List of Papers Browse by Subject Areas Author List
Abstract
Multi-modal neuroimaging studies are essential for exploring various brain disorders; however, they are typically limited in sample size owing to the cost of image acquisition. Meta-analysis is an underutilized method that integrates the findings from multiple studies derived from large samples to assist individual studies. Neuroimaging studies are increasingly adopting transformer architecture for network analysis; however, they tend to overlook local brain networks. To address these gaps, we propose the Meta-analysis Enhanced Graph Attention TransFormer (MEGATF), a novel method for performing multimodal brain analysis built on a graph transformer framework aided with meta-analysis information derived from NeuroSynth. Our method adapts a graph neural network with a transformer attention mechanism that favors local networks and multimodal interactions using PET or cortical thickness. Our method achieved a state-of-the-art classification performance on mild cognitive impairment and attention-deficit/hyperactivity disorder datasets, distinguishing individuals with brain disorders from controls. Furthermore, it identified disease-affected brain regions and associated cognitive decoding that aligned with existing findings, thereby enhancing its interpretability. Our code is at https://github.com/gudtls17/MEGATF.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2372_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/gudtls17/MEGATF
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ChoHyo_Integrating_MICCAI2025,
author = { Choi, Hyoungshin AND Kim, Sunghun AND Lee, Jong-eun AND Park, Bo-yong AND Park, Hyunjin},
title = { { Integrating meta-analysis in multi-modal brain studies with graph-based attention transformer } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15971},
month = {September},
page = {416 -- 425}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes MEGATF, a graph transformer framework that integrates meta-analysis data from NeuroSynth to enhance multimodal brain analysis. By incorporating local brain network structures and multimodal inputs like PET and cortical thickness, MEGATF addresses sample size limitations and improves classification of brain disorders such as MCI and ADHD. It also identifies disease-relevant brain regions and cognitive functions consistent with existing studies, enhancing interpretability.
- 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.
This paper’s key strengths include a integration of meta-analysis data into a graph transformer framework, enhancing multimodal brain analysis with improved interpretability and robustness. It models local brain networks using edge features that capture region-by-region multimodal interactions, addressing limitations of previous transformer methods. The proposed MEGATF model shows clinical relevance by classifying brain disorders such as AD and ADHD, while identifying meaningful disease-related regions consistent with prior studies.
- 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 authors did not introduce methodological innovations, and they primarily employed existing methods to model the edge-wise connections between brain regions. This work is somewhat lacking in methodological novelty. In addition, the compared methods are all based on Transformer architectures, without including methods related to multimodal brain network construction or fusion.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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
(1) The formatting of all equations and variables in this paper should be carefully checked and revised. Matrix symbols should be in bold uppercase, vector symbols in bold lowercase, and scalar quantities in italic lowercase. However, this convention is not consistently followed throughout the manuscript and needs to be corrected accordingly. (2) The generation of term-weighted node features is based on combining structural imaging with cognitive activation maps through weighted averaging, to obtain the representation of each brain region across different cognitive dimensions. However, does this method overlook individual variability? Moreover, since a single brain region may contain multiple functional subregions, does the use of simple weighted averaging neglect the structural diversity within each ROI? (3) Three modalities are ultimately fused and processed as input to the Edge-GAT, where node features are derived from structural images combined with cognitive activation maps, and edge features are based on functional connectivity. What is the rationale behind constructing the graph in this way? (4) Among the compared SOTA methods, there are no other methods specifically designed for multimodal brain network analysis. Please include comparison results with related multimodal methods to better demonstrate the advantages of the proposed method. (5) The formatting of the references is inconsistent, some journal names are written in all uppercase letters while others are in all lowercase. Please check and revise them for consistency.
- 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.
(3) Weak Reject — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The authors did not introduce methodological innovations, and they primarily employed existing methods to model the edge-wise connections between brain regions. This work is somewhat lacking in methodological novelty. In addition, the compared methods are all based on Transformer architectures, without including methods related to multimodal brain network construction or fusion.
