Learning to Fuse Multiple Brain Functional Networks for Automated Autism Identification

Round 1

Reviewer 1 Report

This paper presents a novel multi-functional connectivity network (FCN) fusion framework to classify autism spectrum disorder (ASD) from resting-state functional MRI (rs-fMRI) data. Traditional FCN estimation methods only capture a single relationship between brain regions of interest (ROIs), failing to model complex interactions. The proposed approach estimates multiple FCNs from different perspectives, assigns fusion weights based on label information (ASD vs. healthy), and applies these weights to an adaptively fused FCN to distinguish ASD from healthy controls.

This study has several strengths. First, the use of multiple FCN estimation methods captures the complex interactions between brain regions, which can provide a more holistic view of brain dysfunction in ASD. Second, the use of label information to guide the learning of fusion weights is a novel approach that can increase the discriminative power of the fused FCN. Third, the adaptively weighted fused FCN was validated on the ABIDE dataset, a well-established and widely used dataset in ASD research, which enhances the credibility and replicability of the study.

However, several areas could be improved:

·        The paper lacks a clear comparison of the proposed method with other multi-FCN fusion methods. Comparing results directly with other recent works would allow readers to better understand the improvement brought by the proposed method.

·        Statistical analysis of the results should be performed to evaluate confidence and reliability.

·        Further elaboration on the FCN estimation methods used, how the label information was incorporated, and how the fusion weights were learned would be beneficial for replicability and understanding the robustness of the study.

·        The authors do not discuss how their findings could be applied in clinical settings, such as how this might change current ASD diagnosis procedures or improve patient outcomes.

·        While the authors mention plans to learn different weights for each subject, they do not discuss other potential improvements, such as exploring other types of networks or optimizing the fusion process.

In summary, the paper proposes an innovative method for improving ASD identification using rs-fMRI data. However, additional methodological detail and discussion of clinical implications and future directions could enhance the paper's impact and relevance to the field of ASD research.

Author Response

请参阅附件。

Author Response File: Author Response.pdf

Reviewer 2 Report

This topic is interesting, but the authors should make some revisions in the document. Consider these points:

- Lines 97-119. It is not clear what is the purpose of this paper. Revise it.

- Lines 271-273: "s. The classification results for NYU and UM sites are shown in Tables 4 and 5, respectively." Please report results in the text too and improve this part.

- Lines 319-324: "This verifies that 319 multi-FCN fusion is helpful for the improvement of classification performance. At the same... unknown real brain networks". This part must be improved. In this part authors must consider this very important and recent papers in Pubmed: -- PMID: 35886651 -- doi: 10.3390/ijerph19148799  -- -- PMID: 36424395 --  DOI: 10.1038/s41380-022-01854-7

- Lines 331-332: "The thickness of the solid line connecting two ROIs indicates the strength of the connection" This sentence is not well connected to the context.

- Lines 352-354. It is not clear is this manuscript has some limitations. Add a "strengths and limitations" section of the paper and revise it.

- Lines 356-358. In the conclusion section authors wrote that they "propose a simple multi-FCN fusion strategy, in which the fusion weights are optimized under the guidance of label information for ASD identification". Can the authors explain how to reproduce this model for further research and improve the conclusion in this regard?

- Figure 5 requires more explanation. Improve its figure legends.

Author Response

请参阅附件。

Author Response File: Author Response.pdf

Reviewer 3 Report

Reviewer’s Report on the manuscript entitled:

Learning to Fuse Multiple Brain Functional Networks for Automated Autism Identification

The authors proposed a multi-functional connectivity network fusion model for autism spectrum disorder classification. I found the method and results interesting, but the presentation can be improved. Also, the proposed algorithm needs further clarification. Please see below my comments.

Line 98. Figure 1 should be in method Section 2.2 not in Introduction.

Lines 100-112. These should be in the conclusion section not the introduction. Instead, please highlight the main contributions. What is the objective and what will be proposed and for what purpose?

Figure 1. What is “HC”? Please define all the acronyms the first time they appear and also consider including an acronym table at the end of the manuscript.

Style/editorial issue. The location of Table 1 should be after line 127.

Table 2. Please remove “the” before “Eq. (1)”

Table 2. How do you guarantee that the “If” condition (|Fi+1-Fi|=epsilon) can be reached? I think it should be “<” not “=”. Also, the term in the “if condition” may never get less than epsilon. How is the convergence guaranteed? I think you need to add other stopping conditions as well.

Figure 2. What is Lnner Loo in the flowchart? Please define in the caption.

Line 271. Please add the following articles that describe all these statistical metrics mathematically with their applications in medicine and brain sciences:

https://doi.org/10.1109/JSEN.2023.3237383

https://doi.org/10.3390/pr11041210

https://doi.org/10.1016/j.aci.2018.08.003

What about the computational time of your method with respect to other methods? Please list it in a table and/or discuss.

Please use the past verbs in Conclusion. For example, line 156. We proposed… line 358. incorporated, etc. Please also mention the limitations of your method.

Regards,

The English grammar is generally OK, but there are typos/punctuation issues that should be checked and corrected.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

-The algorithms and the functions of the classification models have to be explained more clearly.

-There are many acronyms which are also defined  with words the first time taht they appear. It would also be useful to provide a table with the acronyms together with their descriptions.

-Algorithm  of Table 2: It is not clear why and how the test

abs(F/sub(i+1)-F/sub i)=epsilon

in the final part of Step 2 can be always processed with "equality" to epsilon when running the algorithm.

-In the loss functions (1), (2), (4), (5) , it is not clear the choice of the values y/sub i= +1, -1.  Are the loss functions performed with the 2**M possibilities for the two eventual  values (+1, -1 ) of  the whole possible sets of y/sub i?. Identify clearly the data and the results when minimizing the  objective functions.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors have addressed the comments and improved the paper. The paper can be accepted for publication.

Reviewer 2 Report

Authors solved all my criticisms.

Good

Reviewer 3 Report

I thank the authors for addressing my comments and improving their manuscript.

Please carefully proofread the manuscript and also fix the following minor issues:

Line 108. Grammar issue: It should be: "... in an unified ..."

Line 114. It should be "In Section 2,..." not "...the second 2"

Line 379. In the third big O, it is little L not I. Please fix.

Thank you!

There are some grammar/punctuation/typo issues that should be checked and corrected. For example:

Line 108. Grammar issue: It should be: "... in an unified ..."

Line 114. It should be "In Section 2,..." not "...the second 2"

Line 379. In the third big O, it is little L not I. Please fix.