List of Papers Browse by Subject Areas Author List
Abstract
Accurate and interpretable assessment of cognitive functions from neuroimaging data is critical for understanding cognitive health. Existing methods either rely on subjective assessments or struggle to capture the complex, time-varying interplay between brain regions, hindering both predictive accuracy and neurophysiological interpretability. In this paper, we introduce BrainWeaver, a novel framework designed to overcome these limitations by weaving task-relevant neural patterns from dynamic connectomes. Specifically, BrainWeaver uses an ROI-guided dynamic multi-graph fusion module to capture spatial dependencies and temporal fluctuations in fMRI data. To enhance interpretability, we integrate neuroscience priors via a hierarchical graph representation module that reinforces key ROI contributions. Moreover, we employ a contrastive learning paradigm to disentangle cognitive patterns from modality-specific noise and inter-subject variability. Extensive experiments on the large-scale HCP S1200 benchmark show that BrainWeaver significantly outperforms state-of-the-art baselines across three distinct cognitive domains, setting a new standard for interpretable dynamic brain network-based cognitive assessment.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3425_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/ZJUDataIntelligence/BrainWeaver
Link to the Dataset(s)
HCP S1200 dataset: https://www.humanconnectome.org/study/hcp-young-adult/data-releases
BibTex
@InProceedings{LiuTao_BrainWeaver_MICCAI2026,
author = { Liu, Tao AND Zhu, Fanwei AND Han, Haijun AND Zheng, Zengwei AND Zhou, Binbin},
title = { { BrainWeaver: Weaving Dynamic Brain Networks with ROI-Guided Attention for Cognitive Assessment } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16894},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes BrainWeaver, a framework for cognitive assessment from fMRI that combines dynamic graph attention, ROI-guided hierarchical representations, and contrastive learning. The main contribution is the integration of these components into an end-to-end model aimed at capturing spatiotemporal brain dynamics while improving interpretability and robustness.
- 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.
- Clear and well-structured paper The method, experiments, and ablations are easy to follow, and the overall structure is solid.
- Well-motivated design The combination of dynamic graphs, ROI priors, and contrastive learning is reasonable and fits the problem setting well.
- Good experimental coverage Evaluated on multiple cognitive tasks with a range of baselines. The ablation study is helpful and shows the contribution of each component.
- Interpretability The ROI importance and connectivity analyses are a nice addition and align with known neuroscience patterns.
- 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.
- Small performance gains Improvements over baselines are marginal and sometimes within standard deviation; no statistical significance analysis is provided.
- Overstated claims on stability The claim of “consistently lower standard deviations” is not fully supported by the results.
- Insufficient data description While the paper mentions the Neurograph benchmark, key details are missing (e. g. , number of subjects used, data splits, definition of prediction targets).
- Minor typos Minor writing issues (e. g. , Section 2.1: ‘an ROI-guided’ → ‘a ROI-guided’)
- 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.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/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 — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper is solid overall: it is well-written, clearly structured, and addresses a relevant problem in fMRI-based cognitive assessment. The proposed method is reasonable, and the experimental evaluation is generally thorough, including ablation studies and interpretability analyses.
However, the main limitation is that the contribution appears largely incremental, combining existing components rather than introducing a clearly novel formulation. While the method shows consistent improvements over baselines, the gains are small and in some cases within reported standard deviations, and no statistical significance analysis is provided. In addition, missing details in the dataset description (e.g., subject counts and data splits) reduce clarity and reproducibility.
Overall, this is a competent and complete study, but the level of novelty and is limited.
- Reviewer confidence
Somewhat confident (2)
- [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
This paper proposes BrainWeaver, a framework for cognitive score prediction from dynamic fMRI brain networks. It combines three modules: (1) ROI-guided dynamic multi-graph temporal attention fusion, which learns attention weights across temporal graph snapshots via intra-graph spatial attention and inter-graph temporal attention to produce a fused spatiotemporal graph representation; (2) Key ROI-specific hierarchical graph representation, which initializes importance weights using neuroscience priors (Multiple Demand system) and performs neuroscience-guided graph pooling; and (3) Individual invariance contrastive learning, which aligns fMRI representations with clinical metadata via symmetric InfoNCE to reduce modality-specific noise. Evaluated on HCP S1200 (~1200 subjects) for three continuous cognitive score regression tasks: working memory (ListSort), inhibitory control (Flanker), and general cognitive status (MMSE).
- 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.
