Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Currently in the field of computer-aided diagnosis, graph or hypergraph-based methods are widely used in the diagnosis of neurological diseases. However, existing graph-based work primarily focuses on pairwise correlations, neglecting high-order correlations. Additionally, existing hypergraph methods can only explore the commonality of high-order representations at a single scale, resulting in the lack of a framework that can integrate multi-scale high-order correlations. To address the above issues, we propose an Inter-Intra High-order Brain Network (I2HBN) framework for ASD-assisted diagnosis, which is divided into two parts: intra-hypergraph computation and inter-hypergraph computation. Specifically, the intra-hypergraph computation employs the hypergraph to represent high-order correlations among different brain regions based on fMRI signal, generating intra-embeddings and intra-results. Subsequently, inter-hypergraph computation utilizes these intra-embeddings as features of inter-vertices to model inter-hypergraph that captures the inter-correlations among individuals at the population level. Finally, the intra-results and the inter-results are weighted to perform brain disease diagnosis. We demonstrate the potential of this method on two ABIDE datasets (NYU and UCLA), the results show that the proposed method for ASD diagnosis has superior performance, compared with existing state-of-the-art methods.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/1263_paper.pdf

SharedIt Link: pending

SpringerLink (DOI): pending

Supplementary Material: N/A

Link to the Code Repository

N/A

Link to the Dataset(s)

http://fcon_1000.projects.nitrc.org/indi/abide/

BibTex

@InProceedings{Han_InterIntra_MICCAI2024,
        author = { Han, Xiangmin and Xue, Rundong and Du, Shaoyi and Gao, Yue},
        title = { { Inter-Intra High-Order Brain Network for ASD Diagnosis via Functional MRIs } },
        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

    The paper proposed a new method to integrate individual- and population-scale models into a single framework with high-order correlations for ASD diagnosis. The experimental results showed that the integration of high-order intra-correlations at the individual scale and high-order inter-correlations at the population scale into a unified framework had superior classification performance.

  • 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 motivation of the proposed method was good. The explanation of the proposed method and figures in the methods section were easy to understand.

  • 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.
    • It is difficult to undertand why the proposed method constructed a hypergraph because it looks like just a graph.
    • The loss function was not specified in the paper.
    • It did not explained how inter-hypergraph handed a new data during the test.
  • 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 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
    • There was no loss function for learning the proposed model.
    • In the paper, the authors used LASSO for intra-hypergraph modeling, however, LASSO may find the pair-wise relationship, i.e., weighted edge, between two ROIs. Why is it called hyperedge in the paper?
    • If a new person’s data comes in the inter-hypergraph, the inter-hypergraph will change, and the neighbors of all existing people on the hypergraph will likely change. How did the proposed method test a new data?
    • What is the difference between Result_intrai in equation (7) and Evi in equation (4)?
    • Minor correction: there is a typo in the title of section 3.3 (ablation -> ablation).
  • 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

    Reject — should be rejected, independent of rebuttal (2)

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

    The proposed method did not sufficiently implement the motivation presented in the introduction.

  • Reviewer confidence

    Somewhat confident (2)

  • [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 multi-level graph-based framework for ASD prediction from resting state fMRI data. Instead of using pairwise correlation, the authors have used higher-order statistics to create the edges between brain regions. The higher-order statistics can capture complex relationships between regions, which is a novelty of this paper. The constructed graph is fed to the first level of graph convolutions to obtain subject-level feature representations. The individual feature representations are further fed into a second layer of graph convolution that exploits population-level characteristics to make predictions.

  • 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 authors have used higher order statistics to capture relationships between brain regions. The use of higher order statistics can capture complex relationships and may create graph representative of the underlying disorder.
    2. Another strong component of this model is that the authors use a population graph to make predictions. The population graph can exploit the population level characteristics of the disorder to improve risk prediction.
  • 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.

    This paper shows good results in a 5-fold cross-validation setting but lacks implementation details and adequate descriptions of their methodology. Here are the major issues:

    Issues with methodology:

    1. In the paper, the authors never mentioned their training loss or their optimization strategy.

    2. The authors are optimizing 2 GCN, and whether it is done jointly or separately is unclear.

    3. The authors need to provide details about the node features. Are node features generated from fMRI time series?

    4. The authors need to provide further motivation on removing edges with \alpha_i <0. A higher negative value also indicates that the ROI is a good predictor. Would it make more sense to use absolute values?