- 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 #2
- Please describe the contribution of the paper
In this study, a prior knowledge guided graph attention transformer model was proposed for brain disease classification using mutli-modal neuroimaging data, by incorporating meta-analysis information and graph attention transformer.
- 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.
Prior knowledge from meta-analysis was used to guide feature extraction.
- 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.
Limited technical novelty in the graph-based attention transformer.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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?
- Node features. The node features were from PET or cortical thickness only, why not also including fMRI based features for multi-modal learning?
- Meta-analysis information. How were the 64 terms selected to get the activation maps? It’s better to provide a demonstration of the selected terms. BTW, it seems that the activation maps included not only cognition-related terms but also other behavioral terms (e.g., emotion related), as shown in Fig.2. Maybe “cognition-related term” should be updated with other description.
- It is not clear how the region-level and term-level importance were obtained.
- In Table 3, the model w/o edge feature (1st model) had better performance than models with FC or NFC only (2nd and 3rd model). Some discussion regarding this would be helpful.
- 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 present work introduces several key methodological and conceptual advancements in the field of computational neuroimaging, with a focus on enhancing multimodal brain network analysis through meta-analytic integration, graph-based attention mechanisms, and clinically interpretable findings. The principal contributions are delineated as follows.
-
Augmentation of Multimodal Neuroimaging via Meta-Analytic Synthesis Prior research in brain network analysis has predominantly relied on single-cohort or unimodal datasets, potentially limiting generalizability. To address this constraint, a meta-analytic framework was incorporated, enabling the aggregation of findings from diverse neuroimaging studies. This approach facilitated a more comprehensive representation of brain connectivity patterns by leveraging existing large-scale datasets. Consequently, the robustness and external validity of the derived models were significantly improved, ensuring that the analytical framework accounted for cross-study heterogeneity while preserving biologically meaningful signals.
-
Graph Neural Network-Based Attention Mechanism for Dynamic Multimodal Fusion A critical innovation of this study was the development of an attention-based graph neural network (GNN) architecture designed to model both localized brain network dynamics and cross-modal interactions. Traditional methods often treat brain regions in isolation or impose rigid connectivity constraints, neglecting the hierarchical and context-dependent nature of neural processing. The proposed GNN attention mechanism dynamically weighted intra-regional and inter-modal relationships, allowing for adaptive information flow across structural and functional neuroimaging modalities. This design not only enhanced predictive accuracy but also provided an inherently interpretable representation of how different brain systems interact.
-
Empirical Validation and Neuropathological Correlates in Brain Disorders The efficacy of the proposed framework was rigorously evaluated against state-of-the-art alternatives, with results demonstrating consistent superiority in both discriminative performance and interpretability. Beyond methodological advancements, the study yielded clinically relevant insights by identifying disorder-specific neural signatures associated with Alzheimer’s disease (AD) and attention-deficit/hyperactivity disorder (ADHD). These findings aligned with existing neurobiological evidence while also uncovering novel connectivity patterns, thereby contributing to a deeper understanding of the pathophysiological mechanisms underlying these conditions.
-
- 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.
In summary, the strength of this research lies in its advancement of multimodal brain network analysis through the integration of meta-analytic data, a novel GNN-driven attention mechanism, and clinically grounded interpretations. The proposed framework not only improved computational modeling in neuroimaging but also bridged the gap between data-driven discoveries and neuroscientific understanding, offering a robust tool for future investigations into brain disorders.
- 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.
A notable limitation in the described methodology is the heterogeneity in dataset composition and relatively small sample sizes for certain classification tasks, which may affect the generalizability and robustness of the findings.
