Abstract

Brain network analysis based on functional Magnetic Resonance Imaging (fMRI) is pivotal for diagnosing brain disorders. Existing approaches typically rely on predefined functional sub-networks to construct sub-network associations. However, we identified many cross-network interaction patterns with high Pearson correlations that this strict, prior-based organization fails to capture. To overcome this limitation, we propose the Brain Hierarchical Organization Learning (BrainHO) to learn inherently hierarchical brain network dependencies based on their intrinsic features rather than predefined sub-network labels. Specifically, we design a hierarchical attention mechanism that allows the model to aggregate nodes into a hierarchical organization, effectively capturing intricate connectivity patterns at the subgraph level. To ensure diverse, complementary, and stable organizations, we incorporate an orthogonality constraint loss, alongside a hierarchical consistency constraint strategy, to refine node-level features using high-level graph semantics. Extensive experiments on the publicly available ABIDE and REST-meta-MDD datasets demonstrate that BrainHO not only achieves state-of-the-art classification performance but also uncovers interpretable, clinically significant biomarkers by precisely localizing disease-related sub-networks. Our code is available at https://github.com/Jingfeng-Tang/BrainHO.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2916_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Jingfeng-Tang/BrainHO

Link to the Dataset(s)

https://fcon_1000.projects.nitrc.org/indi/abide/ http://rfmri.org/REST-meta-MDD

BibTex

@InProceedings{TanJin_Learning_MICCAI2026,
        author = { Tang, Jingfeng AND Cao, Peng AND Wen, Guangqi AND Yang, Jinzhu AND Liu, Xiaoli AND Zaiane, Osmar R.},
        title = { { Learning the Hierarchical Organization in Brain Network for Brain Disorder Diagnosis } },
        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 BrainHO (Brain Hierarchical Organization Learning), a framework for brain disorder diagnosis from resting-state fMRI data. The central claim is that existing methods for brain network analysis rely on predefined functional atlases (e. g. , Yeo7, Smith10) to partition brain regions into fixed sub-networks and then model intra- and inter-subnetwork interactions independently. The authors argue that this rigid parcellation misses important cross-network interactions, for instance, regions from different predefined communities (such as VAN and SMN) may exhibit high functional connectivity that is clinically relevant but architecturally invisible to atlas-bound models. To overcome this, BrainHO introduces a hierarchical attention mechanism that operates at three levels: node-to-node (standard self-attention over brain regions), node-to-subgraph (learnable subgraph tokens aggregate information from nodes via Sparsemax attention), and subgraph-to-graph (a global graph token aggregates subgraph-level representations via Softmax attention). The key design choices are: (a) learnable subgraph tokens that discover sub-networks from data rather than using fixed atlas labels; (b) Sparsemax activation to enforce sparse, interpretable attention; (c) an orthogonality constraint loss that encourages diverse, non-redundant subgraph representations; and (d) a hierarchical consistency constraint that distills global graph-level semantics back down to refine local node-level representations via KL divergence, alongside an auxiliary classification loss on node features. The method is evaluated on two public datasets: ABIDE (N=1009, ASD vs. HC) and REST-meta-MDD (N=2380, MDD vs. HC), both using the Craddock 200 atlas. The authors report state-of-the-art accuracy of 69.68% on ABIDE and 64.71% on REST-meta-MDD compared to a set of recent Transformer-based and GNN-based baselines (BNT, Com-BrainTF, ALTER, DHGFormer, LHDFormer). An ablation study validates the contribution of each component. The paper also provides interpretability analysis showing that learned sub-networks partially align with known functional networks and additionally capture disease-relevant cross-network patterns.

