List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Hongmin Cai, Zhixuan Zhou, Defu Yang, Guorong Wu, Jiazhou Chen
Abstract
Previous studies have shown that neurodegenerative diseases, specifically Alzheimer’s disease (AD), primarily affect brain network function due to neuropathological burdens that spread throughout the network, similar to prion-like propagation. Therefore, identifying brain network alterations is crucial in understanding the pathophysiological mechanism of AD progression. Although recent graph neural network (GNN) analyses have provided promising results for early AD diagnosis, current methods do not account for the unique topological properties and high order information in complex brain networks. To address this, we propose a brain network-tailored hypergraph neural network (BrainHGNN) to identify the propagation patterns of neuropathological events in AD. Our BrainHGNN approach constructs a hypergraph using region of interest (ROI) position encoding and random-walk-based sampling strategy, preserving the unique identities of brain regions and characterizing the intrinsic properties of the brain-network organization. We then propose a self-learned weighted hypergraph convolution to iteratively update node and hyperedge messages and identify AD-related propagation patterns. We conducted extensive experiments on ADNI data, demonstrating that our BrainHGNN outperforms other state-of-the-art methods in classification performance and identifies significant propagation patterns with discriminative differences in group comparisons.
Link to paper
DOI: https://doi.org/10.1007/978-3-031-43904-9_23
SharedIt: https://rdcu.be/dnwG1
Link to the code repository
N/A
Link to the dataset(s)
N/A
Reviews
Review #1
- Please describe the contribution of the paper
The author proposed a hypergraph neural network to represent the brain network. Then, random walk is used to generate hyperedges and the GNN method is used to update the information of entity and edge. The goal is to identification of propagation patterns of neuropathology and early diagnosis of AD.
- 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 random walk is used to generate the hyperedge. It is very interesting, which can consider the different field of brain and construct the lcoal hyperedge and global hyperedge. It is very useful in this work.
- The propagation patterns can be fund. Which is usefule for researchers. It can help us to research the main reasion of AD.
- 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 generation of hypeedge is not clear. How to utilize the random walk to handle this problem. I hope the author can provide more information to illustrate the problem.
- I do not find experiment to evoluate the performance of propagation patterns prediction. 3.3 may be the related experiment. However, it lacks the related analysis.
- Fig 1i lacks the illustractions. The caption should be added more information to introdduce the framework. It is very help for readers.
- 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
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: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html
- The generation of hypeedge is not clear. How to utilize the random walk to handle this problem. I hope the author can provide more information to illustrate the problem.
- I do not find experiment to evoluate the performance of propagation patterns prediction. 3.3 may be the related experiment. However, it lacks the related analysis.
- Fig 1i lacks the illustractions. The caption should be added more information to introdduce the framework. It is very help for readers.
- 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?
n/a
- 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 #2
- Please describe the contribution of the paper
The paper presents a novel graph neural network for the classification of Alzheimer’s Disease. With the proposed hyperedge generation strategy and ROI position encoding module, the proposed method achieves reasonable improvements comparing with traditional GCN.
- 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 is well organized and the written is clear.
- The motivation is clear and the innovation is reasonable.
- 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 performance is not good enough. With PET image, the classification accuracy is usually above 80% for NC vs. MCI, above 70% for EMCI vs. LMCI. The performance of the proposed method is worse than many SOTA methods, and the less than 70% accuracy makes the analysis about propagation pattern questionable.
- Please rate the clarity and organization of this paper
Excellent
- 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
The data is from a public dataset and the details for reproduction is provided.
- 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 results of SOTA methods (including GCN and CNN) should be provided to give the readers information about the current situation about AD diagnosis.
- The reason why the proposed method is worse than SOTA methods should be discussed.
- 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?
The unsatisfactory performance
- Reviewer confidence
Very confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
4
- [Post rebuttal] Please justify your decision
The propoesed method may be better than other ROI socre based methods, but still far worse than the image based methods, raising doubts about such methodology. Especially with the limited accuracy (less than 70% in most scenarios), its main advantage of providing insight about the connection of different brain ROIs is unconvincing. Exploring such connection through image based network may be a better solution.
Review #3
- Please describe the contribution of the paper
A hypergraph neural network is applied for improving connection information in brain networks, and a two-stage message passing is proposed. The approach is reasonable.
- 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 hypergraph neural network is proposed explicitly tailored for brain networks. Statistical results are provided. A two-stage message passing is introduced.
- 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 compared approaches are out-of-date. Some nessessary methods are needed, such as BrainGNN, dynamic GNN.
The auther states that each hyperedge is actually a subset of V randomly sampled based on ¯A. This is still not clear how to sample on a graph to obtain a new edge in detail.
- 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
The detailed parameter settings are listed. Apart from biomarker interpretation part, the method is reproducible.
