List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Carlo Amodeo, Igor Fortel, Olusola Ajilore, Liang Zhan, Alex Leow, Theja Tulabandhula
Abstract
Graph theoretical analyses have become standard tools in modeling functional and anatomical connectivity in the brain. With the advent of connectomics, the primary graphs or networks of interest are structural connectome (derived from DTI tractography) and functional connectome (derived from resting-state fMRI). However, most published connectome studies have focused on either structural or functional connectome, yet complementary information between them, when available in the same dataset, can be jointly leveraged to improve our understanding of the brain. To this end, we propose a function-constrained structural graph variational autoencoder (FCS-GVAE) capable of incorporating information from both functional and structural connectome in an unsupervised fashion. This leads to a joint low-dimensional embedding that establishes a unified spatial coordinate system for comparing across different subjects. We evaluate our approach using the publicly available OASIS-3 Alzheimer’s disease (AD) dataset and show that a variational formulation is necessary to optimally encode functional brain dynamics. Further, the proposed joint embedding approach can more accurately distinguish different patient sub-populations than approaches that do not use complementary connectome information.
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16431-6_39
SharedIt: https://rdcu.be/cVD6V
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 proposes a novel method, the Functionally Constrained Structural Graph Variational Autoencoder (FCS-GVAE), capable of combining information from the functional and structural connectomes in unsupervised learning. The method is evaluated on OASIS-3 Alzheimer’s disease (AD) dataset. The results show that this method has a better performance than the baseline model.
- 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 is very interesting, which combines the functional and structural connectomes of individuals to inferences about their differences. -The methodology is explained beautifully, albeit it could be improved.
- 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.
-Presentation and Writing Quality: The quality of the paper needs to be improved: figures/ diagrams, and the overall presentation need significant modification before publication. -The Experiments (datasets\ ablation experiment) of the proposed methods can be further validated. -The discussion (weakness and limitation) of the proposed methods can be further revealed.
- 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
-Reproducibility: Code 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: https://conferences.miccai.org/2022/en/REVIEWER-GUIDELINES.html
This paper proposed a new idea in brain networks. However, in my opinion, there lacks of insightful thinking about the results and some minor comments need to be addressed. -Figure 3 is not explained further in the article, what is the role of Figure 3? -The classification methods ( SVM, MLP, and RFC)of AD need a detailed description. And why not compare with CNN for the AD classification tasks? In Table 1, what mean is the RFC?
- Why not do an ablation experiment? How much does this node feature(fMRI) affect the downstream tasks?
- How do the proposed methods and tuning strategies generalize to other datasets?
- 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?
-More experiments are needed to prove that the combination of functional and structural connectomes can make the downstream task better. The Presentation and Writing Quality are poor. Not good enough. Limited experiments. Needs more work and/or revisions. I believe it should be rejected.
- Number of papers in your stack
2
- 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
5
- [Post rebuttal] Please justify your decision
In the previous review, I have raised several concerns about the clarity of the paper, the presentation of Fig3 and Table1 as well as some experimental details. The authors’ response has addressed my concerns. Although this rebuttal remains concerned with the generalization of this model on other datasets, I think this work should be of interest to a sub-community of the MICCAI. I would suggest a weak acceptance of this paper.
Review #2
- Please describe the contribution of the paper
The author proposed a function-constrained structural graph variational autoencoder model to learn the joint embedding of both structural and functional information. To further evaluate the proposed model, they used the joint embeddings to classify different populations.
- 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 author proposed a novel application of graph variational autoencoder model in incoporating the functional and structural connections and identifying joint embeddings.
- 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 author claims that their model can establish a unified spatial coordinate system for comparing across different subjects, their analyses still rely much on the registration process in preprocessing stage. In addtion, the necessity of including an AE model is questionable and needs further clarification.
- 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 reproducibility of the paper is limited.
- 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 would be necessary to discuss the necessity of including an AE model in the proposed framework.
- Did author use separate training and testing dataset for classification task? Please add more details.
- Lack the visualization of joint embeddings of structural and functional connectome, which makes it difficult to understand the specific correspondence between them.
- 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 novel application of graph VAE model.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
2
- 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 work proposes a function-constrained structural graph variational autoencoder (FCS-GVAE) capable of incorporating information from both functional and structural connectomes in an unsupervised fashion.
- 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.
Contributions: 1). It is very interesting to employ an approach of incorporating function-constrained structural graphs from both functional and structural connectome in an unsupervised fashion to generate a variational graph autoencoder. 2). In this work, the authors provide a comprehensive experimental study and hyperparameter tuning.
- 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.
Major Concerns: 1). Methodological validation
The reviewers are curious about validating proposed methods with other peer deep neural networks.
2). More details of the sampling technique in this paper need to be included.
3). More explanation that orthogonality cannot allow for better compression is needed.
- 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 reproducibility of this paper is limited.
There is no source code released in this work.
- 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
Major Concerns: 1). Methodological validation The authors provide the validation based on two embedding techniques for AD classification in this work. In addition, the reviewers are interested in comparisons of other peer deep models. For instance, the authors can evaluate Deep Boltzmann Machine (DBM) to replace the AE in Fig. 1. Here are some references for methodological validation: Salakhutdinov, R., & Hinton, G. (2009, April). Deep Boltzmann machines. In Artificial intelligence and statistics (pp. 448-455). PMLR. Furthermore, did the authors perform cross-validation for classification validation? Unfortunately, there have not been more details of classification validation included in this work. Moreover, reviewers are curious about the comprehensive comparisons. In detail, can authors validate the proposed GVAE with other peer methods in terms of time-consuming and reconstruction accuracy? This validation would further benefit the clinical translational application in the future.
