Abstract

The connectivity structure of brain networks/graphs provides insights into the segregation and integration patterns among diverse brain regions. Numerous studies have demonstrated that specific brain disorders are associated with abnormal connectivity patterns within distinct regions. Consequently, several Graph Neural Network (GNN) models have been developed to automatically identify irregular integration patterns in brain graphs. However, the inputs for these GNN-based models, namely brain networks/graphs, are typically constructed using statistical-specific metrics and cannot be trained. This limitation might render them ineffective for downstream tasks, potentially leading to suboptimal outcomes. To address this issue, we propose a Customized Relationship Graph Neural Network (CRGNN) that can bridge the gap between the graph structure and the downstream task. The proposed method can dynamically learn the optimal brain networks/graphs for each task. Specifically, we design a block that contains multiple parameterized gates to preserve causal relationships among different brain regions. In addition, we devise a novel node aggregation rule and an appropriate constraint to improve the robustness of the model. The proposed method is evaluated on two publicly available datasets, demonstrating superior performance compared to existing methods. The implementation code is available at https://github.com/NJUSTxiazw/CRGNN.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/1009_paper.pdf

SharedIt Link: pending

SpringerLink (DOI): pending

Supplementary Material: https://papers.miccai.org/miccai-2024/supp/1009_supp.pdf

Link to the Code Repository

https://github.com/NJUSTxiazw/CRGNN

Link to the Dataset(s)

http://adni.loni.usc.edu/ http://fcon_1000.projects.nitrc.org/indi/abide/

BibTex

@InProceedings{Xia_Customized_MICCAI2024,
        author = { Xia, Zhengwang and Wang, Huan and Zhou, Tao and Jiao, Zhuqing and Lu, Jianfeng},
        title = { { Customized Relationship Graph Neural Network for Brain Disorder Identification } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
        year = {2024},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15002},
        month = {October},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    Here are the contributions of the paper: 1.A novel CRGNN framework is proposed that integrates graph structure learning with downstream tasks within a unified framework, addressing the discrepancy between graph structure and downstream tasks. 2.The introduction of a CRB that adaptively learns the most suitable brain networks/graphs for various downstream tasks in a customizable manner. 3.A significant advantage over traditional statistical-specific methods by capturing nonlinear interactions among brain regions, providing a more comprehensive understanding of brain dysfunction.

  • 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 CRGNN introduces a novel approach to learning brain network structures by dynamically capturing causal relationships among brain regions, which is a significant advancement over static, correlation-based methods. In addition, the method proposed by this paper integrates graph structure learning with downstream tasks allows for an end-to-end learning process that closely aligns the network structure with the objectives of disease identification. Moreover, the interpretability of the model has always been something we admire. The CRGNN model proposed by this paper focuses on causal effects rather than just correlations, so the identified discriminative connections not only enhance the interpretability of the model, but are also consistent with established neurological research, further highlighting the performance of the model.

  • 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 readability of the article needs to be improved, For example, the explanation of the symbols (tanh()) in the formula and some wording in the article need to be improved.
    2. The description of the Directed Graph Convolution(DGC) is not detailed enough. In the formula process of disseminating message, nodes and adjacency matrices are generally needed, but the two definitions are not given in detail in the article. Secondly, why does this use a directed graph? How is the direction of this directed graph defined? Is it one-way or two-way? If this is based on other papers, please provide corresponding literature support. Otherwise, please describe it clearly. Finally, what role do the two parameters 𝜃 1 and 𝜃 2 play in DGC?
    3. The method used to transform a two-dimensional matrix into a vector is not clearly explained.
    4. The font of the formulas in the main frame diagram(Fig.1.) is too small and difficult to read clearly.
    5. The meaning of the brain map obtained after model training is not well explained, even if the author says that it is a causal relationship between brain areas.
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Do you have any additional comments regarding the paper’s reproducibility?

    Please describe the GDC method clearly or provide the source code directly.

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
    1. Re-edit and revise the descriptions and explanations in the article.
    2. Explain the DGC method in detail.
    3. Explain why there will be a larger improvement after adding LG loss.
    4. Change the text in the figures to the appropriate size.
    5. Provide a more reasonable explanation for this brain map generated through model training.
    6. Disclosing the source code of the model allows readers to better understand this work.
  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making

    Weak Accept — could be accepted, dependent on rebuttal (4)

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

    The readability of the article needs to be improved, and the main methods in the article are not explained clearly, which brings obstacles to understanding the model proposed in this article. Secondly, some details in the article can be improved to improve the quality of this article. Finally, the causal relationship between the parameters obtained by training as brain intervals proposed in this article needs further explanation.

  • Reviewer confidence

    Very confident (4)

  • [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 #2

  • Please describe the contribution of the paper

    The authors introduce a novel approach called the Customized Relationship Graph Neural Network (CRGNN), designed to concurrently optimize graph network learning and downstream task training. Comprising two key components, CRB and RAB, the framework functions as follows: CRB constructs a graph G by capturing causal relationships among brain regions, while RAB aggregates node features within a shared latent space determined by G. Evaluation of the method is conducted on two widely accessible datasets, ADNI and ABIDE.

