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
Alzheimer’s disease (AD) is characterized by the propagation of tau aggregates throughout the brain in a prion-like manner. Tremendous efforts have made to analyze the spatiotemporal propagation patterns of widespread tau aggregates. However, current works focus on the change of focal patterns in lieu of a system-level understanding of the tau propagation mechanism that can explain and forecast the cascade of tau accumulation. To fill this gap, we conceptualize that the intercellular spreading of tau pathology forms a dynamic system where brain region is ubiquitously wired with other nodes while interacting with the build-up of pathological burdens. In this context, we formulate the biological process of tau spreading in a principled potential energy transport model (constrained by brain network topology), which allows us to develop an explainable neural network for uncovering the spatiotemporal dynamics of tau propagation from the longitudinal tau-PET images. We first translate the transport equation into a backbone of graph neural network (GNN), where the spreading flows are essentially driven by the potential energy of tau accumulation at each node. Further, we introduce the total variation (TV) into the graph transport model to prevent the flow vanishing caused by the l_2-norm regularization, where the nature of system’s Euler-Lagrange equations is to maximize the spreading flow while minimizing the overall potential energy. On top of this min-max optimization scenario, we design a generative adversarial network (GAN) to depict the TV-based spreading flow of tau aggregates, coined TauFlowNet. We evaluate TauFlowNet on ADNI dataset in terms of the prediction accuracy of future tau accumulation and explore the propagation mechanism of tau aggregates as the disease progresses. Compared to current methods, our physics-informed method yields more accurate and interpretable results, demonstrating great potential in discovering novel neurobiological mechanisms.
Link to paper
DOI: https://doi.org/10.1007/978-3-031-43898-1_8
SharedIt: https://rdcu.be/dnwAF
Link to the code repository
N/A
Link to the dataset(s)
N/A
Reviews
Review #2
- Please describe the contribution of the paper
This paper proposes a novel physics-informed deep neural network, named TauFlowNet, to discover spreading flow of tau propagation and achieve promising result by comparing to current 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.
- Conceptualizing the intercellular spreading of tau pathology as a dynamic system and formulating it as a potential energy transport model constrained by brain network topology.
2/ Developing an explainable neural network called TauFlowNet to uncover the spatiotemporal dynamics of tau propagation from longitudinal tau-PET images.
-
Introducing the total variation (TV) into the graph transport model to prevent flow vanishing caused by ℓ2-norm regularization and designing a generative adversarial network (GAN) to depict the TV-based spreading flow of tau aggregates.
-
Evaluating TauFlowNet on the ADNI dataset to predict future tau accumulation and exploring the propagation mechanism of tau aggregates as the disease progresses.
-
Demonstrating that the proposed method yields more accurate and interpretable results compared to current methods and has the potential to discover novel neurobiological mechanisms.
- 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 Fig.3 looks good, however, it could be better if I can see more quantity evaluation on those ROI.
- Please rate the clarity and organization of this paper
Excellent
- 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 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/2023/en/REVIEWER-GUIDELINES.html
This paper proposed a novel physical-based model discover spreading flow of tau propagation. And it fixes over-smoothing issue during the training of GNN. In the evaluation part, this paper show the model expandability of region-to-region spreading flow.
- 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?
I give this score mainly based on the soundness of the methodology and the experimental design.
- 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
7
- [Post rebuttal] Please justify your decision
I believe this MICCAI paper is well organized and shows enough contribution in medical image field.
Review #3
- Please describe the contribution of the paper
The paper proposes a model to predict the dynamics of tau propagation, an indicator of Alzheimer’s disease, from PET images. The paper is using calculus of variation to derive the flow equations (the Euler-Lagrange PDEs), then solves it as an iterative message passing process via simple Graph Neural Network framework. The GNN is latent representation of the flow map solution, so it also uses a GAN to generate the tau flow maps from the encoding of the GNN. The model is evaluated on a public dataset, the ANDI dataset, and as ablation study is conducted.
- 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.
• Tau-PET is an important measurement for Alzheimer’s disease. • The model is taking a calculus of variation perspective, and attempt to solve the E-L equations within a leaning model. • Evaluation is done on public dataset, and providing the code is released, future papers can compete against this model.
