List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Xi Chen, Wenwen Zeng, Guoqing Wu, Yu Lei, Wei Ni, Yuanyuan Wang, Yuxiang Gu, Jinhua Yu
Abstract
As one of the common complications, vascular cognitive impairment (VCI) comprises a range of cognitive disorders related to cerebral vessel diseases like moyamoya disease (MMD), and it is reversible by surgical revascularization in its early stage. However, diagnosis of VCI is time-consuming and less accurate if it solely relies on neuropsychological examination. Even if some existing research connected VCI with medical image, most of them were solely statistical methods with single modality. Therefore, we propose a graph-based framework to integrate both dual-modal imaging information (rs-fMRI and DTI) and non-imaging information to identify VCI in adult MMDs. Unlike some previous studies based on node-level classification, the proposed graph-level model can fully utilize imaging information and improve interpretability of results. Specifically, we firstly design two different graphs for each subject based on characteristics of different modalities and feed them to a dual-modal graph convolution network to extract complementary imaging features and select important brain biomarkers for each subject. Node-based normalization and constraint item are further devised to weakening influence of over-smoothing and natural difference caused by non-imaging information. Experiments on a real dataset not only achieve accuracy of 80.0%, but also highlight some salient brain regions related to VCI in adult MMDs, demonstrating the effectiveness and clinical interpretability of our proposed method.
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16443-9_64
SharedIt: https://rdcu.be/cVRzi
Link to the code repository
N/A
Link to the dataset(s)
N/A
Reviews
Review #1
- Please describe the contribution of the paper
In this paper, the authors intend to propose a novel graph-based method of diagnosing VCI using two modalities (fMRI and DTI). The main contribution in methodology is two-fold: one is the dual-modal GCN to firstly process the two-modality images independently (graph construction and convolution) and fuses them together for the diagnosis, the other is the node-based normalization and constrain mechanism to resolve the over-smoothing issue in GCN and incorporation of the non-imaging information. Experiments such as ablation studies, SOTA comparisons and biomarker interpretations also demonstrate the effectiveness of their works.
- 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 methodology of dual-modal GCN is reasonable. The motivation for designing the node-based normalization and constraint is sound, although the design of such mechanisms is not quite clear to me
- 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 is quite a general graph-based tool, which is not specifically designed for VCI diagnosis, and to resolve the potential issues in this research topic. The implementation of the node-based constraint item is confusing and should be further explained. Experiments were not conducted in cross-validation, so the validity of the results is questionable, and training number is also limited to construct the proposed model in a proper way.
- 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
It seems that the authors have provided the necessary information to reproduce the methods, but it is still strongly suggested to publish the source code online if available, which can help the readers to fully understand its mechanism.
- 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 major issue in the methodology section, is its descriptions of the node-based constraint item, especially the similarity constraint mechanism. The motivation to design the fusion of imaging and non-imaging features in such manner (using imaging feature as node and the similarity between two non-imaging features as edge) is not clear to me, as you can just simply concatenate them together for constructing the diagnosis model, which can also be effective and can fully utilize the two features. The computation of loss2 is also missing, which makes it difficult for readers to reproduce this work. Table 1 shows the ablation study, where the last line seems to be the results of the proposed method, but it is unclear what the meaning is for that (1) and (2), which is also not clearly stated in the methodology and experimental sections. Therefore it remains confusing to me why the last line can achieve such a high performance compared with the others. Also, in some lines the performances from the testing set seem to be better than those from the validation set, which is also against my expectation, and it is suggested that the authors explain more about this. The authors claim that they use 156 subjects for training, which might be sufficient for some conventional machine learning methods such as SVM, but definitely not for GCN. Limited training data will result in overfitting of the model training, which can hinder the credibility of the proposed method. To solve this problem, the usual way is data augmentation, from which there are already some ways available in this field. I am wondering if the authors realized such issues and if they have any solutions to them. In Section 3.3, the authors provide the biomarker interpretations, from which they found some ROIs which are highlighted for VCI diagnosis, but it remains unknown if such findings are valid when there are no related references included in this paper to support 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
4
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
There are some unclear parts in the methodology section, and the explanations to the experimental results are also missing. The rest of the paper seem to be OK to me, there are some contribution and novelties in this paper but are not quite significant to me.
- Number of papers in your stack
4
- 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
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #2
- Please describe the contribution of the paper
This paper proposed a dual-modal GCN framwork to integrate imaging and nonimaging information for VCI identification in adult MMDs.
- 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 design different ways to extract complementary information from rs-fMRI and DTI when constructing graphs. The personal biomarker interpretation is interesting.
- 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 process of the graph construction is not clear in this paper. The author should use the cross-validation method and repeat the experiment some times to finally obtain the generalization ability of the proposed method.
