List of Papers Browse by Subject Areas Author List
Abstract
In conjunction with graph neural networks (GNNs), functional connectivity analysis based on fMRI data can provide insights into the interaction and communication patterns in brain network, which has gained increasing attention in the diagnosis of neuropsychiatric disorders. However, traditional GNN based models focus primarily on brain regions, with limited attention given to changes in brain connectivity induced by diseases, and often lack specific methods to address noise and outliers. To accurately preserve and analyze connections in brain networks and retain the structure information in the original graph over message passing, we propose an Residual-Posterior Line Graph Network (RP-LGN). RP-LGN innovatively re-models each edge as a node to highlight functional connectivity information. Subsequently, it integrates residual blocks and a single-pass, low-variance Bayesian variational inference method to approximate the true posterior distribution. Bayesian variational posterior facilitates the quantification of uncertainty in model predictions and enhances model robustness in the presence of noise and anomalous data, ultimately promoting more accurate clinical decision-making. Compared with other models,the performance of RP-LGN was validated on the ABIDE and ADHD-200 dataset, with significant accuracy improvements, and revealed significant site-specific differences and unique connection patterns associated with diseases.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3697_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/YeDbae/RP-LGN
Link to the Dataset(s)
ABIDE: https://fcon_1000.projects.nitrc.org/indi/abide/abide_II.html
BibTex
@InProceedings{ZhaXin_The_MICCAI2025,
author = { Zha, Xinbei and Zhang, Jiaming and Gu, Jin},
title = { { The Refining of Brain Connectivity Features on Residual Posterior Patterns } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15971},
month = {September},
page = {648 -- 658}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a Residual-Posterior Line Graph Network (RP-LGN) for brain disorder diagnosis, which converts the edges into nodes to focus on critical brain connectivity patterns and incorporates the residual posterior patterns to facilitates the quantification of uncertainty in model predictions.
- Please list the major strengths of the paper: you should highlight 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 method proposed in this paper presents an idea of integrating the residual structure and the low-variance Bayesian variational inference, taking into account the changes in brain connectivity caused by diseases and the noise and outliers existing in the fMRI data.
- The performance of the model is evaluated on both the ABIDE and ADHD datasets.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- The paper is a straightforward applications of existing methods, lacking innovation.
- The method lacks detailed descriptions on the implementation approach, making it difficult to understand the proposed model.
- Due to the lack of detailed description in the methods, the proposed method appears challenging to replicate. Moreover, the incomplete ablation experiments make it hard to determine the effectiveness of each module.
- Please rate the clarity and organization of this paper
Poor
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
- The paper claims that the key contribution is the integration of the line graph convolutional network and Bayesian variational posterior. However, these techniques may be straightforward applications of existing ones, lacking innovation.
- The literature review is not enough sufficient, failing to cover relevant research on the variations of brain functional connectivity. e.g., “Spatio-Temporal Graph Hubness Propagation Model for Dynamic Brain Network Classification”. This may lead to the innovation of the paper not being established on the basis of a comprehensive understanding.
- As stated in the paper, ``The incorporation of residual posterior patterns improves the discriminative power and interpretability of functional connectivity matrices.” However, the paper does not provide a detailed explanation of how the proposed method improve the interpretability.
- There is a lack of ablation experiments on the line graph method, which cannot fully prove the effectiveness of the constructed line graph.
- There is a lack of a detailed description of how to use it to incorporate uncertainty measures and its actual effect on the model.
- Many symbols in this paper are unclear and mixed, which makes it difficult to follow the paper.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
Although the author reiterates the innovation of this paper in the rebuttal, it is still not convincing and solid and my major concen remains. Specifically, the author states that “this is the first instance that converts brain connections into nodes”. However, many existing edge-centric functional brain network studies [1-3] for very similar tasks have already explored modeling interactions between edges by considering them as nodes. In this case, the key contribution and the claimed novelty of this paper does not sound and the authors seem overclaim it. Moreover, the method of this paper is mainly a direct application of an existing method. I do not recommend acceptance considering these two factors.