Clinically valuable problem formulation. Continuous cognitive score regression across three cognitive domains is more informative and clinically actionable than binary classification. Predicting specific cognitive scores enables tracking of cognitive trajectory and early detection of decline, which has direct clinical utility beyond diagnostic labels.
Principled use of neuroscience priors. The Multiple Demand (MD) system theory [Duncan, 2010] — which posits that diverse cognitive tasks share a core neural scaffold anchored in prefrontal and parietal cortex — provides well-established neuroscience grounding for the key ROI selection. The differential initialization (Eq. 5: key ROIs=1, non-key=ε) is an elegant compromise that preserves prior knowledge while allowing data-driven refinement, avoiding the brittleness of hard constraints.
Comprehensive and convincing interpretability analysis. Fig. 2 and Fig. 3 together provide both node-level (regional importance) and edge-level (topological connectivity) interpretability across all three tasks. The distinct spatial patterns are neuroscientifically plausible: working memory shows a hub-centric pattern centered on left PFC with fronto-parietal connections; inhibitory control shows a balanced mesh topology between frontal eye fields, PFC, and parietal regions reflecting the attention network; general cognitive status shows widespread DMN and associative cortical involvement. These distinct patterns across tasks provide evidence that the model captures task-specific neural substrates rather than a generic signature.
Large-scale evaluation with thorough baselines. HCP S1200 (~1200 subjects) is substantially larger than most brain network studies. The comparison against 11 baselines spanning both static (GCN, GIN, GAT, GraphSAGE) and dynamic (ASTGCN, DySAT, EvolveGCN, TGCN, STAGIN, STNAGNN, Neurograph) architectures is thorough and well-organized. The consistent superiority of dynamic over static models validates the importance of temporal modeling.
- 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.
Improvements over SOTA are marginal and lack statistical significance testing — the most critical concern. BrainWeaver vs. Neurograph: Working Memory RMSE 13.40±0.34 vs. 13.60±0.49 (1.5% improvement — confidence intervals overlap); Inhibitory Control RMSE 10.26±0.28 vs. 10.30±0.25 (0.4% improvement — well within noise); General Cognitive Status RMSE 1.08±0.08 vs. 1.13±0.15 (4.4% improvement — more convincing but still moderate). No statistical significance tests (paired t-test, Wilcoxon, or bootstrap confidence intervals) are provided. Without p-values, the claimed SOTA cannot be substantiated. This is a critical omission for a paper whose primary claim is superior performance. The consistently lower standard deviations of BrainWeaver suggest improved stability, but stability alone does not constitute a sufficient contribution.
Each component represents incremental variation of existing techniques. (a) The temporal attention fusion (Eq. 1-4) is an attention-weighted averaging over temporal graph snapshots, similar to DySAT [21] and STAGIN [14]. The specific attention computation differs in using separate graph-attention for query/key (Eq. 2), but this is a minor architectural variation. (b) The hierarchical graph representation (Eq. 5-6) uses prior-guided initialization followed by MLP-based importance learning and GAT-based clustering — conceptually similar to DiffPool with a prior-informed initialization. The paper does not compare against or discuss these related pooling methods. (c) The contrastive learning (Eq. 7) is standard symmetric InfoNCE without modification.
The “clinical metadata” component raises a potential information leakage concern. The contrastive learning module aligns fMRI representations with “clinical metadata” (Section 2.3), but the specific variables constituting this metadata are never disclosed. If any cognitive-correlated variables (e.g., education level, other cognitive test scores, age) are included, the model gains access to information correlated with the prediction target during training, constituting information leakage. This must be clarified — it could fundamentally invalidate the entire experimental evaluation.
Critical experimental details are missing, compromising reproducibility. (a) The key ROI set R is not listed — which specific brain regions from the Schaefer parcellation are designated as MD system nodes? (b) The number of temporal windows M, window size, and stride for dynamic graph construction are not stated. (c) Hyperparameters ε, τ, λ are not specified and have no sensitivity analysis. (d) Variable n in Eq. (2) is undefined. These omissions make independent reproduction impossible despite code availability.
