List of Papers Browse by Subject Areas Author List
Abstract
Neurodegenerative disorders, notably Alzheimer’s Disease type Dementia (ADD), are recognized for their imprint on brain connectivity. Recent investigations employing Graph Neural Networks (GNNs) have demonstrated considerable promise in diagnosing ADD. Among the various GNN architectures, attention-based GNNs have gained prominence due to their capacity to emphasize diagnostically significant alterations in neural connectivity while suppressing irrelevant ones. Nevertheless, a notable limitation observed in attention-based GNNs pertains to the homogeneity of attention coefficients across different attention heads, suggesting a tendency for the GNN to overlook spatially localized critical alterations at the subnetwork scale (mesoscale). In response to this challenge, we propose a novel Disentangled Attention GNN (DAGNN) model
trained to discern attention coefficients across different heads. We show that DAGNN can generate uncorrelated latent representations across heads, potentially learning localized representations at mesoscale. We empirically show that these latent representations are superior to state-of-the-art GNN based representations in ADD diagnosis while providing insight to spatially localized changes in connectivity.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/2861_paper.pdf
SharedIt Link: https://rdcu.be/dV54h
SpringerLink (DOI): https://doi.org/10.1007/978-3-031-72117-5_21
Supplementary Material: N/A
Link to the Code Repository
https://github.com/gururgg/DAGNN
Link to the Dataset(s)
N/A
BibTex
@InProceedings{Gam_Disentangled_MICCAI2024,
author = { Gamgam, Gurur and Kabakcioglu, Alkan and Yüksel Dal, Demet and Acar, Burak},
title = { { Disentangled Attention Graph Neural Network for Alzheimer’s Disease Diagnosis } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
year = {2024},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15010},
month = {October},
page = {219 -- 228}
}
Reviews
Review #1
- Please describe the contribution of the paper
1.Propose a novel Disentangled Attention GNN model trained to response to the challenge that the traditional GNN may overlook spatially localized critical alterations at the subnetwork scale. 2.Design a disentanglement loss to generate spatially distinct attention coefficients across heads attention coefficients which may be similar in the traditional GNN. 3.Limite the study to discriminate the Mild Cognitive Impairment subjects from ADD subjects using sNETs to assess the capacity of DAGNN.
- 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.
A novel approach was employed to address the issue of homogenization across heads attention coefficients, using a disentanglement loss.
- 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 ablation experiments were conducted, and there is a lack of comparable methods. Additionally, it would be more convincing to test the effectiveness and generalizability of the proposed disentanglement loss on different models.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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
It is recommended that the author conduct more comprehensive experiments. Additionally, providing a more detailed exploration and analysis of the designed loss function, such as mathematical proofs or decomposition, would be beneficial.
- 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?
This work addresses the problem of homogenization across heads attention coefficients and proposes a simple yet effective method that can be generalized to other models. Unfortunately, the authors did not delve deeper into the topic, but I am highly interested in the future developments of this research.
- 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
Although I still believe the experiments in the paper could have been more thorough, the existing results are sufficient to support their conclusions. The supplementary explanation of the loss function also addressed my concerns.
Review #2
- Please describe the contribution of the paper
The author found that GNNs tend to overlook spatially local critical changes in the subnetwork scale, and proposed the DAGNN model with disentangled attention to address this issue.
- 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.
There is no significant innovation.
- 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 problem highlighted by the author regarding GNNs is a common issue in current GNN research, and it seems to be forcefully applied to the field of AD-related recognition. For MICCAI, I believe we should focus on addressing specific issues in the medical domain rather than general issues in the field of artificial intelligence.
(2) I think the Disentanglement Loss is the highlight of this paper, but I have some concerns. In equation (9), the value of λ has a huge impact on the experimental results. The manuscript should conduct a thorough ablation study on the setting of λ (including setting it to 0). The author set it to 0.1, which makes me concerned about whether the Disentanglement Loss is effective during gradient backpropagation.
(3) Have the same parcellation atlas (Destrieux atlas) been used in all comparative models for fair comparison?
(4) The font in Figure 1 could be enlarged to improve readability.
(5) The letter C in Figure 2 could be enlarged, and further explanation could be provided. Alternatively, the size of A could be reduced, and C could be placed in the first row, with B in the second row.
- 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
(1) The problem highlighted by the author regarding GNNs is a common issue in current GNN research, and it seems to be forcefully applied to the field of AD-related recognition. For MICCAI, I believe we should focus on addressing specific issues in the medical domain rather than general issues in the field of artificial intelligence.