  • 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 offers a clear and coherent narrative, guiding readers through the significance of the problem, existing methodologies, their constraints, and how the proposed approach mitigates these limitations. 2) The idea/methodology is interesting and promising. 3) A thorough evaluation of the framework is conducted, including comparisons with five baseline methods and three ablation studies, which meticulously demonstrate the significance of each module within the framework. The discriminative connectivity showcased, coupled with alignment with the relevant literature, provides compelling evidence of the method’s efficacy.

  • 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) Section 2.3 is a bit confusing. For example, the authors mention that G(l) = G(l)+I in equation 2. Does this mean that G always includes self-connections? At which stage of the pipeline does this happen? 2) Regarding the ablation studies and evaluation, did the authors test only the RAB module, while using Pearson in the CRB module? In my understanding, it would be meaningful to explore whether the gates/causal relationships as a starting point, actually contribute to the result.

  • Please rate the clarity and organization of this paper

    Very Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Do you have any additional comments regarding the paper’s reproducibility?

    N/A

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html

    See weaknesses above.

    Also, for the TopK pooling layer, how was the retention of the ⅓ of the original number of nodes decided?

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making

    Weak Accept — could be accepted, dependent on rebuttal (4)

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

    I see potential in the paper and find it promising.

    However, my primary concern revolves around the efficacy of the CRB module in enhancing performance. Once the authors provide an explanation or present the results of the ablation study/experiment addressing this concern, I am open to reconsidering my rating.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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

    This paper proposed a Customized Relationship Graph Neural Network that can bridge the gap between the graph structure and the downstream task.

  • 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 presented a novel GNN model for brain disorder identification that integrated graph structure learning and downstream tasks within a unified framework.

  • 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.

    Please describe how to treat the unbalanced data in your experiments.

  • Please rate the clarity and organization of this paper

    Very Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Do you have any additional comments regarding the paper’s reproducibility?

    N/A

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html

    Please describe how to treat the unbalanced data in your experiments.

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making

    Accept — should be accepted, independent of rebuttal (5)

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

    The method is innovative. The paper is well wrote.

  • Reviewer confidence

    Very confident (4)

  • [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




Author Feedback

We thank all reviewers for their encouraging and constructive comments. Our responses to the major concerns are itemized as follows. 1) The specific meaning of the symbol tanh() in Formula 1 (R1): The symbol tanh represents an activation function that is widely used in the field of deep learning. 2) The description of DGC should be improved (R1): The node is defined by the AAL atlas [14], where each node corresponds to a specific brain region. Additionally, the CRB module is developed to estimate the causal relationships between brain regions. This is why the adjacency matrix (i.e., a directed graph) is adopted to aggregate node information. The two parameters θ1 and θ2 represent the parameters of two fully connected layers. These layers are utilized to convert the node features into a more compact dimension. 3) The method for transforming a 2D matrix into a vector (R1): This is implemented by the built-in function, namely torch.tensor.reshape, of the PyTorch framework. 4) The meaning of the brain map is not well explained (R1): Each node in Fig. 2 represents a brain region, and the causal relationships between brain regions are inferred from the CRB module. The directed edges in Fig. 2 demonstrate significant differences in the strength of connections between these brain regions among the two groups of subjects. 5) Explain why there will be a larger improvement after adding LG loss (R1): The directed graph G is a part of the model’s parameters to be optimized, and imposing a specific constraint on G helps reduce the risk of overfitting. The LG loss acts as a regularization term, guiding the optimization process to favor simple models that generalize better to new data. As a result, the addition of the LG loss leads to a more robust model, resulting in a significant enhancement in performance. 6) Disclosing the source code of the model (R1): Yes, we will release the code when the work is formally published. 7) Does G always include self-connections (R4)? Yes, we add self-loops to the obtained graph when performing the directed graph convolution operation, a technique that was also frequently utilized in similar studies [8]. The addition of a self-loop is utilized to retain valuable information from the nodes themselves when updating node features. 8) Did the authors only test the RAB module in the ablation studies, while using Pearson in the CRB module (R4)? Yes, there is only one combination option remaining when testing the effectiveness of the CRB module, as indicated in the final row of Table 2. This is because, for this ablation experiment, the directed graph was replaced by a brain network/graph constructed using the Pearson correlation coefficient, which results in an undirected graph. Therefore, the DGC layer is currently unsuitable for aggregating node features. Instead, the graph convolution operation [8] is suitable for undirected graphs. In addition, the loss term LG is not included in this experiment because the Pearson brain network is not part of the model’s parameters in this setting, and adding the constraint does not affect the model’s performance. Therefore, this is the reason why there is only one combination of ablation experiments regarding the CRB module. 9) How was the retention of 1/3 of the original number of nodes decided (R4)? This is a hyperparameter of the model, similar to the learning rate, and is typically determined through empirical optimization. (10) Please describe how to treat the imbalanced data in your experiments (R5): The ADNI dataset utilized in this paper exhibits some imbalance in terms of the number of subjects in the two categories, with a ratio of approximately 1:1.5. From the results in Tables 1 and 2, it is evident that our method demonstrates good performance even on unbalanced data.




Meta-Review

Meta-review not available, early accepted paper.



back to top