List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Yilin Leng, Wenju Cui, Chen Bai, Zirui Chen, Yanyan Zheng, Jian Zheng
Abstract
Constructing structural brain networks using T1-weighted MRI (T1-MRI) presents a significant challenge due to the lack of direct regional connectivity. Current methods with T1-MRI rely on predefined regions or isolated pretrained modules to localize atrophy regions, which neglects individual specificity. Besides, existing methods capture global structural context only on the whole-image-level, which weaken correlation between regions and the hierarchical distribution nature of brain structure. We hereby propose a novel dynamic structural brain network construction method based on T1-MRI, which can dynamically localize critical regions and constrain the hierarchical distribution among them. Specifically, we first cluster spatially-correlated channel and generate several critical brain regions as prototypes. Then, we introduce a contrastive loss function to constrain the prototypes distribution, which embed the hierarchical brain semantic structure into the latent space. Self-attention and GCN are then used to dynamically construct hierarchical correlations of critical regions for brain network and explore the correlation, respectively. Our method is trained on ADNI-1 and tested on ADNI-2 databases for mild cognitive impairment (MCI) conversion prediction, and acheive the state-of-the-art (SOTA) performance.
Link to paper
DOI: https://doi.org/10.1007/978-3-031-43993-3_12
SharedIt: https://rdcu.be/dnwNd
Link to the code repository
https://github.com/Leng-10/DH-ProGCN
Link to the dataset(s)
Reviews
Review #1
- Please describe the contribution of the paper
In this paper, the authors intend to propose the DH-ProGCN method which can dynamically construct disease-related dtructural brain network based on T1-MRI, which starts with a prototype learning method for finding the critical brain regions and a contrastive loss function for region embedding, then the self-attention mechanism with GCN for finding their correlations and conduct classification task.
- 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 idea of conducting brain network analysis using T1-MRI can have merits in the clinical scenario. The methodology seems to be novel, and the classification performance has demonstrated its effectiveness.
- 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 claim of “dynamic” in the brain network graph construction, and the “hierarchical” in the prototype learning are not clear enough to me, as I haven’t found any hierarhical or dynamic strucutres in Fig.1 or descriptions in the methodology section.
- 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 authors claim to have the source code in github, and use the public dataset ADNI which I think is eligible for reproducing.
- 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
Experiments haven’t mentioned if it is cross-validation, which is important and can demonstrate if the method’s performance is robust or not. More experiments are also recommended for NC-MCI classification, which also plays an important role for early-stage diagnosis for AD, and is also considered a harder task than the MCI classification. It is better to show more details about the scores for the ablation studies, for better evaluations of the proposed mechanisms in the proposed method, which can be added in the supplematary material.
- 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?
This is a complete work with no significant flaws, and have novelties which I think can bring merits to the communities.
- Reviewer confidence
Very confident
- [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
A dynamic structural brain network construction method based on T1-MRI for diagnosing 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.
A novel dynamic structural brain network construction method named hierarchical prototypes embedding GCN (DH-ProGCN) to dynamically construct disease-related structural brain network based on T1-MRI. And this method has been validated in the ADNI database with excellent 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.
1.As is well known, T1 MRI is structural imaging and does not contain dynamic information. Why do authors need to learn dynamic brain networks and capture dynamic information. 2.The introduction section needs to be taken seriously, as some explanations are incorrect and references are outdated.
- 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 author has published some parameter settings, but the code has not been fully disclosed, making it difficult to determine 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/2023/en/REVIEWER-GUIDELINES.html
Please provide the scientific basis for capturing dynamic information from the structure T1-MRI.
- 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?
As is well known, T1 MRI is structural imaging and does not contain dynamic information. Why do authors need to learn dynamic brain networks and capture dynamic information.
- Reviewer confidence
Very confident
- [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 #3
- Please describe the contribution of the paper
The author proposes a dynamic hierarchical graph for disease classification based on T1-MRI. Self-attention as well as GNN are implemented.
- 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 outperformed the competing methods considered in the work. The authors performed the ablation study from different perspectives.
- 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 optimazation process is not stated clearly.
- The GNN applied in the framework is confusing.
- The biomarker explanation is not clear.
- 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 appraoch is reproducible, although some parameter settings are ignored.
- 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
-
Concerning the node definition, the feature maps are generated by a CNN, with each channel corresponding to the response of a distinct brain region. However, it is unclear how the peak response location and image clustering are determined. Additionally, since the feature maps are mixed by the CNN backbone and do not represent raw regions, it is unclear what the obtained regions are representing.
-
The adjacent matrix is built using Q, K, and V. It appears that multiplying by V may not be necessary.
-
The use of GNN in the framework is unclear. The author built hierarchical edges through self-attention and clustered channels into nodes. Why not use a Transformer framework instead? It seems that a Transformer would be a more suitable choice.
