List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Tingting Dan, Minjeong Kim, Won Hwa Kim, Guorong Wu
Abstract
A confluence of neuroscience and clinical evidence suggests that the disruption of structural connectivity (SC) and functional connectivity (FC) in the brain is an early sign of neurodegenerative diseases years before any clinical signs of the disease progression. Since the changes in SC-FC coupling may provide a potential putative biomarker that detects subtle brain network dysfunction more sensitively than does a single modality, tremendous efforts have been made to understand the relationship between SC and FC from the perspective of connectivity, sub-networks, and network topology. However, the methodology design of current analytic methods lacks the in-depth neuroscience underpinning of to what extent the altered SC-FC coupling mechanisms underline the cognitive decline. To address this challenge, we put the spotlight on a neural oscillation model that characterizes the system behavior of a set of (functional) neural oscillators coupled via (structural) nerve fibers throughout the brain. On top of this, we present a physics-guided graph neural network to understand the synchronization mechanism of system dynamics that is capable of predicting self-organized functional fluctuations. By doing so, we generate a novel SC-FC coupling biomarker that allows us to recognize the early sign of neurodegeneration through the lens of an altered SC-FC relationship. We have evaluated the statistical power and clinical value of new SC-FC biomarker in the early diagnosis of Alzheimer’s disease using the ADNI dataset. Compared to conventional SC-FC coupling methods, our physics-guided deep model not only yields higher prediction accuracy but also reveals the mechanistic role of SC-FC coupling alterations in disease progression.
Link to paper
DOI: https://doi.org/10.1007/978-3-031-43898-1_9
SharedIt: https://rdcu.be/dnwAG
Link to the code repository
N/A
Link to the dataset(s)
N/A
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes to use a generalized version of a basic coupled dynamic oscillators model to generate (nodal mean) BOLD signals, where the coupling comes from the structural brain connectome. The model is fit / optimized with respect to the actual observed (mean) BOLD signals. The coupled oscillators model utilizes time varying intrinsic state variables per node. Global sychronization level of the brain is quantified as a function of time using nodal state variables and exploited to define new biomarkers. The validity of the dynamic model is shown by comparison to conventional functional connectomes, while the diagnostic power of the proposed biomarkers (AD-vs-CN) are compared with 3 standard non-AD-specific DL models.
- 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 main contributions of the paper are offering a generalization of Kuramoto model for coupled oscillators where the natural frequency term in the original model is replaced with a parametric FCN and the coupling term is replaced with a GNN. The model thus built is used to describe the dynamics of an intrinsic nodal state variable, which is used to generate a “simulated” BOLD signal. The model output is compared with the observed BOLD signal to drive the optimization/learning task. The authors have adopted the main idea in the original model that explains the dynamics of phase of an oscillator and exploited it to describe an intrinsic state variable for each oscillator, where each brain region (connectome node) is modeled as an oscillator. This parametric, learnable generalization and qualitative assessment of the model outputs (with respect to the observed functional connectome) is the main novel contribution.
- 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.
- My main concern is about the applicability of such physical oscillator models to regional mean fMRI BOLD signals. BOLD signals are very indirect measurements of brain function and have very poor temporal resolution, let alone the fact that due to technological limitations there is a good deal of interpolation in time to build one snapshot of the whole brain. Such physical models are more likely to be applicable to signals such as EEG. Hence, I am skeptical about the conclusion with regard to AD-vs-CN.
- The diagnostic classification performance hits around 75-80% accuracy which is not better than AD-vs-CN classification performances reported in literature. The authors have compared their method with generic models (GCN, RNN and LTCNet) whereas the comparison should have better been done with respect to methods developed for this specific task or at least for the domain of dementia.
- The paper refers to prior work on structure-function coupling but misses an important fundamental approach which is based on graph laplacians and graph fourier transform. One such main publication that should have been cited is “Human brain networks function in connectome-specific harmonic waves, Nature Communications 2016 7:1, (2016), 1-10, 7(1)”.
- Although the methods section is generally comprehensible and does convene the main components to the reader, the lack of some details and the clarity of writing prohibits reproducibility. For example, (a) Section 2.2: The learnable parameter in the diffusion process is not clear, (b) Section 4.3: The definition of the order parameter Phi_t has a term as “sum_i e^{i v(t)}”, I think it must be “sum_i e^{j v_i(t)}”.(c) Section 2.2: Inverse f function is not specified.
