List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Houliang Zhou, Yu Zhang, Brian Y. Chen, Li Shen, Lifang He
Abstract
The interconnected quality of brain regions in neurological disease has immense importance for the development of biomarkers and diagnostics. While Graph Convolutional Network (GCN) methods are fundamentally compatible with discovering the connected role of brain regions in disease, current methods apply limited consideration for node features and their connectivity in brain network analysis. In this paper, we propose a sparse interpretable GCN framework (SGCN) for the identification and classification of Alzheimer’s disease (AD) using brain imaging data with multiple modalities. SGCN applies an attention mechanism with sparsity to identify the most discriminative subgraph structure and important node features for the detection of AD. The model learns the sparse importance probabilities for each node feature and edge with entropy, L1, and mutual information regularization. We then utilized this information to find signature regions of interest (ROIs), and emphasize the disease-specific brain network connections by detecting the significant difference of connectives between regions in healthy control (HC), and AD groups. We evaluated SGCN on the ADNI database with imaging data from three modalities, including VBM-MRI, FDG-PET, and AV45-PET, and observed that the important probabilities it learned are effective for disease status identification and the sparse interpretability of disease-specific ROI features and connections. The salient ROIs detected and the most discriminative network connections interpreted by our method show a high correspondence with previous neuroimaging evidence associated with AD.
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16452-1_45
SharedIt: https://rdcu.be/cVVp0
Link to the code repository
N/A
Link to the dataset(s)
N/A
Reviews
Review #1
- Please describe the contribution of the paper
This paper presents a (GCN) Graph Neural Network-based approach for classification and detection of Alzheimer’s disease from multi-modalities of brain images. The authors propose the sparsity to capture subgraph representations attending on most important features for better discrimination between AD and healthy groups. Ablations and experiments in the paper show that the SGCN method provides better classification than related work when using multi-model data.
- 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.
- Applying Graph convolution networks (GCN) on multi-modal neural-images is an interesting idea as it shows evidence that a single modality of images is not sufficient.
- Imposing sparsity in GCNs (SGCN) as attention shows that the method is better than the conventional GCN.
- The experiments show performance metrics higher than related works.
- 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.
- Sparse representation in GCN is not a novel contribution. The SGCN method has been used widely on other computer vision applications, which makes the novelty of the approach moderate. Few examples are: Renjie Wu et al. 2018 “k3-Sparse Graph Convolutional Networks for Face Recognition” and Liushuai Shi et al. 2021 “SGCN:Sparse Graph Convolution Network for Pedestrian Trajectory Prediction”. See also the survey in David Ahmedt-Aristizabal et al. 2021. “Graph-Based Deep Learning for Medical” Diagnosis and Analysis: Past, Present and Future
- Some items of the loss function and mathematical notions are not clear.
- There is no explanation of the architecture and implementations.
- Please rate the clarity and organization of this paper
Satisfactory
- 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 paper includes analysis of results and performance and the dataset used are cited. However, the paper lacks implementation details for reproducibility.
- 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/2022/en/REVIEWER-GUIDELINES.html
- It is unclear why the loss in EQ. 3. is called mutual information. In my knowledge, the mutual information describes divergence between two probabilities. EQ. 3 looks more like a conditional probability based on joint attention matrices. Could the authors clarify more this ambiguity?
- The regularization/cross-entropy loss in EQ. 5 need to be checked. What are the ground truth and predicted variables?
- Please check the mathematical notation in section 2.2. n and N look referring to the same thing.
- The authors have chosen K=10 but they didn’t justify this choice or discuss the effect of K smaller or larger on the performance of the GCN model.
- The authors might also compare their method with existing Graph attention-based convolution networks which are close to SGCNs.
- The presentation has some English typos to be checked in (e.g., section 2.3).
- 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 contribution in terms of originality of the approach and its significance are moderate. Sparsity in GCNs is not a new framework/approach. I am not confident about the efficiency and robustness of the approach since the paper doesn’t show clear evidence about time complexity of the SGCN computations and the impact of K (in KNN) on sparse matrices. My rating could change if the authors clarify my concernes.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
2
- 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
6
- [Post rebuttal] Please justify your decision
Based on all reviewers comments and the responses of the authors I decided to increase my rating to accept. The authors have answered my concerns and clarified the differences between their contribution and existing methods.