- 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 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
1）At the beginning of the abstract part, it is suggested that the author briefly describe in one sentence what the existing methods for VCI research are usually and what are the shortcomings, and then lead to the research of this paper. 2）In the second paragraph of the introduction, the research on VCI using rs-fMRI and DTI and their shortcomings are introduced respectively. What is the traditional research on multi-modality data fusion, how to fuse multi-modality data, and what are the similarities and differences between the research in this paper and the previous multi-modal fusion research? The author can briefly explain. 3) In the third paragraph of the introduction, what is the traditional multi-modality method of the graph, and what are the advantages of the model proposed in this paper. The author can briefly explain. 4) In Section 2.2, graph construction based on rs-fMRI and DTI, the authors should indicate whether a fully connected graph or a sparse graph is used. In addition, whether fully connected graph or sparse graph should be reflected in Fig. 1 and Fig. 2. Besides, it should be stated whether a uniform graph structure was used for all subjects, or different graph structures were used between subjects. The authors can briefly explain how the FN matrix is calculated. 5) In section 2.3, how the input and output dimensions of each layer change, the author should give a brief description. 6) In the experimental parts of 3.1 and 3.2, because the sample size is small, the experimental results obtained by only dividing the training set, the validation set, and the test set cannot explain the performance of the model. And please clarify the usage of validation set in this work. More importantly, the author should use the cross-validation method and repeat the experiment some times to finally obtain the generalization ability of the proposed method. 7) In Section 3.3, please supplement the explanation of how the self-attention pooling mechanism preserves the important brain regions of the subjects.
- 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 theoretical novelty of the paper is limited. The experiment strategy is not rigorous.
- 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
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #3
- Please describe the contribution of the paper
- This paper designs different ways to extract complementary information from rs-fMRI and DTI when constructing graphs, which maximizes the utilization of characteristics of different modalities.
- Node-based normalization and similarity constraint item are proposed to improve performance by solving the problem of over-smoothing and integrating non-imaging information, respectively. 3) Some salient biomarkers for VCI identification are selected by introducing self-attention pooling mechanism which combines node features and graph topology, showing the clinical value of the proposed 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.
This paper tackles the recognition of VCI in MMD by an integrated GCN. By constructing graphs that extract complementary information from two modalities that contain the corresponding important information, node-based normalization and similarity constraint item are combined to improve performance. The results from the private dataset demonstrate this method can highlight some salient brain regions related to VCI in adult MMD while achieving high accuracy.
- 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 dataset is not sufficiently described.
- 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 dataset is not sufficiently described.
- 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 description of the method section is not clear.
- 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
7
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper is novel in method, clearly organized and full of results.
- 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
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.
Details of constraints could be added and cross-validation is recommended.
- 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
Thanks for all the comments to help us improve our manuscript and here is our response after integrating suggestions and confusion from all reviewers. 1.Concern about the limited dataset size and validity of the results. On the one hand, some common augmentation tricks such as flip and cropping are not suitable for rs-fMRI and DTI since they are 4D. Some previous works covering fMRI and GCN were also finished with limited dataset and the total numbers of subjects were 170, 109 and 184, respectively (reference 1~3). On the other hand, we will perform cross-validation in the revised version, hoping to increase the credibility of the method. Ref.:1)Graph convolution network with similarity awareness and adaptive calibration for disease-induced deterioration prediction (MIA,2021).2)H3K27M Mutations Prediction for Brainstem Gliomas Based on Diffusion Radiomics (MICCAI,2021).3)Graph Convolutional Network Analysis for Mild Cognitive Impairment Prediction (IEEE ISBI, 2019).
2.More details about the constraint item in method. The constraint item is designed to reduce the disturbance caused by non-imaging information. In general, some non-imaging information like manufacturer may lead to natural difference of medical image and obstruct correct classification. Therefore, unlike our graph-level backbone that regards each subject as a graph, we propose an extra node-level item to constrain the classification of main backbone. In detail, for node-based item, we take imaging features extracted from graph-level backbone of each subject as a node and the similarity between two non-imaging features as edge, which can make subjects with similar non-imaging feature closer and narrow the scope of final classification to a certain extent. For example, without node-based constraint item, the model needs to classify the sample equally in the whole sample group, but with constraint, the model tends to achieve classification in the sub-groups with similar non-imaging information, thus we can repress the non-imaging disturbance. We believe such method is better than simple concatenation since low-dimensional non-imaging features could be ignored if concatenated with high-dimensional imaging features. It should be noticed that the final classification result is the prediction result of graph-level backbone and the node-based item is only designed to constrain the prediction through the cross entropy loss function loss2.
3.Confusion about Table 1. The last line (2) in the Table 1 is our proposed method, and the difference between (1) and (2) is the graph construction. We construct the graph in (1) using method mentioned in reference 12 which has been clarified in manuscript. The final performance indicates that graph construction is a fundamental but important part in a GCN model.
4.Existing references about biomarker interpretation. We have listed some references to support our findings. On the one hand, reference 1 and 2 suggest that olfactory cortex and hippocampus-related regions are related to mild cognitive impairment. On the other hand, reference 3 shows that executive dysfunction is one of the important characteristics for VCI in adult MMD. Ref.:1)Profound loss of layer II entorhinal cortex neurons occurs in very mild Alzheimer’s Disease (J. Neurosci.,1997).2)Hippocampal and entorhinal cortex volume changes in Alzheimer’s disease patients and mild cognitive impairment subjects (IEEE EMBS,2018).3)Effect of moyamoya disease on neuropsychological functioning in adults (Neurosurg,2008).
5.The description of the method section is not clear. We will revise method part including: 1)Details about the dataset like imaging parameters. 2)Explanation about how the self-attention pooling works. 3)Clarifying the sparse graph construction. 4)Details about the dimensions of each layer. Other concerns will also be clarified including brief description of existing methods and their shortcomings for VCI research, traditional methods for multi-modal data fusion.