  • 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.
    • Well-motivated problem formulation. The paper clearly articulates a genuine limitation of atlas-based approaches: predefined sub-network boundaries are rigid and cannot accommodate the well-documented phenomenon of cross-network functional interactions that are disease-relevant. The motivating example, that Frontal_Inf_Oper_L and Temporal_Sup_L are hard-coded into separate Yeo7 communities despite known ASD-related connectivity between them, is concrete and well-supported by neuroscience literature. This makes the transition from fixed to learnable sub-networks a natural and compelling direction.
    • Principled hierarchical design. The three-level hierarchy (node → subgraph → graph) is architecturally clean and mirrors the known modular-hierarchical organization of the brain. The use of Sparsemax at the node-to-subgraph level is a thoughtful choice: it enforces sparsity in the attention distribution, which both aids interpretability (each subgraph attends to a limited subset of nodes) and provides a natural mechanism for discovering discrete-like community structure. This is more principled than simply applying Softmax, which would produce diffuse attention weights.
    • Complementary regularization strategy. The combination of orthogonality constraint (encouraging diversity among subgraph tokens) and hierarchical consistency (distilling graph-level semantics back to node features) is well-designed. The orthogonality loss formulated as a cross-entropy over the cosine similarity matrix (Eq. 4) is elegant and differentiable. The teacher-student setup for hierarchical consistency (Eq. 5) with temperature-scaled KL divergence is a standard but effective mechanism for ensuring that local node features are informed by global context. The ablation study convincingly shows that each component contributes to performance.
    • Interpretability analysis. The paper goes beyond classification accuracy to provide meaningful neuroscientific interpretation. Figure 3 shows that some learned sub-networks align with predefined functional networks (e. g. , No. 7 for MDD overlapping 66.7% with SMN), while others span multiple predefined networks, capturing the very cross-network interactions the method was designed to find. The disease-related sub-networks identify regions consistent with prior literature (e. g. , fusiform gyrus and amygdala for ASD, hippocampus and anterior cingulate for MDD), which strengthens the clinical relevance claims.
    • The method is tested on two distinct disorders (ASD and MDD) with different datasets, demonstrating generality. On both datasets, BrainHO achieves the best accuracy among the compared baselines, and notably does so using only the static PCC matrix while several competitors also require raw BOLD signals as additional input.
  • 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.
    • Narrow and potentially non-representative baseline selection. This is the most significant concern. Table 1 compares BrainHO exclusively against a narrow set of recent Transformer/GNN-based methods (BNT, Com-BrainTF, ALTER, DHGFormer, LHDFormer), all from roughly the same methodological lineage. However, the broader literature on ABIDE classification includes numerous methods that report substantially higher accuracy. For example, Subah et al. (2021) reported 88% accuracy using DNNs [Subah et al. A deep learning approach to predict autism spectrum disorder using multisite resting-state fmri. Applied Sciences (2021)]; Chen et al. (2024, DeepASD) reported 87–88% accuracy using adversarial-regularized graph learning with multimodal data [Chen et al. Deepasd: a deep adversarial-regularized graph learning method for asd diagnosis with multimodal data. Translational Psychiatry (2024)]; Saponaro et al. (2024) achieved 85% using a joint fusion approach [Saponaro et al. Deep learning based joint fusion approach to exploit anatomical and functional brain information in autism spectrum disorders. Brain Informatics (2024)]; Epalle et al. (2021) reported 78.1% with hinge-loss-trained deep architectures [Epalle, Thomas Martial, et al. “Multi-atlas classification of autism spectrum disorder with hinge loss trained deep architectures: ABIDE I results. “ Applied soft computing 107 (2021): 107375]; and even simpler methods such as DFS+GCN (Shao et al. , 2021) achieved 79.5% on ABIDE with the C-PAC pipeline [Shao, L. , Fu, C. , You, Y. , & Fu, D. (2021). Classification of ASD based on fMRI data with deep learning. Cognitive Neurodynamics, 15(6), 961]. Against this broader landscape, BrainHO’s 69.68% accuracy appears modest rather than state-of-the-art. At minimum, the paper should acknowledge these higher-performing methods and explain why they were excluded from the comparison or