2). Sampling Technique Issues In Fig.1, the authors described a vital computational pipeline. However, the authors do not discuss the sampling technique in detail. Which sampling technique do the authors utilize in this work? Can authors validate their sampling techniques with Gibb’s sampling? Or some dimensionality reduction techniques can be thought of as an alternative way to replace the sampling techniques?
3). Why is orthogonality better? In Section 2.5, the authors emphasized, ‘Note that, unlike PCA, a single layer AE with no-nonlinearity does not impose orthogonality, allowing for better compression. Originally, orthogonality could maintain the lowest dimensionality of feature space. From the reviewer’s perspective, there is probably extensive overlapping existing in feature spaces generated by AE. Alternatively, can the authors provide references to support their conclusion?
- 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 approach comprehensively employs the functional and structural information is promising.
Further validating the proposed GVAE with the other peer deep neural networks is required; more details and an explanation of the technique should be provided. Moreover, clarify the terminology such as no-nolinearity.
- Number of papers in your stack
5
- What is the ranking of this paper in your review stack?
3
- Reviewer confidence
Very confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
5
- [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 proposed idea of embedding structural and functional connectome data using a generative graph model has novelty and technically sounds. However, the paper needs to be improved, and the authors are expected to answer critical questions in the rebuttal. The importance of a single AE is not much evident, although the authors explained it in terms of no-nonlinearity. Discussion regarding Table 1 and Fig. 3 needs further clarification. Reviewers raised detailed and constructive comments regarding figures, experimental settings, the sampling technique, and preprocessing. Authors should answer those questions in the rebuttal as well.
- 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).
2
Author Feedback
We thank all reviewers for recognizing the novelty of our approach, and the area chair for giving us a chance to detail how we plan to rectify concerns. Certain comments regarding writing clarity, more robust ablation studies, and insights from results are well taken and will be included in the manuscript revision. We focus here on aspects which caused confusion or needed explanation:
Clarifying results and downstream classification task: Reviewers had questions regarding the importance of Figure 3 and Table 1, as well as the main aspects of the downstream classification task. We apologize for the lack of clarity; they show the main difference between the latent spaces obtained when embedding both function and structure of the brain together and independently. Figure 3 qualitatively shows how including the variability of the FC leads to a regularization of the latent space. If it is not included, the latent distribution appears random. However, when FC is included, it highlights different populations. Table 1 shows the effect of regularization. Different classification models overcome the unbalancing of the datasets and improve the outcome of the classification task itself, in terms of F1-score, precision, and recall. The three models employed are: support vector machine (SVM), multi-layer perceptron (MLP), and random forest classifier (RFC). They are used as a measure for the quality of representations obtained by our method. The hyperparameters used are standard from the sklearn package in Python. Classification has been done by randomly splitting the data into training/test sets (80%/20%), with reported performance over the test set. A cross-validation loop was used with the training data to find the best models. A NN (even CNN) may have been used here, and will be included in the revision.
Expanding on the experimental setting (sampling and further experiments), design choices, and data processing: We thank the reviewers for the suggestion to expand on these elements to strengthen the manuscript. The sampling employed in this work is similar to a traditional VAE setup, with straightforward Gaussian sampling (Sampling is taken care of during training using the reparameterization trick). In choosing this, minimal assumptions are made (other than standard Gaussianity of the latent variables). This choice simplifies the pipeline in line with extant work, and other distributional assumptions (and their sampling strategies) may be explored in the future. The key benefit of using a variational architecture is that it allows for generative modeling (e.g., generating synthetic structural connectomes for further analysis) The second component of our model, the autoencoder, is flexible in the proposed workflow. As rightly pointed out it may be replaced, for instance by a PCA, or even eliminated. However, PCA introduces the constraint of orthogonality (implicit in its definition), which is not introduced by an equivalent single layer AE and would lead to a higher reconstruction error for the former. We apologize for the confusion language in the manuscript; if the AE was eliminated, one of the key aspects of this work, namely the proper dimensionality reduction of the brain networks down to an interpretable common coordinate space, wouldn’t be easy to accomplish.
We perform minimal processing on the structural graph and generate functional graphs via Pearson coefficients of BOLD signals in a particular time window (without thresholding). Thus, the choice of time windows is the only preprocessing step that may influence downstream results. Embedding dimensions, number of layers, and others for the FCSGVAE have been justified as much as possible in the submission (and will be improved upon in the revision). The link prediction and classification tasks support the tenet that the representations learned from the joint embeddings are non-trivial by demonstrating that they are useful for both reconstruction as well as prediction tasks.
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 of the proposed method for embedding structural and functional connectivity using GVAE was recognized by reviewers. The rebuttal satisfactorily addressed reviewers’ critiques, including clarification of the proposed architecture and detailing downstream tasks, experimental setting, and preprocessing.
- 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).
7
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 paper presents a novel application of graph VAEs that can joinly learn structural and functional information. The rebuttal has addressed major concers raised by the reviewers. This paper can raise insightful discussions at MICCAI.
- 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).
7
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 propose a functionally constrained VAE for common embedding of functional and structural connectivity. The authors provide good responses to the reviewer comments.
- 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