    5. In Eq. 3 the authors need to explain how X and H are constructed.

    6. The authors mentioned that they are combining the results from itra-GCN and iter-GCN. Are they taking a sum of the probabilities? If yes, how do they binarize the class probabilities to get cases and controls?

    Issues with hyperparameters:

    1. The authors need to provide the details of the hyperparameters they are optimizing and how they are optimizing. They also need to provide the ranges of the hyperparameters.

    Issues with baselines:

    1. From the paper, it is unclear what the inputs to the baselines are. Are the inputs constructed from pairwise correlation maps?
  • Please rate the clarity and organization of this paper

    Satisfactory

  • 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 provide sufficient information for reproducibility.

  • 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
    1. I would recommend the authors reshuffle the methodology and provide the necessary information about their training strategy. Considering the restrictions on space the authors can remove equation 4,5 and 6 because they are simply estimating and thresholding Euclidean distance.

    2.Describing the hyperparameters are essential to reproduce the results. Currently, it’s hard to understand the optimal hyperparamters and how they are optimized.

    1. While explaining the baselines, the authors should try to provide a single line description about the input features. This will help the readers to understand the importance of the model.
  • 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 Reject — could be rejected, dependent on rebuttal (3)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
    1. Lack of information about the methodology made it hard to understand and appreciate the novelty of the model.
    2. Without the hyperparameters it was hard to understand if the results were super optimistic or it is generalizable. Additionally, it would be hard to reproduce.
    3. The inputs to the baselines are not adequately described, making it hard to appreciate the performance boost.
  • 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

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

  • [Post rebuttal] Please justify your decision

    The authors have answered most of my questions. However, there are still some questions whether the authors have made a fair comparison (appropriately optimizing hyperparameters) to the baselines.



Review #3

  • Please describe the contribution of the paper

    The authors describe a strategy to use hypergraphs (graphs whose edges can join any number of node vertices) in two ways to analyze fMRI data connectivity: within a subject (intra-hypergraphs) and across subjects (inter-hypergraphs) in an integrated manner (termed inter-intra hypergraph computation (I2HBN) to perform “disease diagnosis” by classifying Autism Spectrum Disorder (ASD) subjects. Testing is performed on the public ABIDE dataset, with evidence of classification improvements over other SOTA methods.

  • 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.

    Hypergraph construction using the sparse linear regression strategy is interesting and appropriate to yield the inter-hypergraph edges. Furthermore, the multiscale embedding that is described for intra-Hypergraph Learning is certainly current and interesting, although the basic idea appears to have been already published in reference 4. Furthermore, the high-level-correlation-based Intergraph learning is also interesting, but relatively straightforward

    Testing on the ABIDE dataset is well-organized and appropriate state of the art comparisons are made to strategies published in the computer vision literature, including GCN, GAT, HGNN+ and GiGCN.

  • 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.

    While algorithm comparisons to SOTA techniques are reasonable, more direct comparisons to graph neural net strategies published in the medical image analysis literature would have been appreciated.

    The conclusions drawn in section 3.2 are likely overstated. While it is true that I2HBN outperforms the other methods, it is difficult to fully interpret why.

    On the other hand, the ablation studies are more telling regarding the utility of intra vs inter stages and the results and discussion are appreciated.

  • 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 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?

    Reproducibility testing should be included and would strengthen this paper. It should be doable given access to the ABIDE datasets.

  • 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

    It is assumed that the inter-graph learning is performed across both ASD and control classes. If this is done separately within the classes this would be helpful to know. Other than that the paper generally reads well, although comparison directly to already-published ASD classification methods in the medical image analysis literature would have been helpful.

  • 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?

    This paper reads well and the integration of subject-based and population-based hypergraphs is a novel contribution that yields promising results. It adds higher-level correlations to what has been previously published and compares well to other SOTA graph neural network methods in the literature. I would have liked to have seen comparisons to already-published ASD classification strategies in the medical image analysis literature, but realize this can sometimes be difficult to obtain code.