The ablation study reveals inconsistencies that are not discussed. Table 2 shows that removing the Fusion module barely affects General Cognitive Status performance (RMSE: 1.08 vs. 1.06 for full model — in fact, removing Fusion slightly improves RMSE). This suggests the temporal fusion module contributes inconsistently across tasks, contradicting the claim of a universally beneficial framework. The paper should honestly discuss task-dependent utility rather than claiming uniform superiority.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/en/REVIEWER-GUIDELINES.html
The paper exclusively evaluates on healthy young adults (HCP). Given that BrainWeaver is motivated as a cognitive assessment tool, evaluation on a clinical population (e.g., MCI patients, elderly cohorts) would substantially strengthen the clinical relevance claims. The regression framework handles three tasks independently. A multi-task learning formulation predicting all three cognitive scores simultaneously could leverage shared neural substrates and potentially improve performance. The contrastive learning module deserves more careful analysis: what is the effect of different metadata variables? Does removing specific variables change performance?
- 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 — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
BrainWeaver addresses a clinically relevant problem with a well-motivated framework and provides thorough interpretability analysis grounded in neuroscience. However, three major concerns prevent recommendation: (1) Improvements over Neurograph are marginal (0.4-4.4%) and lack statistical significance testing — without p-values, the SOTA claim is unsubstantiated. (2) Potential information leakage through unspecified clinical metadata in the contrastive learning module could invalidate the experimental evaluation. (3) Each component (temporal attention, hierarchical pooling, InfoNCE) is an incremental variation of existing techniques without sufficient differentiation. If the authors can (a) provide statistical significance tests demonstrating meaningful improvements, (b) fully disclose clinical metadata and rule out information leakage, and (c) provide ablation on hyperparameters (ε, τ, λ, M), the paper could reach the acceptance threshold.
- 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 paper proposes a framework, BrainWeaver, for predicting cognitive scores from dynamic brain connectomes from rs-fMRI connectivity. The framework components are: (1) an ROI-guided multi-graph temporal attention fusion module that aggregates a sequence of dynamic brain graphs using learned spatial and temporal attention, (2) a hierarchical graph representation module that uses neuroscience priors (the multiple demand system) to bias node importance toward known prefrontal and parietal ROIs and then pools via graph clustering, and (3) a contrastive learning module that aligns fMRI graph embeddings with clinical metadata to reduce inter-subject variability. The framework is evaluated on three cognitive prediction tasks from the HCP S1200 dataset (working memory, inhibitory control, general cognitive status), where it achieves some improvements over baselines.
- 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 hierarchical graph representation module (Section 2.2, Eq. 5–6) initializes ROI importance weights differently for multi-demand system nodes versus other ROIs. This is a specific and well-motivated design choice grounded in established cognitive neuroscience (Duncan, 2010). The ablation in Table 2 confirms this matters, where removing it causes the largest degradation across all three tasks (e.g., inhibitory control RMSE rises from 10.26 to 12.02).
- Overall, the ablation studies are thorough (Table 2). The authors remove each of the three components individually and results show all three contribute significantly to improving performance, with a clear ordering (Prior > CL > Fusion).
- The baselines are also thorough (Table 1), with comparing BrainWeaver against 11 methods from static GNNs (GCN, GIN, GAT, GraphSAGE) to dynamic spatiotemporal models (EvolveGCN, DySAT, TGCN, ASTGCN, STAGIN, STNAGNN, Neurograph).
- Figure 2 shows distinct spatial importance maps per cognitive domain (fronto-parietal for working memory, distributed frontoparietal-salience for inhibitory control, DMN for general cognitive status). These align with known clinical literature on functional neuroanatomy. Figure 3 provides a complementary topological view.
- For reproducibility, the authors shared an anonymous code repository (including the weights of the trained BrainWeaver model).
- 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.
- I couldn’t find the exact paradigm of fMRI used in the entire paper; the authors should explicitly state resting-state fMRI and refer to it as such. When you reference HCP datasets or mention dynamic, it’s very important to distinguish resting vs. task-based fMRI. Also mention the # of ROIs/parcellation used so readers do not have to go searching another paper.
- Looking at Table 1, the improvements shown between BrainWeaver and Neurograph are marginal (at most is the 4.4% RMSE drop for the general cognitive status task). There’s also no reported statistical significance testing, and considering the overlapping standard deviations between Brainweaver and some baselines, the differences in results could likely be noise.
- There are also inconsistent numbers between Tables 1 and 2.The full model row in Table 1 reports RMSE/MAE/MAPE of 13.40/10.67/10.73 for working memory and 1.08/0.85/2.97 for general cognitive status, while the full model row in Table 2 reports 13.38/10.66/10.77 and 1.06/0.81/2.85, respectively. These should be the same model. The discrepancy suggests different runs or seeds, and it is not explained.