-
It is not clear how Figure 3 was obtained from the hierarchical critical regions.
-
- 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 proposed method has some merits, also some points needed to be claimed.
- 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.
Strength
- This paper proposes a novel brain network construction method (DH-ProGCN) to dynamically model disease-related structural brain network based on T1 MRIs. The paper is well-organized.
- The authors perform extensive ablation studies from different perspectives.
Weakness
- The claim of “dynamic” in the brain network graph construction is not clear enough. We cannot find any dynamic structures in Figure 1 or descriptions in the methodology section.
- It is unclear how the peak response location and image clustering are determined. In addition, since the feature maps are mixed by the CNN backbone and do not represent raw regions, it is unclear what the obtained regions are representing.
- The visualization and interpretation of attention score A is welcome.
- Experiments haven’t mentioned if it is cross-validation, and no std (CI) is provided.
- It is not clear how Figure 3 is obtained from the hierarchical critical regions.
- Some method details are confusing, e.g., “[tix, tiy, tiz] represents the peak response coordinate of the i-th image” seems that each image only has one peak response.
Author Feedback
Thanks for your valuable comments on our submission. We appreciate the time and effort you dedicated to reviewing our work. In this rebuttal, we wish to address your concerns. 1.For Q (question) 1 from AC (Area Chair), R (Reviewer) 1 and R2, we will clarify the description of “dynamic” below. We conclude the methods section with an elaboration on “dynamic”: “Notably, as prototypes are dynamic, constructed brain network graphs are also dynamic, rather than predefined and fixed” with the aim of “providing a more personalise brain network representation for every subject”. The “dynamic” involves personalized identification of critical regions and the dynamic determination of edge features. In other words, the proposed brain network is a sample-wise dynamic model that processes each sample using data-dependent architectures and parameters. This is distinct from common dynamic fMRI-based brain networks which operate in a temporal-wise dynamic manner. Further, our approach differs from previous methods that rely on pre-defined connectivity relationships. 2.For Q2, Q6 from AC and Q1 from R3, we apologize for any lack of clarity in the clustering process description, especially the peak response. We first get convolution features of input images from backbone, where the channel number of features is C. As each channel can correspond to a certain type of visual patterns, we propose a channel clustering module to cluster spatially-correlated subtle patterns as compact and discriminative parts from a group of channels whose peak responses appear in neighboring location. Intuitively, each channel can be represented as position vector whose elements are coordinates from peak responses over all training images, where [txi, tyi, tzi] is the coordinate of the peak response of the i-th channel. In optimization process, we clustered coordinates of channels by K-means to initialize clustering results. We obtain multi-hierarchy clustering by repeating the clustering centers obtained from the last clustering. For the convenience of training, the clustering process is approximated by two fully connected layers replacing K-means. In this way, a group of channels whose peak response appear in neighboring locations are clustered together by clustering their coordinates. 3.For Q3 and Q5 from AC and Q4 from R3, we perform some brief elaborations. Figure 3 shows visualization results based on graphs before GNN. We first use peak response coordinates of cluster centers as node coordinates which represents the coordinates of critical regions, then correlation matrix A between nodes is obtained by self-attention as edge features. Finally, node coordinates and edge features are inputed into BrainNet Viewer software to generate Figure 3. The thickness of edges represents the correlation coefficient between nodes, i.e., the connected strength between brain regions. The higher correlation coefficient of A, the thicker the edge. Nodes belonging to the same parent node are assigned the same color. The size of nodes represents its hierarchy, the higher the hierarchy the bigger the node. 4.For Q4 from AC, Q2 from R1, and weakness from R3, are about experiments. The validation way is described at dataset section. To prove the robustness, we train the model on ADNI1 and validat it on independent ADNI2 dataset following SOTA works we compared in section 4.1. 5.For Q2 and Q3 from R3, we have two reasons for using GNNs instead of transformers. Firstly, the objective of our work is to construct brain network graphs, converting euclidean geometric image into a graph structure. GNN is well-suited for processing graph, making it a natural choice for our task. Secondly, we utilize self-attention only to generate edge features, although both GNNs and transformers can capture long-range dependencies, GNNs have the strength of learning underlying topological information which has not been extensively explored in related works of transformers.
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.
Most of the major concerns (such as clarification of “dynamic” and description of peak response) have been well addressed.
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.
I agree with the AC and the reviewers that the term of “dynamic” is very confusing. After read the paper, it is not clear what is the practical value of this dynamic brain network. It is also not clear what performance this dynamic brain network can achieve but traditional method can not. There is significant missing details in the method. I recommend to reject.
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.
Although the authors tried to address all the concerns, the key question “the description of dynamic” is still not clear. Brain network is of course personalized. If the process of building such a network is different across subjects, how can we compare brain network among people? Therefore, I recommend reject.