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

Abstract

The graphical representation of the brain offers critical insights into diagnosing and prognosing neurodegenerative disease via relationships between regions of interest (ROIs). Despite recent emergence of various Graph Neural Networks (GNNs) to effectively capture the relational information, there remain inherent limitations in interpreting the brain networks. Specifically, convolutional approaches ineffectively aggregate information from distant neighborhoods, while attention-based methods exhibit deficiencies in capturing node-centric information, particularly in retaining critical characteristics from pivotal nodes. These shortcomings reveal challenges for identifying disease-specific variation from diverse features from different modalities. In this regard, we propose an integrated framework guiding diffusion process at each node by a downstream transformer where both short- and long-range properties of graphs are aggregated via diffusion-kernel and multi-head attention respectively. We demonstrate the superiority of our model by improving performance of pre-clinical Alzheimer’s disease (AD) classification with various modalities. Also, our model adeptly identifies key ROIs that are closely associated with the preclinical stages of AD, marking a significant potential for early diagnosis and prevision of the disease.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/1153_paper.pdf

SharedIt Link: https://rdcu.be/dV18R

SpringerLink (DOI): https://doi.org/10.1007/978-3-031-72086-4_48

Supplementary Material: https://papers.miccai.org/miccai-2024/supp/1153_supp.pdf

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{Sim_MultiModal_MICCAI2024,
        author = { Sim, Jaeyoon and Lee, Minjae and Wu, Guorong and Kim, Won Hwa},
        title = { { Multi-Modal Graph Neural Network with Transformer-Guided Adaptive Diffusion for Preclinical Alzheimer Classification } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
        year = {2024},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15005},
        month = {October},
        page = {511 -- 521}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes a novel end-to-end framework called GTAD (Graph Transformer-based Aggregation and Diffusion) for graph classification tasks, providing interpretability on the brain networks by visualizing the localized scales captured by the diffusion kernels. This can help facilitate early diagnosis and prevention of Alzheimer’s disease.

  • 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.
    • Interesting novel method for interpretable visualization of brain networks that involves visualizing the localized scales captured by the diffusion kernels
    • This can potentially facilitate early diagnosis and prevention of Alzheimer’s disease
  • 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 could discuss potential future research directions -The limitations of the current GTAD framework could be more explicitly stated (for example,the computational complexity of the diffusion kernel and transformer components)

  • 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?

    While the details on hyperparameter settings are provided in the supplementary material, the lack of publicly available code may make it difficult for others to fully reproduce the results. The authors could consider open-sourcing the code to increase the impact and reproducibility of their work.

  • 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
    • The sentence “These conventional methods rely on homophily condition that node features locally connected by the edges are behave similarly” contains a typo and could be rephrased for clarity.
    • The statement “however, these methods often disregard sufficient expressive power of the central nodes, lacking interpretation of the result” could be expanded to better explain the motivation for the proposed GTAD approach. -The experiments on structural brain networks from Diffusion Tensor Imaging (DTI) and ROI measures from functional imaging in the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset demonstrate the potential of the developed GTAD framework for graph classification and interpretation. However, the claim of “practical results for pre-clinical AD classification”(Introduction section, last paragraph) would require further evaluation involving clinicians to validate the clinical utility of the approach. The authors could mention that incorporating clinician feedback and assessment is an important future direction to solidify the practical applications of GTAD -The specific combinations of biomarkers used for the 3-way classifications in the ADNI dataset could be provided for reproducibility.
  • 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?

    The paper presents a novel and promising framework for graph classification, with a strong focus on interpretability for brain network analysis. Addressing the suggested comments, particularly the future directions, limitations could help to further strengthen the paper.

  • Reviewer confidence

    Not confident (1)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

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

  • [Post rebuttal] Please justify your decision

    I would like to thank authors on taking time to properly address comments and elaborating more on limitations of GTAD. Making an effort to release code will further improve method reproducibility and enhance its adoption.



Review #2

  • Please describe the contribution of the paper

    The authors proposed a framework to aggregate both short-and long- range properties for better prediction of graph labels, 2) demonstrating superior performance on graph classification in comparisons to the state-of-the-art methods, and 3) showing interpretability on the brain networks in a scenario with multiple imaging biomarkers.

  • 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.

    Good writing and well organized experiments.

  • 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.

    Some parts are not clear enough. The figure should show how to generate the gnn graph. The author does not mention the computer resources they used.

  • 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?

    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

    some parts are not clear.

    1. How does tractography work here? What kind of tractography algorithm?
    2. Figure 1 should add MRI and DWI at the beginning since this is multi-model work. The figure should point it out.
    3. Figure 2 color bar, why the highest scale value is 1.13? Need more explanation.
  • 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?

    writing, experiment

  • 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

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

  • [Post rebuttal] Please justify your decision

    After saw the feedback from authors, I changed to accept.



Review #3

  • Please describe the contribution of the paper
    1. proposing a framework to aggregate both short- and long- range properties for better prediction of graph labels.
    2. demonstrating superior performance on graph classification in comparisons to the state-of-the-art methods.
    3. showing interpretability on the brain networks in a scenario with multiple imaging biomarkers.
  • 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 proposed method learns node-centric parameters of a diffusion kernel which are governed by a transformer to capture both local characteristic and global graph-level information.

  • 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 symbol K denotes both the convolution layers and the key of the self-attention module, which raises confusion.
    2. The standard deviations of GTAD is at the same level with other methods, even higher in the experiment of Cortical Thickness & \beta-Amyloid. This is inconsistent with what the author stated. What is the reason for this situation?
  • 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?