- The experiments are limited to the HCP S1200 dataset with the Neurograph preprocessing pipeline and the Schaefer parcellation. The generalizability claim would hold a lot stronger where this study also examined on additional datasets, especially since healthy young adults from HCP are far from clinical populations with cognitive impairment.
- The contrastive learning module (Section 2.3) is unclear. What clinical metadata variables are included? Age, sex, education, handedness? How are they encoded? The projection heads for $u_i$ and $v_i$ are also not described architecturally nor is a dimension mentioned in the Implementation details. The temperature $\tau$ and weighting factor $\lambda$ in Eq. 8 are also not reported. While there is code, it’s not clear whether the arguments there are the final optimized hyperparameters vs. initialized parameters, hence why they should be reported for reproducibility and to evaluate whether the contrastive objective is well-calibrated.
- The cross-graph temporal attention fusion (Section 2.1, Eq. 2) is difficult to follow and potentially overcomplicated. For Eq. 2, why is this particular form of bilinear attention used over standard multi-head attention across temporal steps? There is no comparison to simpler temporal aggregation strategies (e.g., mean pooling, LSTM, standard transformer temporal attention) in the ablation. While removing the entire Fusion module is tested in Table 2, the ablation does not show whether this specific attention design outperforms simpler temporal fusion alternatives.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/en/REVIEWER-GUIDELINES.html
- For additional references if there’s space, the neuroscience motivation for using MD system priors in a dynamic framework could be further strengthened by citing recent work showing that MD system dynamics specifically predict cognitive control abilities on HCP data (Cai, W., Taghia, J., & Menon, V. “A multi-demand operating system underlying diverse cognitive tasks.” Nature Communications 15, 2185 (2024)).
- State the train/validation/test split ratio. It’s not clear if you kept Neurograph’s 70/20/10 with the fixed stratified split seed.
- Also state you applied 5-fold CV; it’s not clear from results Tables whether the standard deviations here are applicable to CV-averaged results or multiple runs of CV folds.
- 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 — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses a meaningful problem of predicting cognitive scores from dynamic brain graphs and makes a reasonable contribution by integrating neuroscience priors into a dynamic graph framework. The ablation study supports the value of each component, and the interpretability analysis shows task-differentiated spatial patterns that are consistent with known functional neuroanatomy. However, the quantitative improvements over the Neurograph benchmark are marginal and not statistically validated, the evaluation is limited to a single dataset of healthy subjects, the contrastive learning module and temporal attention mechanism are not well-described, and there are some unexplained numerical discrepancies between Tables 1 and 2.The paper does well with combining domain knowledge and interpretability analysis, but there is a thin margin of improvement and a narrow evaluation scope that led to my rating.
- 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 sincerely thank the reviewers and meta-reviewer for their constructive comments. We appreciate the recognition of BrainWeaver’s clinically relevant problem formulation, neuroscience-grounded design, comprehensive baselines, ablation study, and interpretability analyses. We will revise the camera-ready version to improve clarity, reproducibility, and the positioning of our contributions.
Regarding the concern about clinical metadata and potential information leakage, we clarify that cognition-related metadata have been deliberately excluded from the contrastive learning module. The metadata used in our framework does not include the target cognitive scores, other cognitive assessment scores, or variables directly derived from the prediction labels. It is used only for subject-level contrastive regularization rather than providing label-related information to the prediction head.
We agree with the reviewers that statistical validation is important. In the camera-ready version, we will make our wording more precise and cautious, supplement missing implementation details to improve reproducibility, and correct minor textual issues, notation inconsistencies, and numerical discrepancies.
Finally, we will better position BrainWeaver as an integrated neuroscience-guided dynamic graph framework rather than claiming novelty for each individual component. We will also discuss the task-dependent contribution of temporal fusion and the limitation that the current evaluation is based on healthy HCP subjects, with validation on clinical cohorts left as an important direction for future work.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. 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.
This paper presents a well-executed and clinically relevant framework for cognitive score prediction using dynamic fMRI, with strong experimental coverage and meaningful interpretability analysis. However, reviewers note that performance gains are modest and lack statistical significance, and the overall novelty is somewhat incremental. Important implementation details and the use of clinical metadata require clarification to ensure reproducibility and avoid potential concerns such as information leakage. With these revisions and clearer positioning of contributions, the paper meets the bar for acceptance.