- 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 compared approaches are out-of-date. Some nessessary methods are needed, such as BrainGNN, dynamic GNN.
The auther states that each hyperedge is actually a subset of V randomly sampled based on ¯A. This is still not clear how to sample on a graph to obtain a new edge in detail.
- 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
5
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper is well organization and well written.
- 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
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.
Strengths:The author proposed a hypergraph neural network to represent the brain network. Then, random walk is used to generate hyperedges and the GNN method is used to update the information of entity and edge to diagnose AD. Weaknesses: The generation of hyperedge is not clear, how to utilize the random walk to handle this problem; The compared approaches are out-of-date, some nessessary methods are needed, such as BrainGNN, dynamic GNN; We do not find experiment to evoluate the performance of propagation patterns prediction; The auther states that each hyperedge is actually a subset of V randomly sampled based on ¯A. This is still not clear how to sample on a graph to obtain a new edge in detail.
Author Feedback
We appreciate the reviewers’ enthusiasm for our work on discovering brain network dysfunction in Alzheimer’s disease using hypergraph neural network. The main concerns are 1) unclear illustration of hyperedge generation and 2) lack of comparison with other SOTA methods. We first address these major questions and then the specific questions from each reviewer. Major question 1: Unclear illustration about the generation of hyperedge Due to the page limit, we only provide the main idea of hyperedge generation based on second-order biased random walk sampling. It mainly has three steps: 1) Calculate the transition probability matrix with the original edge weight matrix and parameter p according to Eq.(2). 2) For each node, set it as the beginning of the walk, and calculate which neighbor to be traversed according to transition probability. Then we sample this neighbor to the hyperedge and set this neighbor as the current node. Repeat this procedure L times to generate a hyperedge. Iterate all nodes to construct a hypergraph. 3) Execute steps 1-2 for local sampling and global sampling to generate local and global hypergraphs respectively. Fuse two hypergraphs by concatenation to form the final hypergraph. Major question 2: Lack of experiments compared with other SOTA methods We thank you for the constructive comments. We conducted experiments on stratifying CN, EMCI, and LMCI groups using our BrainHGNN and two SOTA approaches (BrainGNN MedIA’2021 and DynamicEdgeConv CVPR’2019) on amyloid-PET and FDG-PET datasets. For amyloid-PET data, BrainHGNN (0.751) achieved higher accuracy than BrainGNN (0.745) and DynamicEdgeConv (0.735) in CN/LMCI comparison. BrainHGNN (0.695) also had superior power in classifying EMCI and LMCI groups, compared to BrainGNN (0.668) and DynamicEdgeConv (0.672). In CN/EMCI comparison, BrainHGNN (0.647) outperformed BrainGNN (0.639) and DynamicEdgeConv (0.634). For FDG-PET data, the experiments showed similar results, implying that BrainHGNN achieves better performance than the other two SOTA methods on both datasets. We will add their performance in the final version. #R1# Q1: Section 3.3 lacks a related analysis of propagation patterns prediction. -: Since AD-related propagation patterns are unknown, we cannot quantitatively analyze our results. We qualitatively discuss the patterns identified by BrainHGNN. Here are some analyses of section 3.3. In this experiment, first, we repeat 10-fold cross-validation 30 times for amyloid-PET (or FDG-PET) data in CN/LMCI comparison. Second, for each of the 300 experiments, we record 10 hyperedges with the highest edge weight. Third, we count the occurrence frequency of each hyperedge in the top-10 list over the 300 experiments. Finally, we choose 4 hyperedges with the highest frequency as potential AD-related propagation patterns. We find that most nodes in these hyperedges are in the AD-related subnetworks, indicating that BrainHGNN can identify neuropathology propagation patterns in AD. Q2: Fig 1 lacks illustrations. -: We will provide detailed illustrations for the framework in the final version. #R2# Q1: The reason why the proposed method is worse than SOTA methods. -: Thank you for this constructive comment. We would respectfully disagree for two reasons. (1) Our input data is a vector of SUVR scores of ROIs from PET image, not the whole PET image. So it is unfair to compare our method with image-based deep learning methods (e.g., CNN). To further demonstrate our method’s performance, we compare BrainHGNN with two SOTA approaches. The results show that BrainHGNN still achieves the best classification performance in major question 2. (2) Another major contribution of BrainHGNN is its ability to discover brain network dysfunction in AD. Results show that hyperedges identified by BrainHGNN are significantly associated with AD-related subnetworks. #R3# The answers to Q1 and Q2 are illustrated in major questions. Hope that our answers can address your questions.
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.
The novelty is good and the rebuttall addresses the main concerns regarding 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.
The rebuttal significantly improved the paper quality. It is suggested that the authors incorporated them in their camera ready paper.
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 authors proposed a hypergraph neural network explicitly tailored for brain networks. Although more details were added in rebuttal, there are still concerns on the clarity and performance of the study.