  • 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

Thanks for all the valuable comments. R1, R4: Hypergraphs and Graphs The hypergraph is an extension of the graph, where an edge in a graph can only connect two nodes, thus limiting its ability to represent correlations. In contrast, a hyperedge in a hypergraph can connect any number of nodes, providing a stronger capability to represent complex correlations. We empirical employ the Inductive strategy, where during training, the model aims to learn a general representation to make predictions on unseen data. R1, R4: Motivation The proposed method is motivated by two key factors. Firstly, cognitive activities in the brain involve complex interactions between different regions. Such complex interactions cannot be represented by graph structures, whereas hypergraph structures can effectively represent these group correlations through hyperedges connecting multiple nodes. Secondly, ASD and healthy subjects exhibit complex group-level high-order correlations, such as differences in behavior and neural activity. Understanding these correlations through a group-level hypergraph structure is crucial. Therefore, this paper integrates individual (Intra-hypergraph) and group (Inter-hypergraph) high-order correlations within a framework and aims to enhance ASD diagnosis and provide biomarkers from a high-order correlations perspective for clinicians. R1, R4: Proposed Method The proposed method employs a two-stage training optimization process. (1) In the first stage, individual subjects are modeled with different brain regions (segmented based on the AAL template) as nodes in the hypergraph structure, and the fMRI signals of brain regions as node features (i.e., X). This establishes a higher-order association structure H among different brain regions. The goal is to learn individual brain function embeddings (Evi in Eq. 4) guided by high-order correlations and to compute individual hypergraph prediction scores (Result_intra in Eq. 7). Notably, LASSO is introduced in the hypergraph modeling to preserve from ROIs with αi values＞0 and form a hyperedge with the i-th ROI. Note that negative values are not deeply explored in this paper, but it is an interesting idea because interactions between neurons involve both activation and inhibition. This kind of relationship may also exist between different brain regions, and this line of thought will continue to be explored. (2) In the second stage, all subjects are modeled together, with each subject as a node. The intra-embedding generated in the first stage is used as the feature to establish high-order correlations among different subjects. The goal is to learn functional and behavior differences at the group level and to generate group-level prediction scores (Result_inter). (3) The final results are obtained by weighted fusion of the two (Eq. 7). Additionally, the experimental results show that our proposed two-stage method outperforms the end-to-end method. This is mainly because the quality of intra-embedding significantly affects the results of the second stage, and generating a better embedding in the first stage leads to a superior group hypergraph structure. R1, R3, R4: Training Architecture and Parameter Settings The proposed method uses cross-entropy as the loss function in both stages. The Intra_model consists of two hypergraph convolutional layers, followed by pooling and two fully connected layers. Dropout is set to 0.5, the optimizer is Adam, with a learning rate of 1e-5, and 100 epochs. In the second stage, the learning rate is set to 1e-2. We will open-source our code after the paper is accepted, along with the ABIDE subject IDs and cross-validation parameters used. R1: Comparative Experiments The methods compared in this paper follow the training protocols mentioned in the cited literature. The graph methods use validated features: graph edge weight matrices as node features. The input features for hypergraph methods and our method are the same, all being fMRI signal information.



Meta-Review

Meta-review #1

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    Upon reading the reviews, paper and the rebuttal, it appears that the authors have addressed a majority of the concerns raised during the review process (barring the comment on negative correlations).

    The paper offers an interesting methodology for using hypergraph representations to discover individual specific (intra-) higher order dependence as well as construct population hyper-graphs to aid discriminability within populations. The methodology, experiments, baselines and results seem to support the main contributions of this work

  • What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).

    Upon reading the reviews, paper and the rebuttal, it appears that the authors have addressed a majority of the concerns raised during the review process (barring the comment on negative correlations).

    The paper offers an interesting methodology for using hypergraph representations to discover individual specific (intra-) higher order dependence as well as construct population hyper-graphs to aid discriminability within populations. The methodology, experiments, baselines and results seem to support the main contributions of this work



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    Reviewers were mixed in their final assessments. They agree on many strengths of the paper including the proposed method that considers both intra- and inter-hypergraph learning to integrate individual and population models, the clear explanation of the motivation and methods, and the good empirical results. However, there were a number of concerns, including some missing methods/implementation details that require clarification and missing comparisons to relevant GNNs in the ASD medical image analysis literature. It also appears that there may have been some reviewer misunderstanding regarding the hypergraph definition that affected their review/rating. Considering these points, I agree with the majority of the reviewers and recommend accept. The authors should take care to include the requested clarifications regarding some methods/implementation details in the final version.

  • What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).

    Reviewers were mixed in their final assessments. They agree on many strengths of the paper including the proposed method that considers both intra- and inter-hypergraph learning to integrate individual and population models, the clear explanation of the motivation and methods, and the good empirical results. However, there were a number of concerns, including some missing methods/implementation details that require clarification and missing comparisons to relevant GNNs in the ASD medical image analysis literature. It also appears that there may have been some reviewer misunderstanding regarding the hypergraph definition that affected their review/rating. Considering these points, I agree with the majority of the reviewers and recommend accept. The authors should take care to include the requested clarifications regarding some methods/implementation details in the final version.



back to top