List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Xiang Tang, Xiaocai Zhang, Mengting Liu, Jianjia Zhang
Abstract
Construction and analysis of functional brain network (FBN) with rs-fMRI is a promising method to diagnose functional brain diseases. Traditional methods usually construct FBNs at the individual level for feature extraction and classification. There are several issues with these approaches. Firstly, due to the unpredictable interferences of noises and artifacts in rs-fMRI, these individual-level FBNs have large variability, leading to instability and unsatisfactory diagnosis accuracy. Secondly, the construction and analysis of FBNs are conducted in two successive steps without negotiation with or joint alignment for the target task. In this case, the two steps may not cooperate well. To address these issues, we propose to learn common and individual FBNs adaptively within the Transformer framework. The common FBN is shared, and it would regularize the FBN construction as prior knowledge, alleviating the variability and enabling the network to focus on these disease-specific individual functional connectivities (FCs). Both the common and individual FBNs are built by specially designed modules, whose parameters are jointly optimized with the rest of the network for FBN analysis in an end-to-end manner, improving the flexibility and discriminability of the model. Another limitation of the current methods is that the FCs are only measured with synchronous rs-fMRI signals of brain regions and ignore their possible asynchronous functional interactions. To better capture the actual FCs, the rs-fMRI signals are divided into short segments to enable modeling cross-spatiotemporal interactions. The superior performance of the proposed method is consistently demonstrated in early AD diagnosis tasks on ADNI2 and ADNI3 data sets.
Link to paper
DOI: https://doi.org/10.1007/978-3-031-43993-3_21
SharedIt: https://rdcu.be/dnwNm
Link to the code repository
https://github.com/seuzjj/ACIFBN
Link to the dataset(s)
Reviews
Review #2
- Please describe the contribution of the paper
This paper proposed to learn common and individual FBNs adaptively with Transformer framwork, in order to joint the feature extration step and the task application step.
- 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 paper proposed adaptively contructing and analyzing method for FBNs, which is solid and sufficient based on the evaluations.
- 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.
The authors provided the difference of DMN, FN , CEN between NC and eMCI that is meaningful for the further early diagnosis. But if the difference could be displayed on the brain, that will make the readers especially the physicians understand the difference intuitively.
- 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
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/2023/en/REVIEWER-GUIDELINES.html
1.It is suggested to highlight the contributions more clearly.
- For Table1, the mean±std performance should be provided instead of only the best performance.
- 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
6
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper is well-organized and provides friendly figures to make the research purpose and algorithm are easy understood,
- Reviewer confidence
Very confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #3
- Please describe the contribution of the paper
The contribution of the paper lies in proposing a new method for constructing and analyzing functional brain networks using rs-fMRI data. This method involves adaptive learning of common and individual functional brain networks within a Transformer framework to address issues related to variability and unsatisfactory diagnosis accuracy. The method also models asynchronous functional connectivities to better capture actual functional connectivities. The performance of the proposed method was consistently demonstrated to be superior in early Alzheimer’s disease diagnosis tasks on ADNI2 and ADNI3 data sets. The paper also identified diagnostic clues through the learned individual functional brain networks. The contribution of the paper is significant in advancing the research on functional brain networks and providing useful guidance for clinical applications.
- 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 main strengths of the paper include:
- Proposing a novel method for constructing functional brain networks using rs-fMRI data that addresses issues related to variability and unsatisfactory diagnosis accuracy.
- Utilizing a Transformer architecture that allows for adaptive learning of common and individual functional brain networks.
- Modeling asynchronous functional connectivities, which results in a more accurate representation of actual functional connectivities.
- Consistently demonstrating superior performance in early Alzheimer’s disease diagnosis tasks on ADNI2 and ADNI3 data sets compared to other state-of-the-art methods.
- Providing diagnostic clues through the learned individual functional brain networks, which can aid in the identification of early-stage Alzheimer’s disease. Overall, the proposed method has the potential to significantly improve our understanding and diagnosis of various neurological disorders.
- 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.