    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

    This paper proposed a framework to aggregate both short- and long-range properties for better prediction of graph labels. I have some concerns as follows:

    1. The experimental results show that Transformer-based methods also achieves competitive performance, and the proposed GTAD achieves approximately no more than 1% improvement. It can be explained that this diagnostic task is relatively easy for all models, and the samples incorrectly classified with other methods is important in this situation. Therefore, I wonder if the authors can analyze these samples with the learned scales.
    2. The symbol K denotes both the convolution layers and the key of the self-attention module, which raises confusion.
    3. How to explain that the standard deviation of GTAD is less than other methods in the experiment of Cortical Thickness & \beta-Amyloid.
  • 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 idea is interesting and the experimental results also show its effectiveness. As mentioned above, I have some concerns that require the authors to provide a more detailed analysis in their rebuttal.

  • 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 #4

  • Please describe the contribution of the paper

    The proposed integrated framework offers a novel approach to guiding diffusion processes within a network. By employing a downstream transformer, the framework effectively combines both short- and long-range properties of graphs through the use of diffusion-kernel and multi-head attention mechanisms, respectively. the key contributions are Integration of short- and long-range information, diffusion-kernel mechanism, multi-Head attention and guidance of diffusion process.

  • 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 paper shows ability to effectively aggregate graph properties that achieve superior performance in graph classification tasks, and provide interpretable results.

  • 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 method is able to detect the key regions in distinguishing the progressions of neurodegenerative brain diseases through modality-wise attentions. However the clinical relevance and biological interpretation is lacking.

  • 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?

    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

    Authors should explain the biological significance of the obtained results (Figure.2 and Figure 3). Statistical tests may be conducted on distribution of attention scores across all brain regions with cortical thickness, β-Amyloid and FDG among CN, SMC and EMCI cases to support the GNN results.

  • 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?

    Novelty in methodology and a comprehensive analysis for preclinical Alzheimer classification. The paper is very well written.

  • 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



Author Feedback

We thank the reviewers for their constructive comments. We will address all the raised concerns to have the paper accepted. [Q] Reproducibility (All) [A] We will publicly release our code and pre-trained model upon acceptance. The ADNI data is publicly available but ROI-wise measures may slightly differ depending on pre-processing pipeline.

[Q] Lack of various analyses (R1,4) [A] We analyzed the critical influence of specific ROIs at the preclinical stage of Alzheimer’s disease using learned scales and attention scores in our paper. However, due to space limitations, we could not include all of our findings. We plan to extend this paper to a journal version to disseminate more detailed analysis including biological, statistical and clinical relevance.

[Q] Computing resource (R3) [A] One NVIDIA RTX A6000 GPU was used for the experiment. [Q] Why is the highest scale value 1.13? (R3) [A] The scale determines the magnitude of kernel convolution, and the ROI with the largest scale had 1.13. It means that the ROI required feature aggregation from the largest neighborhood to explain AD.

[Q] How to generate the input graph? (R3) [A] We ran probabilistic tractography using FSL and FreeSurfer to construct the structural brain networks. 160 regions by Destrieux atlas was defined based on the T1-weighted MR image. Then, we applied surface seed-based probabilistic fiber tractography to the DWI image to generate an anatomical connectivity matrix. For clarity, we will include MRI, DWI and PET as an input in the figure as well.

[Q] Typo, duplicated notation and lack of figure (R3,4,5) [A] We thank the reviewers for pointing out these issues. We will revise our paper accordingly and polish carefully.

[Q] Limitations of GTAD (R4) [A] The notable performance and enhanced interpretability of brain networks, particularly in scenarios involving multiple biomarkers, are indeed significant contributions of GTAD. There may be computational limitations, e.g., accurate kernel convolution requires eigendecomposition of Graph Laplacian (O(N^3) where N: number of nodes), but bypassing the diagonalization with polynomial approximation is well-known. Its clinical practicality may be a limit, as obtaining multi-modal imaging measures per subject is challenging in practice. To enhance understanding, we will make a small section discussing these limitations.

[Q] Motivation for GTAD (R4) [A] In medical analyses, data is often limited due to various reasons. To address this, integrating diverse modality information aims to capture a wide range of features. However, this approach may overlook ROI-wise characteristics in brain analyses. GTAD addresses these issues by learning a transformer-guided diffusion kernel that captures both local and global information simultaneously.

[Q] Why are performance improvements slight? (R5) [A] Transformer-based methods also achieve competitive performance. However, GTAD showed that, as the number of imaging modalities increases, the difference in performance between GTAD and the second-best method becomes more pronounced. Furthermore, unlike existing transformer-based works, GTAD captures local information by learned scales, enhancing the interpretability of brain networks. This allows GTAD to obtain ROI-wise characteristics across all samples, potentially enabling accurate classification of samples that other models may struggle to classify correctly.

[Q] Why are standard deviations of GTAD similar to baselines? (R5) [A] The ‘low standard deviations (std)’ in our paper was meant to highlight the stability of GTAD rather than to claim superiority over other models. As the reviewer probably knows, even if a model yields high accuracy, high std means that the model performance is not robust and may lead to high false-positives. The low std of GTAD, as well as other baselines, means that we thoroughly ran the experiments across different models and implies that the comparative evaluation remains both valid and meaningful.



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’

    Strong rebuttal and in agreement with meta reviewer 1

  • 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).

    Strong rebuttal and in agreement with meta reviewer 1



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’

    Post rebuttal, all reviewers agree that the paper should be accepted. Reviewers noted strengths include the interesting novel method, model interpretability, good experiments, and well-written paper.

  • 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).

    Post rebuttal, all reviewers agree that the paper should be accepted. Reviewers noted strengths include the interesting novel method, model interpretability, good experiments, and well-written paper.



back to top