- 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 technical novelty is limited to the derivation of the E-L equation for this problem. Beyond that, the DNN model seems to be rather standard. Also, I am not convinced that a deeper MLP would be as good as the GNN model. Since no other models exist to the proposed problem, and comparisons are only made against trivial versions, it is hard to assess effective the model is for practical use.
- 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 model seems to be easy to implement and reproduce. Data preparation might be harder to replicate, and I think that publishing supporting code would be very helpful.
- 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
Figures should be enlarged and highlighted. I am not sure if technical novelty meets the MICCAI standard, or if the paper fits a more applicative venue.
- 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?
I am not sure whether technical novelty meets the MICCAI standard.
- Reviewer confidence
Somewhat 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
I have read the feedback and other reviews, my concerns on technical novelty (Sec. 6) still remain, but I’m ok with accepting if this is the majority vote.
Review #4
- Please describe the contribution of the paper
The authors propose Tau FLowNet, a GAN incorporating GNN and TV-based optimisation for solving the differential equation for tau spreading in the brains of Alzheimer’s disease subjects.
They formulate the process of tau spreading in a potential energy transport model over graph framework and modify the classical diffusion along graph equation by replacing l2 with TV norm in its associated functional form. This could, according to the authors, prevent the over smoothing problems that typically appear in GNN. The overall optimisation scheme is embedded in a GAN.
- 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 paper is well written, clear, and proposes an interesting method, which is well characterised and described. It is technically sound and the proposed method, with respect to the specific application of tau spreading in AD, is definitely original.
- 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.
However, I think the experimental section has two main weaknesses:
1) the authors have compared their method against standard NN-based approaches for prediction of end-stage tau patters in CN, MCI and AD subjects but have not compared against other ode/pde-based models for tau spreading (e.g. the model in [8,9] or in Weickenmeier et al, J Phys, 2019; or Garbarino et al, Neuroimage, 2021 - by the way, the last two are other DS models for protein propagation in AD which need citing). Specifically, not all these models assume linear processes (as you declare in the introduction) but rather some reaction-diffusion processes for protein pathology. They do not allow for directionality, as your method remarkably does, but still I believe a comparison on the predicted MSE is due, for at least one of the models (e.g. the ESM is open at https://github.com/illdopejake/data_driven_pathology/blob/master/esm/ESM_utils.py)
2) I wonder whether you could show the value of the estimated propagation parameters of the model (\alpha, \phi) and maybe comment on their meaning/reliability? It seems to me you are showing results on the directionality of the estimated flows but you could also look at the estimated parameters and comment on their meaning wrt the different clinical groups.
- 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
OK
- 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, see my previous list of comments on strengths and weaknesses.
- 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 find the method quite interesting and the topic actual and relevant. I suggest few more tests and discussion of the results (see above).
- 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
7
- [Post rebuttal] Please justify your decision
My concerns were mainly: 1) comparison with other popular and specific methods (NDM, GRAND, PINN..) and 2) clinical interpretation of the learnt parameters.
I believe the authors have excellently replied to both of my concerns and I look forward to reading the complete paper now.
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.
summary
- physics-informed deep neural network, named TauFlowNet, to discover spreading flow of tau propagation. I beleive that paper is interesting but more comparison is necessary. The code for the other method is publically available, so why compare it?
strength
- Developing an explainable neural network called TauFlowNet to uncover the spatiotemporal dynamics of tau propagation from longitudinal tau-PET images
weakness
- As pointed out by R4. Comparison with other method is necessary
- The other point by R4 is important “It seems to me you are showing results on the directionality of the estimated flows but you could also look at the estimated parameters and comment on their meaning wrt the different clinical groups.”
Author Feedback
We thank for all constructive comments. We will incorporate all feedback in the final version. TauFlowNet code has been released in anonymous Github.
First, we address two comments from meta-review.
For lack of comparisons: Due to the page limit, we only compared our model to popular methods such as GCN and DNN. The comprehensive comparisons with other benchmarks, including NDM (mentioned by R4), GRAND (PDE-based model), GCNII, GAT, ResGCN, and DenseGCN, are shown below. We use 5-fold cross-validation along with std of noise level from 0 to 1 (std=0 std=0.02 std=0.04 std=0.08 std=0.1 std =1).