[1]Joshua Faskowitz et al., Edge-centric functional network representations of human cerebral cortex reveal overlapping system-level architecture. Nature Neuroscience, 2020. [2]Youngheun Jo et al., The diversity and multiplexity of edge communities within and between brain systems. Cell Reports, 2021. [3]Ang Sun et al., Identifying autism spectrum disorder using edge-centric functional connectivity, Cerebral Cortex, 2023.
Review #2
- Please describe the contribution of the paper
The paper’s main contribution is the RP-LGN model. The authors change the way brain connectivity is represented by converting the standard brain graph into a line graph, where each connection is treated as its own node. This approach lets the model directly analyze the links between brain regions to better capture their interactions. In addition, the model uses residual connections along with a Bayesian method in the final layer to keep important structural details while also measuring uncertainty—a key factor when handling noisy data and small sample sizes typical in fMRI studies. Also, the authors tested the model on important neuroimaging datasets. Their results show that this new method can diagnose neuropsychiatric conditions more effectively than many existing approaches in terms of performance.
- Please list the major strengths of the paper: you should highlight 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.
One strength of the paper is its approach to modeling brain connectivity. Instead of simply looking at regions separately, the authors turn connections into individual nodes, which lets the method pick up on details in how regions interact—details that might be lost with more traditional methods. Another strong point is the use of Bayesian inference alongside residual connections. This choice not only helps in maintaining important information through the layers but also gives a measure of how certain the model is about its predictions. This is particularly useful when dealing with the noisy data and small sample sizes typical in fMRI studies. Finally, they back up their claims with solid experimental comparisons on well-known neuroimaging datasets, showing that its new technique really does translate into better performance in a clinical context.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
One drawback is that the paper doesn’t offer a very clear, intuitive reason for converting the connections into nodes—even though it’s a fresh idea, it isn’t fully justified compared with more traditional approaches. The authors could have strengthened their argument by comparing their method directly with similar techniques in prior work. Another issue is that transforming the graph into a line graph can add extra computational cost, but the paper doesn’t discuss how this might affect scalability or real-world use. Lastly, while the experimental results are solid, the paper could do more to show how these findings translate into clinical feasibility beyond just the diagnostic metrics.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I’m giving the paper a score of 4 (Weak Accept) for the following reasons. While it introduces a promising new approach, there are some concerns that need addressing. On the plus side, transforming connections into nodes and using Bayesian inference adds an interesting twist that appears to boost diagnostic performance on fMRI data. The experimental results are solid, and the consistent use of preprocessing and training protocols across models helps support the clinical relevance of the method. However, the rationale for converting connections into nodes isn’t entirely intuitive, and some of the mathematical details come off as quite dense. Additionally, the paper doesn’t fully address the potential computational overhead or how these modeling choices might translate into broader clinical feasibility. With a clearer justification of these aspects or more comparisons to related prior work, the paper could be stronger. Thus, I believe a 4 is appropriate, leaving room for improvement pending a detailed rebuttal.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
The authors proposed a Residual-Posterior Line Graph Network, which treats the partial correlations as the node and utilizes the first nearest neighbor to sparsify the graph. The graph neural network uses SAGE convolution and quantify the uncertainty of the model with Bayesian variational posterior. The authors have proven that the proposed model has better performance than other deep learning methods in classification task on ABIDE and ADHD-200 dataset.
- Please list the major strengths of the paper: you should highlight 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 authors reconstructed the brain graph in a different way that facilitates the model to capture the change of the connection induced by the disease.
- The authors have conducted comprehensive experiments on their model for comparison with the baseline models. They have used 2 open datasets and also separated the ABIDE dataset by the university that provided the brain scan, to avoid the confounding factors in the dataset.
- The authors integrated the model with GradCAM to enhance the interpretability of the model and could extract the most significant connections between the ROIs.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- There is no experimentation on the choice of k. It is understandable to use a small value of k to reduce the dimension in the graph. However, k=1 may be an overkill to oversimplify the graph.
