks for your suggestion, and we have adopted the suggestion to add more detail about the shortage of the existing feature selection methods and the advantages of the method proposed here.

Since feature selection methods can not only remove the redundant information or invalid features but also keep the key information of the original data set, the feature selection methods are used widely to functional connectivity network study. Based on the combination of feature selection search and the construction of the classification model, filter and wrapper are two typical types of feature selection methods. Additionally, ensemble is another popular kind of feature selection method. Filter methods mainly depend on the attribute of features, and the evaluation criteria is independent from classifiers. Wrapper methods embed the feature selection search with the classification performance of the classifier in order to improve the accuracy of the classifier, but the drawback of this method is the higher risk of the overfitting and computationally intensive when the classifier has a high computational cost. Ensemble methods are based on different sampling strategies to extract multiple sample sets, and then, a specific feature selection algorithm is used to obtain multiple sets of feature subsets.

For the proposed two-step feature selection method, it couples two effective feature selection methods to assemble an optimal feature set for building the prediction model. This method takes full advantages of the support vector machine recursive feature elimination, which can efficiently reduce noise and irrelevant features in classification task. In addition, by combining the F-score with correlation score, only the discriminative features can be kept. Our results suggest that the two-step feature selection method performs the best in comparing with other six significant filter, wrapper and ensemble feature selection methods.

All the details above have been highlighted in the revised paper.

 

(2) I could not catch authors’ idea to use the dFC (dynamic functional connectivity) measure in their analysis. As authors stated, it is true that many researchers focused on the dynamic changes in functional connectivity strength. However, as shown in LINES 266-269, authors averaged dFC index between each of two ROIs, and they used the averaged index in the following analysis steps.  Why averaged? The process loses almost all dynamic aspects of functional connectivity.  Authors have to justify why they took temporal average of dFC, and have to clearly denotes what is the difference between the sFC (stational functional connectivity) and averaged-dFC, in other words, what is an advantage of using averaged dFC.

Reply: Sincerely here, we quite appreciate your affirmation of our paper. In fact, most previous studies have assumed that the functional brain connectivity is static during scan time, but more and more research shows that the brain connectivity is dynamic rather than static, even at rest. Functional brain connectivity can also change dynamically in tens of seconds, which has better time resolution than sFC, and can capture many details that cannot be displayed under static conditions, so in this article, we utilize dFC for study. The reason why the average dFC is used is as follows. In the experiment, the function connection matrices obtained by each sliding window process are not quite different, so the the average dFC can be used as the plenipotentiary of the function connection matrices. Additionally, with experiments, the classification accuracy of the averaged dFC is significantly higher than that of the sFC. Thus, we used the averaged dFC here.

 

(3) In Figure 7, authors showed the differences of FC between two developmental states. Is this simply taking a subtraction of Pearson’s correlation (or normalized Pearson’s correlation) between two FC values acquired in two states? Even though the subtraction between state YA (Young-adults) and C (Children)is positive, there are several types of changes in connectivity, as you know: Positive (YA) > Positive (C), Zero (YA) > Negative (C), Negative (YA) >Negative (C), and so on. Every change could be interpreted in the different ways. Positive (YA) > Positive (C) could be understood as ‘decreased positive connectivity’, Zero (YA) > Negative (C) could be ‘emergence of negative (inhibitory) connectivity’, Negative (YA) >Negative (C) is as ‘enhancement of inhibitory influence between two networks’. At least for me, it seems to be very difficult to simply conclude that ‘… several decreases in connectivity between FPTCN and SSN for young adults… (in LINES 339-340)’. Authors should declare how they calculated the differences between two FC values, what types of connectivity change was behind them?

Reply: Thank you very much for your question. In fact, for the calculation of the different FCs between the two groups, we first averaged all the feature vectors of the two groups separately, and based on the selected features, the difference between the two groups can be calculated. In which, the positive values represent the connection enhancement and the negative values represents the connection weakened. At the same time, we also considered the different types of changes in connections, and the  results show that the ROIs all have positively relations, i.e., the type of change belongs to Positive (YA) > Positive (C).

 

(4) In addition, relating to this issue, this conclusion might be derived from the results shown in Figure 6, that shows a ‘number of different function connections (LINES 313-314)’. There are two types of difference: increase and decrease represented in red and blue color in Figure 5, respectively. How authors counted the number of these two types of differences, and summarized into one ‘number’ in Figure 6? What the color in Figure 6 represents? No information, so I could not judge whether authors conclusion is meaningful or not. I need more detailed information about this.

Reply: It is very kind of you to give us such valuable advice. Based on the correspondence between the ROIs and the 12 RSNs, the selected 134 FCs are assigned to the 12 RSNs. In Figure 6, the thickness of the line represents the number of different function connections, i.e., the thicker the line is, the stronger functional connection is.

 

(5) Figure 4: Ventral attention and Dorsal attention network pattern seems to be identical, so that it should be a mistake.

Reply: Thank you very much for your comments. The duplicate picture in Figure 4 is due to the careless of us. We have corrected it in the revised paper.

 

(6) Figures 4 and 5: I would like to know the spatial distribution of ‘uncertain network’ shown in Figure 5. Is it have any spatial structure, or spatially randomly distributed? Authors should add the ‘uncertain network’ maps to Figure 4, and it could help potential readers what the ‘uncertain network’ represents. Please consider.

Reply: Thank you very much for your valuable comments. According to the corresponding relationship between ROIs and RSNs, those ROIs with which the functional roles are uncertain, are uniformly merged into the uncertain RSN. Because the functions of this network is still controversial, thus, in this paper, the functional connections within it dose not been further discussed. While, according to your suggestion, we have added the uncertain RSN to make the integrity of this manuscript.

 

(7) Figure 5: It is very difficult for me to distinguish differences of color of X mark, and it makes me hard to understand what network component has connection to other component. This figure, Figure 5, is a basis of Figure 6 that represents summarized information shown in Figure 5, so please try to improve the visibility.

Reply: It is very kind of you to give us such valuable advice. In the original Figure 5, the numbers along the outer perimeter of the figure represented the 264 ROIs. These ROIs correspond to 12 RSNs. In order to make the correspondence between the 264 ROIs and the 12 RSNs clear, we have made some improvement on it. We have adopted your advice to change the color scheme and added the correspondence between the 264 ROIs and the 12 RSNs. In this way, it is easier for readers to understand this figure. Meanwhile, in order to improve the visibility of Figure 6, we changed it to two figures; one is for connection enhancement, and the other is for connection weakened. They are represented by redlines and blue lines respectively. Then, it is easier for readers to understand both Figure 5 and Figure 6.

 

Thank you again for your invaluable suggestions and comments.