List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Zhibin He, Wuyang Li, Tuo Zhang, Yixuan Yuan
Abstract
Novel brain gyrus landmarks, the gyral hinges (GHs), have been identified and demonstrated to be of functional importance, and a subset of them have correspondences across subjects and species. A precise GH alignment is crucial for understanding the relationship between brain structure and function. However, accurate and robust GH alignment is challenging due to the massive cortical morphological variations of GHs between subjects. Previous studies typically construct a single-scale graph to model the GHs relations and deploy the graph matching algorithms for GH alignment but suffer from two overlooked deficiencies. First, they consider only pairwise relations between GHs, ignoring that their relations are highly complex. Second, they only consider the point scale for graph-based GH alignment, which inevitably introduces several alignment errors on small-scaled regions. To overcome these deficiencies, we propose a Hierarchical HyperGraph Matching (H2GM) framework for GH alignment, consisting of a Multi-scale Hypergraph Establishment (MsHE) module, a Multi-scale Hypergraph Matching (MsHM) module, and an Inter-Scale Consistency (ISC) constraint. Specifically, the MsHE module constructs multi-scale hypergraphs by utilizing abundant biological evidence and models high-order relations between GHs at different scales. The MsHM module matches hypergraph pairs at each scale to entangle a robust GH alignment with multi-scale high-order cues. And the ISC constraint incorporates inter-scale semantic consistency to encourage the agreement of multi-scale knowledge. Experimental results demonstrate that H2GM improves GH alignment remarkably and outperforms state-of-the-art methods. Our code will make public later.
Link to paper
DOI: https://doi.org/10.1007/978-3-031-43999-5_52
SharedIt: https://rdcu.be/dnww5
Link to the code repository
https://github.com/CUHK-AIM-Group/HHGM
Link to the dataset(s)
N/A
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes a hierarchical hypergraph framework for the gyral hinges (GHs) alignment, and the proposed H2GM also remarkably improves GH alignment and outperforms state-of-the-art methods.
- 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.
(1)This method consider and extract high-order relations among GHs via a MsHE module while previous work mostly considered only pairwise relations between GHs. (2)The proposed H2GM is the first attempt to implement a multi-scale hypergraph framework to successfully align GHs while most previous work only consider the point scale. (3)The proposed method shows better results than state-of-the-art methods. (4)The paper is well-organized and easy to follow.
- 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)It is not guaranteed that 10 is the optimal value for β, considering the large interval of β values showed in Table 2. Is it necessary to try more β values to find an optimal value and verify that the model is robust to β? For parameter selection, I will suggest that the author can take a reference from HyperMorph, or introduce Bayesian parameter estimation to the proposed framework. (2)The performance evaluations are based on too limited data. It would be much better to add experiments on other datasets.
- 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
Good 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
(1)Major point: Please provide more quantitative results and analyses to properly demonstrate its value. (2)Minor point: Insufficient explanation of characters: What’s the meaning of Q in Formulation.7? Minor errors: In Section 2.1, it may be appropriate to change “the raw features are sent Dynamic Graph CNN (DGCNN)” to “the raw features are sent to Dynamic Graph CNN (DGCNN)”. In Section 2.2, it may be appropriate to change “MsGM” to “MsHM”. Wrong tenses: In Section 3.1, “use” and “performed” appear in “We use L=3 layers of the MsHM and performed 20 inkhorn iterations.” at the same time. In Section 3.2, “conduct” and “recorded” appear in “We conduct experiments with varying hyperparameters to investigate the sensitivity of β in Eq. 8 and recorded the results in Table 2. ” at the same time.
- 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 experiments presented in the paper are convicing.
- Reviewer confidence
Very confident
- [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
The authors have addressed my concerns.
Review #4
- Please describe the contribution of the paper
This paper proposed a multi-scale landmark matching algorithm based multi-scale GCN. The method is applied and validated on brain gyral hinge alignment dataset.
- 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 landmark matching algorithm based on GCN.
- The introduction is thorough and provided good background information.
- The method is overall suitable for this application.
- The inter-scale consistency loss function is a good innovation.
- 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 input and output of each stage of the algorithm are not clearly depicted in the figures.