The main weakness of the paper:
- The proposed method was evaluated on only two data sets (ADNI2 and ADNI3), which may limit the generalizability of the results.
- The sample size of early Alzheimer’s disease cases in the evaluation is relatively small, which may affect the statistical power of the analysis.
- The proposed method is computationally intensive, which may limit its practical applications in real-world scenarios.
- There is a lack of comparison with other deep learning-based methods that deal with the same problem, which may limit the reliability of the performance comparison.
- 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
I think they did well on the reproducibility of the paper
- 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/2023/en/REVIEWER-GUIDELINES.html
The author did a good job, but some articles need to pay attention to some small details, such as too many numbers after DPARSFA and SPM-12, maybe they want to cite but forgot to mark it.
- 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
6
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
They proposing a novel method for constructing functional brain networks using rs-fMRI data that addresses issues related to variability and unsatisfactory diagnosis accuracy.
- Reviewer confidence
Confident but not absolutely certain
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #4
- Please describe the contribution of the paper
The author proposed to learn asynchronous common and individual FBNs adaptively within the Transformer framework. Both the common and individual FBNs are built by specially designed modules, whose parameters are jointly optimized with the rest of the network for FBN analysis in an end-to- end manner.
- 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 asynchronous information among different brain regions is encoded into the framework. Improve the classical kernel self-attention mechanism by introducing a data-driven common FBN module as regularizations.
- 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.
Insufficient description of technical implementation. The AUC results of the proposed method are much worse than BrainNetCNN in two datasets.
- 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
Many technical details are missing. For example, how to construct a common FBN? How to deal with the cross spatiotemporal asynchronous FBN matrix, which dimensions (nNnN) are expanded. In addition, a parameter in experimental settings is confusing. i.e., The author claimed feature dimension D is set as 12, while matrix X_0 belongs to NTD is 90128*6.
- 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/2023/en/REVIEWER-GUIDELINES.html
- The description of Figure 2 cannot fully correspond to the entire figure. i.e., Missing the description of subfigures (d) and (e).
- Equation 3 is inconsistent with subfigure (c). In equation 3, the Sparsemax function is performed on the sum of normalized A and C, while in subfigure (c), the addition operation is performed after the Sparsemax function.
- I’m curious about if synchronous correlations be updated along with time-lagged asynchronous correlations.
- Why only explore asynchronous FC in the first two layers? how to handle asynchronous information in the remaining three layers of the encoder?
- It is not clear how to construct a common FBN.
- Please clarify in detail why the AUC results of the proposed method are much worse than BrainNetCNN in two datasets.
- 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?
This paper focuses on some key issues and conducts extensive experiments, but there are several issues that need further clarification.
- Reviewer confidence
Very confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
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.
The author proposed a functional brain network modeling approach by learning asynchronous common and individual networks adaptively via the transformer model. Both the common and individualized networks are identified by specially designed modules for AD diagnosis. The key strength of this paper is to integrate certain reasonable and novel aspects into the modeling, such as asynchronous information among different brain regions, introducing a data-driven common FBN module as regularizations, etc. Although there are some interesting findings and merits in this paper, the meta-reviewer as well as some reviewers have some concerns and confusions, and invite the authors to provide the rebuttal to clarify these major concerns: 1. The AUC performance compared to other methods; 2. Insufficient technical details and parameter settings; 3. Insufficient of comparisons with other methods. Please also refer to the detailed comments from each reviewer.
Author Feedback
Thank all the reviewers for their comments and constructive suggestions. R#2 Q1：Highlighting the contributions more clearly. A1：Will highlight as follows: 1) Adaptive learning of common and individual FBNs within Transformer framework, improving the diagnosis accuracy; 2) Modeling both synchronous and asynchronous functional connectivities, resulting in richer and more accurate FBNs; 3) Proposing an integrated framework for FBN construction and analysis, increasing the flexibility and adaptivity of the model.
Q2: Providing mean±std in Table 1. A2：The mean±std have been prepared to be provided in the final version. Our method’s ACC stds(2.4 and 1.8 on ADNI2 and 3) are comparable to or less than others and our improvements are statistically significant.