(2) I think the Disentanglement Loss is the highlight of this paper, but I have some concerns. In equation (9), the value of λ has a huge impact on the experimental results. The manuscript should conduct a thorough ablation study on the setting of λ (including setting it to 0). The author set it to 0.1, which makes me concerned about whether the Disentanglement Loss is effective during gradient backpropagation.
(3) Have the same parcellation atlas (Destrieux atlas) been used in all comparative models for fair comparison?
(4) The font in Figure 1 could be enlarged to improve readability.
(5) The letter C in Figure 2 could be enlarged, and further explanation could be provided. Alternatively, the size of A could be reduced, and C could be placed in the first row, with B in the second row.
- 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 rationality of experiments, the innovation of the model, and the clinical application in medicine are the focal points of my consideration.
- 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 Reject — could be rejected, dependent on rebuttal (3)
- [Post rebuttal] Please justify your decision
Thank you very much to the author for partially addressing my concerns. However, I believe this work faces general issues in the field of artificial intelligence rather than specific problems in the medical field. Therefore, I maintain my original rating.
Review #3
- Please describe the contribution of the paper
This paper presents a new model, the Disentangled Attention Graph Neural Network (DAGNN), aimed at improving the diagnosis of Alzheimer’s Disease type Dementia (ADD) by leveraging graph neural networks. This model addresses the limitation of homogeneity in attention coefficients across different heads in attention-based GNNs. DAGNN introduces a disentanglement loss to generate spatially distinct attention coefficients, ensuring that each head contributes uniquely to the overall model output. This approach enhances the quality of the representations and provides better insights into localized changes in brain connectivity.
- 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 are:
-
New Model Design: DAGNN introduces a new approach to reduce redundancy across different attention heads in GNNs, enabling the model to capture more distinct patterns in brain connectivity data.
-
Demonstration of Clinical Feasibility: The model’s efficacy is demonstrated through superior performance on three diverse datasets, validating its effectiveness and potential for clinical use.
-
Strong Evaluation Metrics: The model undergoes rigorous testing using cross-validation and performance metrics, demonstrating its reliability and generalizability across different types of brain data.
-
- 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 the disentanglement approach in the DAGNN is novel, the underlying technology of graph neural networks (GNNs) and attention mechanisms are well-established in the field. For example, the basic framework of using GNNs for neurological analysis is not entirely new as seen in prior works like the application of Graph Convolutional Networks to brain networks (Kipf and Welling, ICLR 2017)
- 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 provide sufficient information for reproducibility.
- Do you have any additional comments regarding the paper’s reproducibility?
Although it is claimed the source code is available on Github, the link is not provided.
- 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 paper presents a novel model, the Disentangled Attention Graph Neural Network (DAGNN), designed to address Alzheimer’s Disease type Dementia (ADD) diagnosis by leveraging disentangled attention mechanisms within a graph neural network framework. The authors argue that their approach helps overcome the limitation of attention homogeneity across multiple heads in conventional attention-based GNNs, thus enabling the model to detect critical, localized changes in brain connectivity.
The use of disentanglement to generate spatially distinct attention coefficients is a significant step forward in this domain, potentially offering a more nuanced analysis of brain connectivity than existing models. The disentanglement concept is an interesting addition to the field of GNNs. But authors should highlight the unique contributions of DAGNN more clearly.
The paper does not also address how the DAGNN model integrates into existing clinical workflows nor does it discuss the scalability of the approach to larger, more heterogeneous populations. For clinical translation, it is vital to understand how such models can be deployed in typical healthcare systems.
The paper is generally well-written and organized. However, it could improve in clearly distinguishing the novel contributions of DAGNN from existing techniques in the introduction and literature review sections. This would help in immediately capturing the reader’s attention to the unique aspects of the research.
Although it mentions the code or data will be made available on Github, it does not provide the link. The details on the preprocessing steps for the structural network data should be also elaborated, which are crucial for replicating the study results.
More details on how DAGNN could be integrated into clinical settings, including potential challenges and solutions, would be valuable. This discussion could include aspects like model training time, computational requirements, and compatibility with existing diagnostic tools.
Given the potential of the proposed DAGNN model in improving the diagnostic processes for Alzheimer’s Disease through an innovative use of disentangled attention mechanisms, I would recommend a weak accept, pending a strong rebuttal that addresses the concerns regarding innovation and clinical applicability. The novelty of the disentanglement approach needs to be better highlighted and differentiated from existing works, and more information is needed on the practical deployment of the model in real-world settings.