- The benchmarking methods are not reasonably picked. In the method section, the authors cited [5][15] as other graph matching frameworks. However, these papers are not benchmarked in the results section. There also isn’t enough explanation of how the proposed method differs from these papers. Instead, many of the benchmarked methods are designed for point cloud matching and cannot efficiently leverage graph data.
- 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 authors agreed to share code.
- 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 perform more meaningful benchmarks with relevant methods.
-
Improve the figures to better illustrate the input and output of each component in the framework.
-
- 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 experiment design is not persuasive enough to prove the effectiveness of this method.
- 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
4
- [Post rebuttal] Please justify your decision
The additional experiments addressed my concerns. However, the explanation for submitting two very similar works is not convincing. The other submission from this same group uses the exact same benchmarking methods. If the authors noticed that one method outperform another, they should have submitted with the superior method only. Such double-submission practice is wasting peer-review resource.
Review #6
- Please describe the contribution of the paper
The paper presents a framework in three parts to match the brain regions’ landmarks gyral hinges. The first part is the modeling of the brain as a multiscale hypergraph. Second comes a matching part which itself has fours subparts. And lastly, there is a scale consistency constraint.
- 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.
Modeling of the brain in a multiscale fashion and describing a framework how to learn a task on such a model is a valuable contribution.
There are ablation studies and multiple recent baseline methods. The mathematical formulations are detailed.
- 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.
Statistical tests over the evaluations would be valuable. It is questionable whether the performance over the baseline is of statistical significance.
The ablation studies suggest that the ISC and Hyperedge Ralation Learning deliver nearly negligible improvements.
Some sections, the hypergraph establishment one in particular, are somewhat confusing.
The discussion of the ablation studies is questionable. The evaluation results show little support for some components of the method.
- 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’s use a public dataset and state that they will make their code public, thus the work should be 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
One could run a statistical test over the evaluation.
The 2.1 section is somewhat confusing.
- What does it mean to expand the GH on the surface by two rings?
- How does one come to the number of 19 vertices as GHs’ raw features?
Additionally on the clarity, in the beginning of 2. section notation is introduced, however, not everything is commented on with it’s introduction e.g. the s/t subscript and the weight matrix.
On the discussion of the ablation studies. It is stated that the experiments suggest “that the hyperedge structure-aware message is instrumental in improving the alignment accuracy”. However, the improvements seem very minor. Same goes for the ISC. The ablation studies are very valuable to the paper, however, I am not sure about the interpretations these particular numbers. In contrast, the results on the multi-scale modeling are much more convincing.
The last sentence in the abstract has funny English. The linking to the figures does not work properly.
- 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 authors propose to model the brain with a multi-scale hypergraph and outline a framework as how to learn matching of brain landmarks with this model. These are valuable contributions and the mathematics is shown in relative detail.
However, some sections could be described better. The evaluation results are not immediately convincing. This is not acknowledged in the discussion, on the contrary, claims are made with little supporting evidence.
- 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.
- very similar to another MICCAI submission. Please clarify if you have similar paper submitted to miccai and how those two work compare and scientific overlap.
- why use gyral hing matching when existing softwares regiter entire cortex? What type of application would benefit?
- comparison with other approaches is lacking as mentioned by reviews.
Author Feedback
We thank the reviewers for approving our method. The common questions are answered first, followed by responses to individual review comments.
Q1(AC): Clarify the differences with other similar work. There is another submission finished by our group members, which was rejected. That work focuses on gyral hinge (GH) matching (GHM) by extracting GHs’ local and global features at a single scale only. Differently, our work utilizes multi-scale hypergraphs to establish high-order relationships for GHM. This approach enables us to capture complex dependencies and interactions, improving matching accuracy.
Q2(AC): Explain the meaning and application of GHM. The ultimate goal of entire cortex registration (ECR) is to achieve functional region alignment across subjects. However, existing software often sacrifices the local functional alignment accuracy to accomplish ECR. The precise matching of GHs, which have been demonstrated to have inter-subject functional correspondences, will improve the local functional alignment accuracy of ECR and advance our understanding of the brain’s functional organization.