- It is not clear to me how an AD-vs-CN classifier can be used as an early diagnostic tool. ADs are already in the late stages. An early diagnosis method should be tested on identifying CNs that turn to ADs, which requires a follow-up study. Otherwise, it’s better not to make such claims.
- 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 manuscript itself is not sufficient for reproducibility but as the authors have declared that they will share the codes, it must be re-runable. Nevertheless, thought the data used is open data, the ADNI dataset is quite complicated and specifying that 250 subjects from ADNI are used is not sufficient to identify these subjects. Hence, reproducing the results may not be straight forward.
- 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
- Be more precise in describing the math behind your model, define all variables clearly and use consistent notation free from typos.
- Provide implementation and running details, such as hyperparameters used.
- Perform comparative assessment with methods developed for the same problem or at least for the same domain.
- Early diagnosis is important. However, training and testing AD-vs-CN is not apt for that purpose. You should either be using a follow-up study or at least use subjects that span the spectrum of dementia and show that your method can monitor the disease progress.
- 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?
- I do not think such dynamic models are applicable to BOLD signals.
- The results do not show an improvement over SoA AD-vs-CN classification.
- There are unclear parts of the methodology, despite the manuscript does convey the main idea.
- 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 physics-guided graph neural network identifies disease induced changes in Structural-Functional connectivity coupling based on neural oscillation. Applied to the ADNI data set, this model is more accurate than conventional coupling methods and reveals the mechanistic role of SC-FC coupling alterations in disease progression
- 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.
- provides neuroscience insight at a system level based on a physical model
- basing deep model on the principle of the Kuramoto model, which allows us to characterize the SC-FC relationships with mathematical guarantees.
- 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.
- terminology (such as oscillation) is mostly used in EEG literature even though EEG is not used in this article, i.e, article needs to be made more accessible to DTI/fMRI readership
- selection criteria for ADNI samples unclear
- definition of test set unclear
- baseline comparison is limited to pearson correlation
- representing a continuous connection between accuracy scores is wrong (fig 4d) and it is unclear why representation classification accuracy is represented as boxplot in Fig 4e
- 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
uses ADNI data but unclear which samples - also unclear if code will be made publicly available
- 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
- comparison to dynamic functional modelling would be helpful
- synthetic example with known ground truth would have strengthened experiments
- account for confounders (such as age and sex differences)
- expand comparison beyond person correlation
- describe Figure 4d and e in the text or table instead of figure
- 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?
I really like the approach and I think it is novel / will spur discussion. The experiments need a further revision before being fully convincing
- 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 authors propose the physics-guided graph neural network (GNN) to uncover the synchronization between structural connectivity (SC) and functional connectivity (FC) in the brain. The Kuramoto model is adapted to be incorporated into the GNN and the resulting SC-FC coupling descriptors are used to predict Alzheimer’s disease. The experimental results on ADNI dataset demonstrated the potential of this new biomarker in terms of diagnosis and prognosis.
- 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 main strength of this paper is the novel physics-informed GNN based on the Kuramoto model in neuroscience. The resulting SC-FC coupling descriptors demonstrated the improved AD classification accuracy showing the potential in neuroimaging application.
- 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 validation can be improved. The SC-FC-META features are fed into SVM but can be applied to more advanced classifiers such as random forests or extreme gradient boosting trees (XGBoost). In addition, the details about GNN parameters are not provided. Also, they need to discuss the impact of hyperparameters (e.g., chaos level) on the ablation study.
- 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
Details about the implementation of GNN and Kuramoto models can be added to improve reproducibility. Providing code should be recommended.
- 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
The authors propose the new graph neural network (GNN) constrained by the Kuramoto model to find the new biomarker which encodes the coupling between structural and functional connectivity in the brain. The paper is well-written and easy to read. The experimental results on ADNI dataset are convincing and showed improved performance on AD diagnosis. Even though I would accept the paper, I suggest the authors provide more details on the implementation (e.g., hyperparameters) of GNN and Kuramoto models for reproducibility. In addition, they can test the SC-FC-META features with the advanced classifiers to justify the additional training on SC-FC-Net.
- 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?
I recommend ‘Accept’ on this paper mainly due to the novelty of Physics-informed GNN. In applying deep learning to medical images, the prior knowledge or physical model can be beneficial to improve the robustness with limited data. In particular, the SC-FC coupling shows great potential for diagnosis of neuro-degenerative disease and the resulting biomarkers seem promising.