why the comparison set is restricted to the Transformer/GNN family. The same concern likely applies to REST-meta-MDD, where recent work such as Liu et al. (2024) reported 77.6% accuracy using GCN on this dataset [Liu S. , & Gui, R. (2024). Fusing multi-scale fMRI features using a brain-inspired multi-channel graph neural network for major depressive disorder diagnosis. Biomedical Signal Processing and Control, 90, 105837], and a recent LGMF-GNN approach (2024) achieved 78.75% on a comparable MDD cohort [Liu S. , et al. (2024). An objective quantitative diagnosis of depression using a local-to-global multimodal fusion graph neural network. Patterns, 5(12)]. The paper’s claim of state-of-the-art performance is therefore overstated relative to the full literature.
    • Potential data leakage and insufficient methodological detail on evaluation. The paper uses 5-fold stratified cross-validation with 25% of the training set reserved for validation (early stopping). However, several aspects of the experimental setup raise concerns about data leakage: (a) Site effects in multi-site data. ABIDE is collected across 17 sites with heterogeneous scanner parameters, protocols, and demographics. REST-meta-MDD similarly aggregates data from 25 sites. If subjects from the same site appear in both training and test folds, the model may learn site-specific confounds rather than disease-related features. The paper does not describe whether site-aware splitting was performed, nor whether any harmonization was applied. This is a critical omission for multi-site neuroimaging studies, where site correction done outside the cross-validation loop can itself constitute leakage. (b) Mixup augmentation. The paper states that Mixup augmentation was adopted following prior works. Mixup is applied during training and generates synthetic examples by interpolating between pairs. If not carefully implemented within folds, interpolation between training and validation samples could introduce subtle leakage. The paper provides no details on this. (c) Hyperparameter selection. The paper reports specific values for α=1.3, K=8, d=384, τ=2.0, and a sigmoid-scheduled β, but does not describe how these were selected. If hyperparameters were tuned by observing test-fold performance (even indirectly through iterative experimentation), this constitutes another form of leakage.
    • Absence of nested cross-validation. Related to the previous comment, the standard recommendation for neuroimaging classification studies, especially those with complex models and multiple hyperparameters, is nested cross-validation, where an inner loop handles hyperparameter optimization and an outer loop provides unbiased performance estimates (Varoquaux et al. , 2017, NeuroImage). BrainHO has numerous hyperparameters (α, β, K, d, τ, learning rate, weight decay, batch size, scheduler parameters, number of layers, number of heads) yet uses flat 5-fold CV. As shown in the methodological literature, flat CV can produce positively biased accuracy estimates when hyperparameters are tuned. Given that the accuracy margins over baselines are relatively small (1–2%), such bias could account for the observed improvements. A nested CV protocol, or at minimum a sensitivity analysis over hyperparameters, is needed to establish that the reported gains are robust.
    • Limited statistical analysis of results. While the paper reports mean ± standard deviation across 5 folds, it does not perform statistical significance testing (e. g. , paired t-tests or Wilcoxon signed-rank tests) against baselines. Given the overlapping standard deviations in Table 1 (e. g. , BrainHO 69.68±2.11 vs. Com-BrainTF 67.69±3.09 on ABIDE), it is unclear whether the improvements are statistically significant. For MDD, the margins are even tighter (BrainHO 64.71±2.01 vs. ALTER/LHDFormer 63.95). Without significance tests, the claim of outperformance is not rigorously supported.
    • Scalability and sensitivity to K not explored. The number of subgraph tokens K=8 is fixed throughout experiments. However, the optimal number of latent sub-networks likely varies across disorders and datasets. The paper provides no sensitivity analysis for K, which is arguably the most important architectural hyperparameter. How does performance degrade for K=4, 6, 10, 12? Does the interpretability of learned sub-networks change qualitatively? Similarly, the paper does not discuss computational complexity or runtime compared to baselines.
    • Minor issues. (a) The term “intrinsicly” on page 2 is a typo (should be “intrinsically”). (b) The interpretability analysis in Section 3.2 is based on mapping CC200 regions to AAL116, but this mapping itself introduces approximation errors that are not discussed.
  • 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