Q3: Visualizing the differences of NC and eMCI on the brain. A3：They will be visualized in future version.
R#3 Q1: Evaluation on only two datasets may limit the generalizability of the results. A1：The ADNI2 and ADNI3 datasets are among the largest and the most commonly used public fMRI datasets for AD. We also conducted evaluation on our private data for subcortical vascular cognitive impairment diagnosis and the results (Our ACC 84.7% vs. the second-best BrainnetCNN of 77.5%) consistently demonstrate the effectiveness and generalizability of our method. They were not included due to page limit. As explained in Section 3.2, to further ensure the fairness and meaningful statistics, we used random 5-fold cross-validation for all methods.
Q2: About the computation and practical applications. A2：The complexity of our method has been carefully considered and significantly reduced by using the kernel self-attention[29], as introduced in Section 2.1. It takes less than 1 hour to train on ADNI2 and less than 1 second to test one sample with a single A6000 GPU, and it should be acceptable in practice.
Q3: Lack of comparison with other deep learning (DL)-based methods. A3：5 SOTA DL-based methods covering the most representative CNN-, GCN- and Transformer-based ones have been compared in our evaluation in Table 1. More SOTA DL methods will be compared in future version.
Q4: Correcting small typos. A4: They will be corrected.
R#4 Q1: About the technical details and reproducibility of paper. A1: We will 1) publicly release the full implementation of our method with example data ASAP to provide sufficient technical implementation details and ensure the reproducibility; 2) clarify the technical details as much as possible, e.g.: i) the common FBN C is parameterized as a learnable matrix with the same size of individual FBN A; ii) the signals of the N ROIs are segmented into nN segments and treated as nN nodes to construct nN*nN FBNs; iii) the feature dimension D is 12 except 6 for the first layer input.
Q2: Missing description of Fig.2(d) and (e). A2: They are explained in Section 2.2.
Q3: Inconsistency between Eq.3 and Fig.2(c). A3: Fig.2(c) is confirmed to be aligned with the actual implementation and Eq.3 will be updated as a(Q,K,V)=(sparsemax(A/(TD)^0.5)+C)V. All other Eqs have been carefully checked.
Q4: if synchronous correlations are updated with asynchronous ones. A4: As shown in Fig.2(d), both synchronous (in blue) and asynchronous (in red) correlations are simultaneously modeled and updated.
Q5: Why only explore asynchronous FC in the first two layers? A5: 1) reducing the complexity of the model with limited training data; 2) the asynchronous info can be well extracted by the first two layers and higher layers correspond to high level features.
Q6: Why our AUC results are worse than BrainNetCNN? A6: Really sorry, the AUCs of BrainNetCNN reported were typos. That’s why they were not consistent with the reported ACC, SEN and SPE. The actual AUCs of BrainNetCNN on ADNI2 and ADNI3 are 74.6 and 71.8 while our method consistently achieves the best (80.0 on ADNI2 and 75.5 on ADNI3). The correctness and consistency of all other results are checked and confirmed.
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.
This paper proposed a functional brain network modeling approach to learn asynchronous common and individual networks adaptively via the transformer model, and applied it on AD diagnosis. The key strength of this paper is to integrate certain reasonable and novel aspects into the modeling, such as asynchronous information among different brain regions, introducing a data-driven common FBN module as regularizations, etc.
The authors have provided detailed rebuttal to address the reviewers and AC’s concerns. A majority of reviewers retain positive comments on this paper. And I also recommend acceptance of this paper
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.
Most of the major concerns (such as technical details) have been well clarified.
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 title of this paper is “Learning Asynchronous Common and Individual Functional Brain Network for AD Diagnosis”. After reading the paper, my feeling is that the authros were trying to propose an interesting but incremental methods based on transformer, and fit it to a classification task using ADNI. I have been working on this fields for many years and I strongly doubt the practical value of this paper in AD diagnosis. I think ADNI 3 does not have the lable of EMCI becauause EMCI and LMCI (these two terms) have been used in ADNI 2 only. It is not clear how the authros used the ADNI 3 for classification.