- 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
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 authors proposed the physics-guided graph neural network (GNN) to uncover the synchronization between structural connectivity (SC) and functional connectivity (FC) in the brain. The Kuramoto model was used to be incorporated into the GNN and the resulting SC-FC coupling descriptors were used to dignoize Alzheimer’s disease. The work had lots of merits and may spur discussion. However, R#1 raised lots of questions regarding the writing and applicability of the work. During the rebuttal period, the authors should address and improve the manuscript quality.
Author Feedback
We are glad that reviewers consider “the work had lots of metris and may spur disucssion”. We sincerely appreciate all the constructive comments. We are committed to incorporating all the valuable feedback into the final version of our paper. Since the major concerns are raised by R1, we will mainly focus on clarifying critiques regarding the writing and applicability of the work.
R1 Regarding the applicability of our method, we would like to clarify as follows. First, the overarching goal is to embark on a novel exploration of addressing open questions in neuroscience using data-driven approaches. We focus on uncovering the biological mechanism behind self-organized functional fluctuations in fMRI data. Inspired by the hypothesis of coupled neural oscillations, we utilize the Kuramoto model to understand the dynamics of functional fluctuations. Our approach optimizes spectral characteristics of local brain nodes, providing insights into these dynamics. Unlike existing studies, we present an end-to-end solution that directly learns a task-specific model from observed neuroimaging data. Our approach aligns with neuroscience principles and is scalable for analyzing data from animal and human studies. Second, this work primarily focuses on the engineering perspective of SC-FC coupling mechanism, i.e., generating putative biomarkers for disease diagnosis. We completely agree with R1 that a physiological signal such as EEG is a good fit for demonstrating the concept of SC-FC coupling. However, there is a large body of SC-FC coupling studies (PubMed 28596608, 19497858, 36207502) using fMRI, due to the higher spatial resolution. The AD vs. CN classification accuracy by our physics-constrained deep model (75.9%) outperforms other methods based on PDE (73.6%), GNN (72.8%), and RNN (69.6%) backbone, using all fMRI data in ADNI. In this regard, our work has the potential to apply to the clinical field. It is important to note that in our training process, we employ disease classification as the loss function, to drive the optimization of the model. However, our primary interest lies in gaining insights into the functional dynamics, as demonstrated in Fig. 3 and Fig. 4 (middle). To assess the model’s performance, we report the diagnostic accuracy, which is comparable to recent reports in the relevant review literature (PubMed: 33942449 (AD vs. NC 74% on rs-fMRI Table 3)). We anticipate further improvements in accuracy by integrating additional AD biomarkers, such as Aβ and tau-PET data.
Regarding the concern of AD vs. CN classification as the early diagnosis tool, we apologize for the confusion. We are at the early stage of translating this SC-FC technique to early diagnosis of AD at EMCI stage. We specifically use AD vs. CN classification to drive the training since the functional alteration is supposed to be profound at the late stage, which allows us to demonstrate a clear pattern of group differences in functional alteration (Fig. 4 middle). We will make it clear in the final version.
Minor comments: We will correct the typos and rephrase the description for variables. We will include Atasoy’s work on connectome harmonics. The detailed parameter setting as follows: epoch=500, learning rate =0.0001, Adma optimizer, dropout=0.5, hidden dimension=64.
R2 In order to cater to the audience at MICCAI, we will rephrase the neuroscience terminology to make it more accessible and engaging. For our study, we specifically chose subjects from the ADNI dataset who had both structural and functional MRI data available. To evaluate the performance of our model, we employed a 5-fold cross-validation manner and reported the classification results on the testing data. Accept all suggestions.
R3 The reason we chose SVM is to mitigate the potential impact of confounders such as hyperparameters in more advanced ML models. The accuracy by random forests is 72.0% (compared to 68.5% by SVM and 75.9% by our method). We will upload the code to GitHub.
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.
In the first round of review, the manuscript received mixed scores. Particularly, R1 asked detailed questions on both algorithm clarification and applicability. The authors made satisfactory rebuttal to address these concerns. No reviewers updated their review after the rebuttal period. The AC acknowledged the novelty of the work and potential impact to the field and therefore recommended its publication in MICCAI 2023.
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.
This paper proposed a physics-guided graph neural network to identify disease induced changes in structural-functional connectivity coupling, and applied it on AD prediction. In general, it is an interesting study with clear presentation and certain technical merits.
The authors have provided detailed rebuttal to address the reviewers and AC’s concerns. A majority of reviewers retain positive comments on this paper. And I also recommend acceptance of this paper.
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 have addressed most of the concerns from reviewers. Although the applicability is still limited, applying Kuramoto model is definitely one of the main merits. Therefore, I would recommend accept.