List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Fatih Said Duran, Abdurrahman Beyaz, Islem Rekik
Abstract
A connectional brain template (CBT) is a normalized representation of a population of brain multigraphs, where two anatomical regions of interest (ROIs) are connected by multiple edges. Each edge captures a particular type of interaction between pairs of ROIs (e.g., structural/functional). Learning a well-centered and representative CBT of a particular brain multigraph population (e.g., healthy or atypical) is a means of modeling complex and varying ROI interactions in a holistic manner. Existing methods generate CBTs by locally integrating heterogeneous multi-edge attributes (e.g., weights and features). However, such methods are agnostic to brain network modularity as they ignore the hierarchical structure of neural interactions. Furthermore, they only perform node-level integration at the individual level without learning the multigraph representation at the group level in a layer-wise manner. To address these limitations, we propose Dual Hiearchical Integration Network (dual-HINet) for connectional brain template estimation, which simultaneously learns the node-level and cluster-level integration processes using a dual graph neural network architecture. We also propose a novel loss objective to jointly learn the clustering assignment across different edge types and the centered CBT representation of the population multigraphs. Our Dual-HINet significantly outperforms state-of-the-art methods for learning CBTs on both small-scale and large-scale multigraph connectomic datasets.
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16431-6_29
SharedIt: https://rdcu.be/cVD5b
Link to the code repository
https://github.com/basiralab/Dual-HINet
Link to the dataset(s)
N/A
Reviews
Review #1
- Please describe the contribution of the paper
In this study, a dual graph convolutional network architecture is proposed to learn connectional brain templates of brain multigraphs, which learns multigraph representations at node-level and module-level simultaneously.
- 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.
Node-level embeddings and hierarchical module-level embeddings are fused to generate multigraph representations.
- 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 proposed method was only evaluated on brain images of single modality (structural), instead of evaluation on different types of brain connnectivity from multi-modality data, such as fMRI and sMRI.
- The generated CBT was only evaluated in terms of representativeness and topological soundness. It is also important to evaluate if the generated CBT can better capture inter-subject differences and the associations between brain connectivity and phenotypes.
- 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
Good.
- 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
- For the hierarchical multigraph clustering, it will be helpful to visualize the assignment matrix to see if the learned clusters are biological meaningful.
- How does the clustering loss L_d affect the performance? It is better to provide an ablation comparison regarding this.
- Both sMRI and fMRI are available in ABIDE dataset, not sure why only sMRI was used. Morphological connectivity is usually computed as region-wise correlation of multiple morphological features, instead of using one single feature seperately. Learn CBT from multi-modal data will be more promising.
- Further evaluation are needed to see if the generated CBT can better capture inter-subject differences and the associations between brain connectivity and phenotypes.
- 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?
The proposed method to learn CBT from brain multigraphs is interesting, but further evaluation is needed to demonstrate its effectiveness.
- 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
Review #2
- Please describe the contribution of the paper
In this paper, the authors propose a novel method, namely Dual-HINet, for dealing with brain multi-graphs while considering the hierarchical structure of neural interactions. The framework consists of several major steps, from which they seem to firstly construct node-level and cluster-level embeddings simultaneously, which are fused together for final analysis. They also use 4 large-scale connectomic datasets (ABIDE I) to validate the performance.
- 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 motivations for constructing brain multigraph in a hierarchical way are sound, and the methodology introduced in this paper is valid, and with sufficient novelty. Experiments also show the validity of the proposed method.
- 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 organization of the paper needs improving, as there are many repetitive parts in the paper.
- 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 main idea and the major steps are provided and described, current contents seem to be OK for reproducing, but it cannot be fully confirmed unless we really dig into it. Publishing source code will be much welcomed which can greatly help in understanding the proposed method.