- The authors didn’t compare the posterior loss with the binary cross entropy.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper proposes a novel approach to reconstruct brain graphs in a way that highlights disease-induced connectivity changes. The methodological contribution is thoughtful and addresses a key challenge in neuroimaging analysis—namely, how to more effectively model subtle alterations in brain connectivity associated with clinical conditions. The authors introduce a new way of reconstructing brain graphs that facilitates the model’s ability to capture connectivity changes driven by disease, which is a meaningful and non-trivial advancement. The experimental design is robust, with evaluations conducted on two publicly available datasets. Importantly, the authors also control for potential confounding factors by separating the ABIDE dataset based on the contributing institutions, which strengthens the reliability of their findings. The integration of GradCAM for post-hoc interpretability adds significant value, as it provides insight into which connections between regions of interest (ROIs) are most influential in the model’s predictions. This helps bridge the gap between performance and explainability, a crucial aspect in clinical applications.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors have replied and briefly explained the results of the missing ablation study. I would accept the paper if the code could be made public and those ablation studies could be replicated.
Author Feedback
Appreciate to the reviewers for their meticulous examination and insightful suggestions on our manuscript. Response to Common Questions: [1] Due to space limitations, some experiments have been condensed, and we will provide corresponding explanations in the following text. [2] We confirm that the code will be publicly available after acceptance. R#1 Q1:Reason for converting the connections into nodes A1:The proposal to convert brain connections into nodes is inspired by the cross-domain line graph models for link prediction. This transformation is fundamentally motivated by the need to explicitly represent edges as entities in GNNs - rather than merely as weight attributes. R#1 Q2 and R#2 Q4: Computional cost, explanation and clinical feasibility A2: Compared to conventional graph conv, line graph conv incurs only a minimal cost for graph transformation. With negligible computational cost added, it can provide more features of brain functional connectivity, and Fig. 3 illustrates the brain connections identified by the model as significantly contributing to ADHD recognition, supported by relevant references, thereby enhancing the model’s interpretability. Moreover, the line graph adaptively constructs individualized brain connectivity patterns for each subject. The minimal computational cost, medical interpretability, and adaptive connectome construction collectively demonstrates clinical feasibility. R#2 Q1:Innovations A1: Our work is not a straightforward combination of existing modules. To the best of our knowledge, this is the first instance that converts brain connections into nodes,which offered more details that might be lost with more traditional methods. Furthermore, while line graphs were initially introduced for link prediction, we integrate GraphSAGE to address a graph-level representation classification task. Concurrently, we have implemented adaptations to the line graph’s construction, aggregation, and convolution methods, tailored for brain representations. All code implementations for the line graph transformation were independently developed in this work, representing original contributions rather than adaptations of existing implementations. Integrating this with a Bayesian framework to address the challenges of noise and small sample sizes common in fMRI data, enhancing the robustness and generalizability of the model. R#2 Q2:Descriptions A2:The overall model is as described above, we will revise the method part more clearly in the final draft and promise to make the code public after acceptance. R#2 Q3: Relevant research A3:We appreciate your insightful suggestion. However, our framework fundamentally models connections as featured entities, which makes it no longer feasible to visualize the process of connectivity changes or assess inter-connection correlations in a matrix form, as is typically done with static or dynamic FC. Therefore, we have reduced the comparison with FC-related studies and instead focused on visualizing important connections in Fig. 3. R#3 Q1, Q2 and R#2 Q5: Ablation experiments A: Due to space limitations, some experiments in this paper have been condensed. For the selection of the K hyperparameter in KNN, our experiments revealed that increasing the number of edges did not significantly improve performance while increasing computational costs; therefore, these results were not reported.Additionally, Fig. 2 presents the ablation results for the residual, posterior loss, and convolutional block modules. In Fig. 2, the central table represents models excluding posterior loss (P), which utilize cross-entropy loss for comparison. Regarding line graph ablation, we only present the results of employing different convolution methods after line graph construction; previous experiments indicated that using only residual posterior patterns with a standard GCN as the model yielded neither favorable results nor the ability to assess connection importance.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
Although the authors have addressed some reviewer concerns by clarifying their motivation and computational considerations, the manuscript still appears to somewhat overstate the novelty of converting brain connections into nodes, as pointed out by R2. I lean towards recommending rejection at this point, suggesting the authors better articulate and substantiate their specific methodological advances in any future revisions.