    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.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    My recommendation balances a well-designed methodological contribution against significant concerns about the empirical claims and evaluation rigor. The following factors drove my score. Factors supporting the paper. The core architectural idea is sound and coherently presented. The hierarchical attention mechanism (node → subgraph → graph) with learnable subgraph tokens, Sparsemax for sparse attention, an orthogonality constraint for subgraph diversity, and a hierarchical consistency strategy via KL distillation form a principled composition. The motivation is concrete and well-grounded in neuroscience: the rigidity of predefined functional atlases (Yeo7, Smith10) is a genuine limitation, and the illustrative example of Frontal_Inf_Oper_L and Temporal_Sup_L being hard-coded into separate communities despite known ASD-relevant connectivity is compelling. The interpretability analysis (Fig. 3) is the paper’s strongest contribution: it shows that some learned sub-networks align with predefined functional networks while others span multiple networks, identifying disease-relevant regions (fusiform, amygdala for ASD; hippocampus, cingulum for MDD) consistent with prior literature. The ablation study is thorough and convincingly validates each component. Factors against acceptance. Several concerns collectively push the paper below the acceptance threshold. First, the state-of-the-art claim is not supported when placed in the broader literature. The 69.68% accuracy on ABIDE falls substantially below numerous previously published results: DeepASD (Chen et al., 2024, ~87–88%), Subah et al. (2021, 88%), Saponaro et al. (2024, 85%), DFS+GCN (Shao et al., 2021, 79.5%), and Epalle et al. (2021, 78.1%). On REST-meta-MDD, the reported 64.71% is below Liu et al. (2024, 77.6% using temporal GCN) and LGMF-GNN (2024, 78.75% on comparable MDD cohorts). The paper compares only against a narrow set of recent Transformer/GNN baselines from the same methodological lineage without acknowledging or discussing this broader performance landscape. This requires explicit discussion rather than omission. Second, the evaluation protocol raises data leakage concerns that are critical for multi-site neuroimaging. ABIDE (17 sites) and REST-meta-MDD (25 sites) have well-documented site effects from scanner hardware, acquisition protocols, and demographic heterogeneity. The paper does not describe whether site-aware splitting was performed, whether any harmonization was applied, or how Mixup augmentation was constrained to avoid cross-fold interpolation. Recent work (Rosenblatt et al., 2024, Nature Communications) has directly demonstrated how such leakage inflates performance in connectome-based machine learning. For a MICCAI submission, explicit treatment of these issues should be standard. Third, the absence of nested cross-validation is a serious methodological gap. The model has many hyperparameters (α, β, K, d, τ, learning rate, weight decay, Mixup, scheduler parameters, architectural choices), yet uses flat 5-fold CV with reported fixed values. Varoquaux et al. (2017, NeuroImage) explicitly recommend nested CV for this scenario. Given that BrainHO’s margins over baselines are 1–2% and fall within overlapping standard deviations, this methodological choice directly undermines the central claim. Fourth, no statistical significance testing is provided. With standard deviations of ±2–3% and differences of 1–2% between BrainHO and the strongest baselines (e.g., Com-BrainTF 67.69±3.09 vs. BrainHO 69.68±2.11 on ABIDE; ALTER/LHDFormer 63.95 vs. BrainHO 64.71 on MDD), paired t-tests or Wilcoxon signed-rank tests are necessary to establish that the improvements are not attributable to random variation across folds. Fifth, the key architectural hyperparameter K (number of subgraph tokens) is fixed at 8 with no sensitivity analysis. Since the entire method rests on the premise of learnable sub-networks, the absence of analysis regarding how performance and interpretability vary with K is a notable gap. Summary. The paper has a clean methodological contribution and genuinely interesting interpretability results, but the empirical case for state-of-the-art performance is not rigorously established given the narrow baseline selection, the missing treatment of site effects and data leakage, the absence of nested cross-validation for a heavily hyperparameterized model, and the lack of statistical significance testing. A convincing rebuttal addressing the comparison to broader literature, providing site-stratified or leave-one-site-out results, and including statistical significance tests and a K sensitivity analysis could move my assessment to Weak Accept.