- 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 innovative aspects of the Disentangled Attention Graph Neural Network (DAGNN) presented in the paper. The novel use of disentangled attention mechanisms to address the issue of attention homogeneity across multiple heads in graph neural networks is particularly compelling. This approach not only enhances the model’s ability to capture localized, critical changes in brain connectivity but also represents a meaningful advancement over traditional attention-based models. Potential Clinical Impact and the quality and clarity of the paper presentation convinced me for my recommendation. If the authors can convincingly address the weaknesses in their rebuttal, the paper would represent a valuable contribution to the MICCAI proceedings.
- 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 sincerely thank all reviewers for their time and effort in reviewing our manuscript. First, we reiterate our contributions: (1) We introduce DAGNN, a novel attention-based model enhanced by a disentanglement loss, designed to identify spatially distinct patterns within brain networks relevant to ADD/MCI classification. (2) Unlike traditional AGNN, DAGNN can learn decorrelated latent representations localized at the mesoscale. (3) DAGNN’s ability to generate spatially distinct patterns enhances its effectiveness and interpretability, making it a potentially valuable tool for other clinical applications involving brain connectomes. Below, we respond to reviewers’ comments. Experiments: R1 suggests conducting more ablation experiments and comparisons with other models. We believe our presented comparisons with models tailored for brain networks (BRAINGB, BRAINGNN), disentanglement-based methods (DisenGCN, FactorGCN*) and AGNN (an ablated version of DAGNN without the disentanglement loss term, i.e., λ=0) provide substantial evidence for the effectiveness of our approach. Additionally, we confirm that the same parcellation atlas was used for all datasets, addressing a concern raised by R3.
Disentanglement Loss: In response to all reviewers’ comments, we would like to elaborate on the disentanglement loss. We propose DAGNN as an improvement on AGNN which partially suffers from the homogeneity problem across heads. While this is a common issue in AI research that extends beyond the medical field (as noted by R3), our proposed resolution is specifically motivated by the desire to identify distinct and possibly overlapping brain regions whose degradation may reflect different aspects of disease progression. That said, as a novel solution to the homogeneity problem, DAGNN may well be useful in other domains. Note that, the loss function is simple yet more effective compared to DisenGCN, which uses the neighbor-routing mechanism, and FactorGCN*, which adds a classification head for disentanglement in latent space. By directly performing disentanglement in coefficient space, it achieves more differentiated latent representations in comparison, yielding superior experimental results. Also, DAGNN utilizes an L1-norm-based loss (well-bounded and promotes sparsity) and does not introduce a new head or routing mechanism, which makes it computationally efficient. R3 raised concerns about the effectiveness of the disentanglement term in the loss function. The coefficient λ=0.1 was indeed determined empirically. Although the details of this hyper-parameter optimization were omitted, its contribution to the performance is evident from our comparison with AGNN. Note that, AGNN is DAGNN without the disentanglement loss. Generalizability of the proposed disentanglement loss (a reservation of R1) to other multi-headed attention models, such as transformer-based AGNNs, is straightforward. Its effectiveness in different settings is a question we hope to address in follow-up projects motivated by the present proof-of-concept study.
Clinical applicability: R3 and R4 raised concerns about integrating our approach into clinical settings.Figures 2.A and 2.C attest to the promise of DAGNN as a biomarker for diagnosing AD and, in the future, possibly other neurodegenerative diseases that alter brain connectivity. Additionally, the distinct spatial patterns (subnetworks) identified may offer deeper insights into underlying mechanisms from a connectionist perspective.
Reproducibility: The preprocessing stage of structural network construction is crucial, as noted by R4. While we summarized our custom preprocessing pipeline in the paper, further details can be found in the paper’s GitHub page which will be made available in due course. We regret that it had to be left out due to the anonymity requirement during submission.
Figures. We thank R3 for the figure corrections/suggestions. They have been implemented.
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’
The majority of the reviewers acknowledged the novelty of the disentangled attention mechanisms. The rebuttal largely resolves the reviewers’ concerns. In the final version, the authors should clearly describe how they chose subjects from ADNI3, which contained more subjects.
- 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).
The majority of the reviewers acknowledged the novelty of the disentangled attention mechanisms. The rebuttal largely resolves the reviewers’ concerns. In the final version, the authors should clearly describe how they chose subjects from ADNI3, which contained more subjects.
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’
The authors were able to reply the reviewers comments
- 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).
The authors were able to reply the reviewers comments