Review #2
- Please describe the contribution of the paper
In this paper, a sparse interpretable GCN framework (SGCN) for the iden- tification and classification of Alzheimer’s disease (AD) is proposed using brain imag- ing data with multiple modalities.
- 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.
Proposed method applies an attention mechanism with sparsity to identify the most discriminative subgraph structure and important node features for the detection of AD. The model learns the sparse importance probabilities for each node feature and edge with entropy, l1 , and mutual information regularization.
- 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.
Comparison of proposed method need to be done with existing state of the art methods for multimodal disease diagnosis
- 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
Publically available data is being used for evaluation.
- 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/2022/en/REVIEWER-GUIDELINES.html
Dataset sample size in each of the modality is less in number. How to make sure the convergence of proposed method?
Edges in functionality connectivity are estimated based on weighting the Gaussian similarity function of Euclidean distance. What is the intuition of selecting this metric for evaluation.
More detail about mutual information loss is needed.
Why the interpretability is being called as sparse. How the proposed method make sure that the sparsity is being imposed on interpretable coefficients?
Interpretation results need to be discussed in more details.
Grammetical errors need to be checked throughout.
- 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 proposing interpretable model for Multi-Modal Diagnosis of Alzheimer’s Disease. Proposed method utilizes Graph Convolutional Network for doing the classification task along with obtaining interpretation on input features.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
1
- Reviewer confidence
Very confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
Not Answered
- [Post rebuttal] Please justify your decision
Not Answered
Review #3
- Please describe the contribution of the paper
This paper proposed a sparse interpretable GCN framework (SGCN) for the identification and classification of Alzheimer’s disease (AD) using brain imaging data with multiple modalities.
- 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 proposed a novel sparse interpretable GCN framework (SGCN), which applies an attention mechanism with sparsity to identify the most discriminative subgraph structure and important node features for the detection 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.
Correlation analysis between multimodal data is slightly lacking.
- 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
The ideas and experiments described in this paper are quite clear, should be repeatable.
- 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/2022/en/REVIEWER-GUIDELINES.html
The research work in this paper is innovative and of practical significance. It will be better if the following questions can be considered.
- Since the work in this paper is based on multimodal data, it would be better to analyze or learn the relationship between multimodal data more deeply.
- Several more datasets can be used to test for greater persuasiveness.
- 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?
The work of this paper has good innovation and practical significance, and it is recommended to accept.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
1
- 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
Not Answered
- [Post rebuttal] Please justify your decision
Not Answered
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.
- The paper proposed a Graph Neural Network approach for detecting AD from Normal on multi-modal data
- There are concern about the novelty of the paper. Similar appraoch has been proposed in the past as suggested by one of the reviewers. It is OK if the method proposed in not novel but then the emphasis of the paper should be shifted toward what new insight can be gained that was not possible before. Otherwise, it is just applicaiton of new method on a well studied dataset
- Also in Alzheimer’s disease the main question in not really about AD vs HC but rather it is about MCI and who converts to AD and how and when. Of course, this paper is not about longtitudinal data, but I recommend authors to expand their work beyond classification to address real clinical needs
- The relationship between multimodal data is not well studied (beyond ablation study of the effect of loss).
- There are several dataset for ADNI that are publically available, focusing on one dataset is one of the limitations of the paper
- Some minor comments:
- detail of mutual information
- “How the proposed method make sure that the sparsity is being imposed on interpretable coefficients?”
- What is the ranking of this paper in your stack? Use a number between 1 (best paper in your stack) and 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).
15
Author Feedback
Responses to Meta-Reviewer and common questions:
Q1: There are concern about the novelty of the paper. Similar approach has been proposed in the past. A1: Our proposed method is novel, because it aims to discover informative subgraphs by jointly learning sparse patterns of brain multi-modality node features and connectivity for individual subject, which is from EQ.2-5. Our model is also unique on the design of sparse mechanism in comparison with other methods. First, most of the existing methods have not focused on the AD diagnosis and the interpretation. Second, those methods only tried to identify the sparse connectivity mask on the adjacency matrix. In contrast, our method can go further to interpret the important sparse patterns on node features and connectivity.