Q3(AC&R1&R2&R3): Comparison with more SOTA with different datasets and more evaluation. SOTA: We add [5, 15] and [35: PCA-GM] to the comparison, where [35] is designed for graph data. Dataset: We use the Chinese HCP (CHCP) dataset for our experiments, including 100/50 subjects as a training/test set. Evaluation: We use Mean Geodesic Errors (MGE), widely used in point cloud matching tasks. The results are as follows {[7, 8, 28, 5, 15, 35], our}: In HCP, Accuracy(ACC, %)={74.71, 71.83, 75.30, 77.10, 77.01, 74.32, 78.03}, MGE(×100)={9.7, 11.2, 8.8, 6.7, 6.7, 10.1, 5.8}. In CHCP, ACC(%)={71.21, 68.01,71.72, 73.37, 73.42, 69.03, 74.77}, MGE(×100)={11.6, 14.2, 11.0, 8.3, 8.5, 12.1, 7.2}. Our method shows superior performance on both datasets, on all evaluation metrics, as well as on methods using hypergraphs [5, 15] and designed for graph data [35], verifying the effectiveness and generalization of our method. Notably, GH data can be treated as point cloud data, allowing us to leverage benchmarked methods designed for point cloud matching.
Q4(R1&R3): More ablation studies. We have conducted extensive experiments to explore the selection of β and list the results. β={5, 7, 9, 10, 11, 13, 15, 20}, ACC(%)={77.84, 77.83, 77.97, 78.03, 78.01, 77.94, 77.91, 77.87}. The highest ACC is achieved when β=10. Hence, we set β as 10 as the paper claimed. To showcase the efficacy of our method’s Inter-Scale Consistency (ISC) and Hyperedge Relation Learning (HRL) modules, we conduct t-tests on the ACC of the ablation studies. Our method is significant with the experiments of {w/o ISC, w/o HRL}-{0.003, 0.001} in terms of ACC, indicating that both ISC and HRL modules can effectively improve matching accuracy.
Q5(R2): Explanation of the difference with [5, 15]. Unlike [5,15] with a single scale, we extract high-order relations among GHs and propagate GH features for alignment across three scales. Moreover, we introduce inter-scale semantic consistency to optimize the feature distribution of GHs and leverage multi-scale knowledge in hypergraph matching, enabling more effective message propagation and enhancing alignment accuracy.
Q6(R3): Statistical test results. By t-tests, our method is significantly (* indicates p-value<0.001) better than the methods of {[7, 8, 28, 5, 15, 35]}-{*, *, *, *, *, *} in terms of matching ACC.
Q7(R1): The meaning of Q in Form.7. The Q is the number of children nodes belonging to the same parent node.
Q8(R3): The meaning of 2-rings and why use 19 vertices as GH’s raw features. GH is based on a single point extracted from the mesh surface. “2-rings” pertains to the collection of neighboring vertices within a two-step distance from the central point on the mesh. The highest ACC is achieved with 19 points (78.03%) in 2-rings as raw features compared to 7 points (73.31%) in 1-ring and 38 points (75.42%) in 3-rings.
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 authors have adaquately addressed the issue of comparisons with other methods.
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 proposes a multi-scale matching framework for graphs of gyral hinges in brains. The framework is novel, and the overall paper is well-written. Comparison to some baseline methods and ablation studies are shown. The rebuttal mostly addressed the reviewer’s concern. The only remaining issue is that the primary meta-reviewer identified a similar submission from the same group, which was an early reject. This submission proposes another method for the same problem and reports the same baselines (according to R#4, this meta-reviewer doesn’t have access to the other submission). Although double-submissions are bad practice, I would recommend acceptance in light of the early reject of the other work and, more importantly, the merits and contributions of this work.
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.
While the authors have addressed the majority of the reviewers’ questions, this meta reviewer has identified two potential violations of the submission and rebuttal guidelines for MICCAI. Firstly, it was explicitly stated in the rebuttal rules that the authors should refrain from providing or promising new or additional experimental results, as the final decision is based on the submitted manuscript. Although the authors’ new experimental results addressed the reviewers’ questions, they violated the rebuttal rules in doing so.
Secondly, as per the MICCAI submission rules, authors are required to confirm that they have not submitted substantially similar work to MICCAI 2023. Violation of this condition may result in rejection. As noted by R2, if the submissions are indeed substantially similar, the authors should have considered submitting only one version of the work.
These violations raise concerns about adhering to the guidelines and policies set by MICCAI. It is important for authors to follow these rules to ensure fair and unbiased evaluation of their work and to avoid duplication of reviewer efforts.