- Imbalanced and Small Cohort Sizes
Task 1 (MCI vs. CN): 117 MCI vs. 132 CN Task 2 (sMCI vs. pMCI): 39 sMCI vs. 47 pMCI Task 3 (ADHD vs. CN): 86 ADHD vs. 68 CN While the sample sizes for MCI vs. CN and ADHD vs. CN are moderately balanced, the progression prediction task (sMCI vs. pMCI) suffers from a very limited sample size (N=86), increasing the risk of overfitting and reducing statistical power. Small datasets in neuroimaging studies can lead to inflated performance metrics due to high variance, making it difficult to ascertain whether the model’s success is due to genuine biological signals or random noise.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
It provided solid and novel input to science of imaging with little or no hard objections from me.
- Reviewer confidence
Very confident (4)
- [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
Author Feedback
We thank the reviewers for valuable comments. Below we address concerns. Q1(R2-1&R4-3): Rationale for node and edge features. A: 1)PET/cortical thickness reflects features that are largely confined within a given region, while functional connectivity(FC) from fMRI reflects inter-regional interactions. Thus, we chose the former as the node feature and the latter as the edge feature. PET/cortical thickness provides localized information that can benefit from large-scale functional association of meta-analysis. However, FC and meta-analysis results are both from fMRI possibly with a large overlap and thus we chose not to enhance FC with meta-analysis. Our enhanced node feature takes a form similar to the fMRI timeseries and is further used to construct node feature connectivity(NFC). Both FC and NFC were used to compute edge features ultimately fusing all three modalities. 2)Our framework uses both node and edge features for learning the weights of the graph.
Q2(R2-2): Rationale for term selection. A: We queried disease-specific terms(Alzheimer’s and ADHD) in Neurosynth and obtained corresponding meta-analytic activation maps. Each map was correlated with other maps tied to a specific term and we chose the top 64 terms. Some of the chosen terms were indeed behavioral. We will replace “cognition-related terms” with “cognition and behavior terms”.
Q3(R2-3): Details on region and term-level importance. A: We define importance using the output of the final layer before the readout layer. The output X’ is NxD, where N is the number of regions and D is the number of terms. Row-wise sum leads to regional and column-wise sum leads to term importance.
Q4(R2-4): Ablation on edge feature. A: The model without explicit edge features slightly outperforms those using only a single edge type(Table 3), suggesting that naively adding handcrafted connectivity measures may introduce noise or limit flexibility. However, our method, which integrates multiple edge features from different modalities, consistently achieves the best performance. This shows that incorporating diverse edge features allows the model to better capture the complex, multi-faceted nature of brain networks.
Q5(R3-1): Imbalance and small cohort size. A: Our datasets are widely used open datasets(ADNI and HBN). Still, they are small. We plan to use larger datasets in future work.
Q6(R4-1,5): Equation, variables, and reference format. A: We will correct these as suggested.
Q7(R4-2): Individual variability and structural diversity. A: Structural features(PET/cortical thickness) are individual measures, while the cognitive maps are group measures. We combine the two by voxel-wise weighted mean of the cognitive map with the structural feature. Due to the individual variability in the structural features, the combined feature still retains the individual variability. As pointed out, one ROI may have many distinct subregions. This can be handled by using a multi-scale atlas such as Schafer, which offers 100, 200, or 400 regions. Results derived from the fine-scale version could depict these subregions within an ROI. This will be our future work.
Q8(R4-4): Model comparison and methodological innovations. A: 1)We compared our model with a GNN-based model(BrainNetGNN) with multimodal input[17] and transformer models. This is because the transformer models performed well in brain network analysis and both GNN and transformers effectively model regional interactions. GNNs do this by graph structure and transformers use attention-based token interactions. 2)Our main innovation lies mostly with feature definition. Although we employed existing methods to model the edge-wise connections between brain regions, our EdgeGAT module includes attention layers more tailored for brain network analysis than the conventional multi-head attention mechanism of the transformer. More importantly, we used the meta-analytic information to enhance structural features for better brain network analysis.
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.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A