- 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
Contents in the last paragraph of Section 1, the start of Section 2, and the rest parts are repetitive. It is like the authors have introduced their framework based on Fig.1 by three times, and some contents are overlapped with each other. It is suggested that you provide only a brief introduction of your method that focuses on the main idea in Section 1, while in Section 2 focuses on the detailed descriptions of each major step. Similar contents in the three parts of the paper should be reduced. In Fig.1, it is suggested to exchange the location of subgraphs E and F, so that the order of these subgraphs can be organized in a more natural way. In the second line of the second paragraph in Section 3, what does that “AD” stand for? Should that be “ASD” instead? It is also suggested to provide table to describe the contents in Fig.2 instead of figure, which can provide more accurate information between different configurations and the DGN. It is also unknown why they choose ABIDE I for validate their performance, since it is specifically collected for ASD disease. As this method is more like a general brain network analysis method, using only one dataset might not be sufficient to fully validate it.
- 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?
The novelty and the methodology is good, but the organization of the paper needs improving.
- 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 Dual Hierarchical Integration Network (Dual-HINet) to simultaneously learn the node-level and hierarchical cluster-level integrations for the connectional brain template (CBT). Through the proposed dual GCN block, the proposed method can group the nodes through hierarchical layers based on their multi-edge interactions. The subject-specific CBT is derived from the concatenated node-level and cluster-level embeddings. Finally, the population CBT is generated by taking the median of all the subject-specific CBTs in the training set.
- 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.
- It is very challenging to simultaneously learn the node-level and hierarchical cluster-level integrations to produce the subject-specific CBT.
- 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.
- It is unclear what is the ‘hierarchical structure of neural interactions’. It is recommended to give a specific definition to avoid confusion.
- The ‘hierarchical structure’ is obtained through the block C in Fig. 1. It is recommended to give more clinical/medical intuitive how and why this learned clustering-level hierarchical structure helps building a better CBT.
- It is unclear how GCN works in the proposed method. Typically, GCN requires a fixed adjacent matrix to perform the graph convolution. However, the proposed method inputs multiple graphs and there are multiple GCNs modules in the proposed method. It is recommended to give more technical details about how each GCN works to make the proposed method more convincing.
- There is a ‘predefined number of clusters in the ‘C) Hierarchical multigraph clustering’. It is recommended to give a detailed discussion about how this predefined hyperparameter impacts the model performance.
- The proposed method is very complex, with multiple GCN and clustering modules. However, the training set is relatively small compared with the proposed method’s complexity. It is recommended to give a more detailed discussion about how and why the proposed method can be trained on such a small dataset without overfitting.
- Please rate the clarity and organization of this paper
Poor
- 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
As mentioned in the weaknesses, a lot of technical details are missing making the reproducibility less convincing.
- 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
Please refer to the weaknesses for details.
- 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
3
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The proposed method is very complex with multiple GCN modules and clustering modules. However, the current writing of this paper did not provide enough technical details making this paper less convincing.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
4
- 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
Since the authors promised to make the intuition of the proposed method more clear, I would like to adjust my score from ‘reject’ to ‘weak reject’. However, the author did not answer my technique questions. Thus, the model’s reproducibility and generalization to other dataset are still doubtful.
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.
This paper proposes hierarchical multi-graph integration for a brain connectome template, which presents novelty. However, several points need to be clarified in the paper and answered in the rebuttal. The rationale and assumption for using the hierarchical structure and its clustering for the template construction need to be clearly described. Technical details are not clear about how the multiple graphs are handled together throughout the GCN-based structure. Fig. 1 should be improved accordingly to help clarify the method. While we are not asking for additional experiments, there are several concerns raised by the reviewers, such as a small single-modality dataset used and lack of comparisons, which need authors’ responses.
- 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).
3
Author Feedback
The pioneering aspects of our work were highlighted by the majority of reviewers and AC:
- Novel: “The paper proposes hierarchical multi-graph integration for a brain connectome template, which presents novelty” (AC). “The fusion of node-level embeddings and hierarchical module-level embeddings to generate multigraph representations is a novel method”(R1 & R2).
