Conditional Variational Autoencoder for Functional Connectivity Analysis of Autism Spectrum Disorder Functional Magnetic Resonance Imaging Data: A Comparative Study

Round 1

Reviewer 1 Report

In this work, the Authors propose a method to improve characterization of autism spectrum disorder (ASD) related differences in functional connectivity. Comparison of different generative models revealed that their proposed variational autoencoder (VAE) incorporating a conditional attribute and leveraging convolutional layers outperforms methods currently used in the - literature. I have found this study interesting in general. However, I suggest some clarification and reevaluation of results as explained in my comments. 

- VAEs are applied to preprocessed data obtained after parcellation according to Schaefer atlases. However, analysis of functional connectivity (FC) was performed between characteristic brain networks. Why was it not possible to do a within-network analysis?

- Were the actual Pearson coefficients high and significant, indicating statistical coupling between networks? The significance criteria for comparing FC between ASD and neurotypical samples was p<0.05. Especially after such dimension reduction, the level of significance should be adjusted to multiple comparisons.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper investigated the feasibility of variable autoencoder (VAE) for analyzing functional connectivity derived from fMRI. The authors compared VAE and conditional VAE (CVAE) in the performance of reconstruction from the input data and functional connectivity networks by using a public dataset of healthy subjects (HS) and subjects with autism spectrum disorder (ASD). The authors concluded that the CVAE showed a better performance than VAE, more specifically CVAE using a convolutional neural network model (CNN) compared with recurrent neural network.

This is an interesting study with robust results. Here are some minor comments.

1. CVAE used the input of critical information of the subjects, such as age, sex, HS or ASD. Therefore, it is not at all surprising CVAE showed the better performance than VAE.

2. Because of the comments above, it would be more interesting to discuss why and how the difference between CNN and RNN.

3. The author concluded 'we believe that CNN-based VAE and CVAE are more effective in reconstruction and generation'. In Table 3, CNN shows the lowest similarity between three methods. Table 1 does not seem to show a large difference between the methods in all 4 indicators. It would be useful to discuss by integrating the results of Table 1, 2 and 3.

4. In Conclusion section, the authors suddenly introduces another model 'GANs', which is never mentioned in the paper and start discussing the difference of VAE and 'GANs'. This part might be better to move to Discussion section.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Review: Conditional Variational Autoencoder for Functional Connectivity Analysis of ASD fMRI Data: A Comparative Study

The aim of this research is to perform fMRI-FC analysis using Variational Autoencoders (VAEs) in Autism Spectrum Disorder (ASD) to detect atypical interconnectivity between brain regions. In the first part of the study, the authors investigated multiple VAEs (Conditional, Recurrent, and CNN-RNN hybrid) to establish their application in FC analysis. In the second part, the authors introduced phenotypical data to assess improvements in VAEs to compensate for gender differences. In the third part, the authors compared the results with the previous literature. The study's major finding is that CNN-based VAEs improve fMRI-FC analysis.

This topic is of significant importance due to the neurodivergent patterns in fMRI data associated with ASD, which come with certain limitations including inherent biases, limited interpretability, and the underrepresentation of ASD in the female population. The authors have made an attempt to tackle these issues. However, the objectives of this study remain unmet, and there is room for improvement in data interpretability. To address these concerns, the authors need to provide a clearer explanation of the data, particularly with regard to the chord diagrams.

Comments:

·        Authors states previous fMRI-FC analysis   had “inherent biases or limited interpretability” but do not explain the same in details. The only explanation was age, gender and varied phenotype of ASD.

·        It is understandable that authors are only focusing on predefined regions only based on previous literature. However has the authors looked at connections in other regions too?

·        Ln: 420-421: In the VAE experiments (Figure 8), a consistent trend of underconnectivity  between the Limbic and DMN networks emerges across all models.

I do not observe underconnectivity in the RNN female group, or perhaps the author is specifically referring to the upper row in Figure 8 (male + female). Please provide a more explicit explanation of the graphs, as this could potentially confuse readers.

·        Ln: 422-424: Similarly, multiple models identified underconnectivity between the salience and visual networks, which has remained similarly apparent in both male and female populations

Connectivity Sal>Vis is neurotypical in RNN model, RNN clearly fails to find underconnectivity.

In RNN model, limbic> somatomotor has underconnectivity in female and males, but neurotypical in male+ female. Could authors explain this?

·        Ln: 428-431: Furthermore, a noteworthy difference between males and females lies in the connectivity between the somatomotor and DMN  networks. In males, the somatomotor-DMN connection tends to be under-connected, while  in females, it is over-connected.

But the results are actually mixed, example SM>DMN in males has under connectivity in CNN and RNN, while over-connectivity in Mixed model, while its overconnected in RNN and Mixed, while neurotypical in CNN. Could the authors provide more precise and specific reporting of their results instead of using general terms?

·        Ln 432-434: In the CVAE experiments, some trends are similar to those identified with VAE models.  For example, a trend of underconnectivity between limbic and DMN is apparent for both the male and female populations.

This statement holds true only for RNN and mixed models not CNN, in CVAE experiments.

·        Ln 434- 436 The trend of under-connectivity between limbic and  DMN in males and over-connectivity between limbic and DMN in females remains true for  CVAE experiments

This statement is also not true as limbic > DMN networks shows no overconnectivity in females, in VAE or CVAE experiments.

·        Ln 436- 437: The trend of overconnectivity between visual and limbic became more

pronounced for both males and females in CVAE experiments compared to VAE.

This statements is also not true, in CVAE experiment mixed model has neurotypical connection in female, and CNN- male.

 

·        I suggest that the authors consider comparing their best model with previous data in ASD for a more comprehensive analysis.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors addressed all concerns raised by this reviewer.