NDM (0.11±0.02, 0.12±0.03, 0.14±0.03, 0.14±0.03, 0.17±0.04, 0.18±0.04) GRAND (0.27±0.04, 0.28±0.04, 0.28±0.04, 0.28±0.04, 0.29±0.04, 0.29±0.05) CCNII (0.23±0.04, 0.23±0.04, 0.23±0.03, 0.23±0.04, 0.23±0.04, 0.23±0.04) GAT (0.41±0.05, 0.41±0.05, 0.41±0.05, 0.41±0.04, 0.41±0.05, 0.65±0.03) ResNet (0.09±0.02, 0.09±0.03, 0.09±0.02, 0.11±0.02, 0.11±0.02, 0.19±0.02) DenseNet (0.08±0.02, 0.09±0.02, 0.10±0.02, 0.09±0.02, 0.11±0.01, 0.12±0.02) Compared with the results in Table 1, it is evident that our method achieves promising performance over current GNN methods. We will show the Github link (the code for all methods) and include all results in final version.
For the implications of parameters in clinic applications: We appreciate the great suggestion. \alpha is a learnable parameter that constrains the estimation of spreading flows. Since the concept of flow is more interpretable from the perspective of biology and physics, we mainly investigate the neuroscience underpinning of flow in disease progression. In addition to Fig. 3, we examined both the inward and outward spreading flows between the entorhinal cortex and subcortical regions. We found significant group differences in inward vs. outward flow deficit across different clinical groups (CN: 78.99 vs 7.42 EMCI: 46.23 vs 3.85 LMCI: 11.07 vs 1.33 AD: 7.04 vs 1.96), which has great potential for putative biomarkers in AD. On the other hand, \phi denotes a non-linear mapping that connects the external environment, including factors like tau accumulation, to the disease-specific states using a MLP. Directly correlating \phi to clinical phenotypes can be a challenging task. However, it is highly valuable to explore the relationship between tau and other pathologies, such as amyloid and neuroinflammation biomarkers, by analyzing the interactions facilitated by \phi.
R1 Thanks. We will add the flow numbers of each ROI to Fig. 3.
R3 For the contributions, we would like to emphasize the following points here:
- We introduce a novel framework that leverages the calculus of variations and the Euler-Lagrange (E-L) equation to derive innovative GNN model. This framework enables us to gain insights into the behavior of GNN models in a continuous space of graph functionals.
- We identify and address the issue of “over-smoothing” in GNNs (underlines the l_2-norm graph diffusion process). We propose a TV-based solution to mitigate this problem. Although we specifically employ TV to alleviate the vanishing flow issue in this paper, our approach can be scaled up and applied to other GNN applications as well.
- We achieve good performance in predicting future tau aggregates compared to existing GNN methods (please see our response to the MetaReview).
For the concern of GNN underperforming DNN, let us clarify the reasons as follows. While DNN takes whole-brain features (in R^N) as input, conventional GNNs initially process scalar features, specifically regional tau SUVR, in the MLP layer. Thus, the result of GNN is inferior to that of DNN due to the limited availability of rich information. In this context, although GNNs leverage graph topology, GNN shows comparatively lower performance than DNN.
R4 Thanks. We have compared NDM and another PDE-based model (GRAND). The performance is listed in response to MetaReview. We will cite other two papers on DS models.
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 main point raised by R4 is addressed.
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 is an interesting paper proposing a physics informed deep neural network for the prediction of tau propagation. The approach is novel and is addressing an important problem. The rebuttal addresses main concerns. The paper is a solid contribution to MICCAI.
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 paper presents a physics-inspired/derived GNN to discover/analyze tau-spread from longitudinal tau-PET scans in the ADNI cohort. I think that’s an interesting idea and the paper does not seem to contain any major flaws. R3 raised some concerns regarding the novelty of the method, which I think could have been better addressed by the authors in the rebuttal. Aside from this aspect, I believe that the rebuttal sufficiently clarifies all other major issues flagged by the reviewers and I, therefore, think it is an interesting contribution.