- Interesting: “to learn CBT from brain multigraphs is interesting” (R1).
- Validity: “The motivations for constructing brain multigraph in a hierarchical way are sound, the methodology introduced in this paper is valid” (R2). ** Clarifying Concerns: 1) Rationale and assumption for using the hierarchical structure and its clustering for the template construction (R3&AC): ==> As mentioned in the introduction, a CBT is a brain multigraph data fingerprint, which can capture the connectivity patterns of a population of brain networks. Such patterns include the hierarchical structure and clustering of brain ROIs (i.e., known as modularity in network neuroscience literature). On a biological level, such properties of brain networks were demonstrated in the neuroscience literature. For instance, brain disorders can affect not only a subset of ROI-to-ROI brain connections but also a subset of brain ROI clusters (i.e., a set of modules) and their low-to-high level interactions. Such neural hierarchical interactions are demonstrated in various studies, which we will add to the paper. This has motivated us to design a hierarchical learning model of the population-driven CBT. The CBT representativeness results of our ablated models support our rationale. We will clarify this point in the Introduction and support it with references.
2) Technical details of how multiple graphs are handled in the GCN-architecture (R3&AC): ==> In our architecture, each GCN module uses edge-conditioned GCN layers, which are capable of integrating node embeddings from multigraphs that share the same number of nodes. Each GCN uses equation 1 then passes its node embeddings to a ReLU layer. Each GCN module learns weights that have different meanings. Parts B & D use embedding GCN modules, formally Z=GCN(A, X), and learn weights to generate node embeddings of their input multigraph. Part C uses a clustering GCN module, S=softmax(GCN(A,X)), where weights are learned for computing cluster assignment probabilities of the input multigraph in a particular hierarchical layer.
** Dataset used for evaluation(R1&R2&AC): 3) We used 2 large-scale datasets ==> We kindly disagree regarding the “small-size” of our dataset. In the neuroscience literature, our dataset is considered as relatively large. To validate our method, we used 4 large-scale multigraph datasets (ASD & NC left and right hemisphere), independently. The order of 300 subjects is considered as large-scale compared to <100 subjects used in SOTA publications (such as DGN). Our model outperformed DGN across all healthy and disordered large-scale brain datasets.
4) Usage of single-modality dataset(R1&R2&AC): ==> We used multi-view data derived from a single modality but that does not negate the generalizability of our method to multigraphs derived from other neuroimaging modalities. Our model works on any type of multigraph population, with heterogeneous views/modalities. We evaluated our model on multi-view data that have different distributions and characteristics. Each morphological connectivity view captures a particular type of relationship between brain regions. Currently, we do not have in our disposal a multimodal brain multigraph data (e.g., derived from DTI and rsfMRI), but this should not diminish the merits of our work. We intend to collect such data, however, we might need to first generalize our model to multi-resolution graphs. For instance, MNI template is usually used to derive 120x120 functional graphs whereas the cortical Desikan-Killiany template, used here, generates 35x35 morphological graphs.
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.
CBT evaluation is still not straightforward after reading the rebuttal, thus less convincing of the effectiveness of the proposed architecture. Graph clustering, as the way for CBT construction, still remains less evident in the current setting of experiments. Moreover, extending architecture for multimodal cases is claimed in the rebuttal as a straightforward extension. However, it will require modifications to the current architecture and introduce more complexity, thus limiting the practical usage of the proposed architecture.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Reject
- 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).
13
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 idea of the paper is interesting, and the rebuttal has addressed the main comments of the reviewers. I think the paper would be an interesting contribution to MICCAI 2022.
- 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).
4
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 strength is the proposed hierarchical multi-graph integration for learning connectional brain template. Most reviewers’ questions were addressed in the rebuttal. Some remaining issues should also be fixed in the revision. I suggest acceptance.
- 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).
6