Q2: The relationship between multimodal data is not well studied. A2: From Table 1, we have demonstrated the multimodality data can provide the complimentary information to improve the prediction performance. In Fig.1, we showed the different modalities share some common and modality-specific salient ROIs. In Fig.3, we also showed the most discriminative connections in neural systems between different modalities.
Q3: Detail of mutual information. A3: In EQ.3, the mutual information quantifies the probability of prediction y when the input graph to the GCN model is limited to the explanatory graph Gs. The intuition behind comes from the traditional forward propagation based methods for the whitebox explanation [1]. [1] Piotr Dabkowski and Yarin Gal. Real time image saliency for black box classifiers. In NeurIPS, pages 6967–6976, 2017.
Q4: Why the interpretability is being called as sparse. How the sparsity is imposed on interpretable coefficients? A4: In EQ.3, we use the mutual information loss to determine the important probabilities to find the important subgraph. In EQ.4 and EQ.5, we apply the L1 and entropy regularization loss to induce the sparsity on those important probabilities. This helps highlight the discriminative brain regions and connections that are important for the diagnosis, thereby improving the biomarker interpretation.
Responses to Reviewer#1:
Q1: Sparse representation in GCN is not novel. A1: Please refer to A1 to Meta-Reviewer.
Q2: There is no explanation of the architecture and implementations. A2: In our work, our model learns the important probabilities on nodes and edges in EQ.2, which are the input. These probabilities will be multiplied with the feature and adjacency matrix to be fed into three GCN layers and an MLP layer to classify.
Q3: Detail of mutual information in EQ. 3. A3: Please refer to A3 to Meta-Reviewer.
Q4: The cross-entropy loss in EQ. 5 need to be checked. A4: In EQ.5, it’s the entropy regularization loss, not the cross-entropy loss. We use it to induce sparsity on those important probabilities.
Q5: The time complexity of the SGCN and the impact of K in KNN. A5: The time complexity of SGCN is similar to the standard GCN. The only difference is SGCN needs to learn two probabilities parameters PX and PA in EQ.2 and optimize its loss function. In our experiment, we selected K from [3,5,7,…,50] in KNN. The results showed that the smaller K is not enough to exploit the intrinsic neighborhood structure to identify the AD, and the larger K brings noisy information to affect the performance.
Responses to Reviewer#2:
Q1: What is the intuition of selecting the Gaussian similarity metric. A1: The Gaussian similarity has been widely used to build the graphs in the GCN model. We follow this standard way to construct the graphs and measure the node correlation strength.
Q2: Detail about mutual information loss. A2: Please refer to A3 to Meta-Reviewer.
Responses to Reviewer#3:
Q1: More datasets can be used to test for greater persuasiveness. A1: We agree it’s important to use more datasets to further test the generalization ability of our model. We plan to use OASIS and PPMI cohorts to test it in our future work.
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.
- I am still not sure what new insight or unmet clinical need is addressed by this architecture, hence not sure this method is actually tested for a use-case
- In my vote is borderline toward reject however the reviewer changed their vote and I respect that
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- 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).
na
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 further clarified the novelty of the proposed joint learning of sparse representation across multimodal data over GCN and addressed other critiques from reviewers. Overall the proposed method presents sufficient novelty. All answered responses in the rebuttal should be included in the final manuscript.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- 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).
1
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.
This paper proposed a sparse interpretation of GNN using multimodal neuroimagines for Alzheimer’s Disease vs. healthy control classification. While all reviewers agree that that paper is acceptable, I do think there is a major concern that the real clinical needs are not addressed in this paper, i.e., the prediction of MCI conversion into AD, how and when. The authors did not answer this question in the feedback. Despite this limitation, the authors have addressed other concerns raised by reviewers, so I recommend accept this paper with the suggestion that authors can address this limitation in discussion.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- 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).
5