  • 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



Review #2

  • Please describe the contribution of the paper

    BrainHO proposes a hierarchical attention mechanism with learnable subgraph tokens to dynamically aggregate brain regions beyond fixed atlas boundaries, combined with an orthogonality constraint and a hierarchical consistency strategy, evaluated on ABIDE and REST-meta-MDD for disorder classification.BrainHO proposes a hierarchical attention mechanism with learnable subgraph tokens to dynamically aggregate brain regions beyond fixed atlas boundaries, combined with an orthogonality constraint and a hierarchical consistency strategy, evaluated on ABIDE and REST-meta-MDD for disorder classification.

  • 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 motivation is clinically grounded and well-illustrated: Fig. 1(a) concisely demonstrates that significant inter-subnetwork correlations exist that fixed atlases systematically discard, providing a clear empirical justification for the learnable aggregation design. The interpretability analysis in Fig. 3 is a genuine strength—showing that learned subgraphs partially recover known functional networks (e.g., SMN, DAN) while also spanning cross-network regions provides meaningful validation beyond classification accuracy alone.

  • 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.

    W1.Insufficient Novelty Positioning Against Prior Hierarchical Brain Network Literature. The paper frames fixed-atlas methods as the dominant paradigm to overcome, but several recent works already achieve data-driven or atlas-free hierarchical brain network learning without predefined subnetwork constraints—including SW-HGL (MICCAI 2024), and hierarchical GNN approaches in the brain network analysis literature (e.g., 10.1109/TNNLS.2023.3282961). DHGFormer, which is used as a baseline in Table 1, itself performs hierarchical learning with dynamic graph structures. The paper does not discuss how BrainHO’s learnable subgraph token design is methodologically distinct from these contemporaries, leaving the novelty claim inadequately substantiated. Furthermore, the core mechanism—using learnable latent tokens as queries in cross-attention to pool node features into higher-level representations—is structurally identical to the Perceiver architecture (Jaegle et al., ICML 2021) and closely related to DiffPool (Ying et al., NeurIPS 2018); neither is cited or discussed.

    W2.The Hierarchical Consistency Loss Lacks Theoretical Justification and Has a Structural Conflict.Lhc​ (Eq. 5) uses the graph-level token X^g​ as a teacher to supervise node-level features X^n​ via KL divergence. However, X^g is itself derived by hierarchically aggregating from X^n through the same shared encoder, meaning teacher and student are computed from the same parameter set within the same forward pass. This creates a circular dependency in the gradient signal: improving node features changes the teacher simultaneously, making the optimization target non-stationary in a way that is neither theoretically motivated nor empirically analyzed. The paper provides no justification for why global-to-local distillation in this self-referential setting should produce stable or meaningful supervision, and no ablation isolates the effect of distillation direction (graph→node vs. node→graph).

    W3.Interaction Between LOC​ and Lhc​ Is Unanalyzed. LOC pushes subgraph tokens toward mutual orthogonality in the representation space, while Lhc pulls node representations toward alignment with the graph token derived from those same subgraph tokens. These two objectives operate on overlapping parts of the representation space and may conflict: orthogonality encourages subgraph diversity, but consistency encourages node features to converge toward a single global summary. The paper neither analyzes this potential tension theoretically nor provides training curve or gradient norm diagnostics to confirm stable joint optimization. The sensitivity analysis for α and β is absent.

    W4.The “w/o HA” Ablation Is Inadequately Defined. Table 1 reports a “w/o HA” variant described only as replacing the hierarchical attention with “an attention module,” yielding 61.74% accuracy. This description is too vague to be scientifically meaningful—it is unclear whether the replacement is a standard self-attention over all nodes, a mean pooling, or another architecture. Without a precise definition, this ablation cannot confirm that the performance gap is attributable to the hierarchical structure specifically rather than to incidental architectural differences such as parameter count or depth.

  • 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

    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 core idea of replacing fixed atlas parcellation with learnable subgraph tokens is a valid and timely direction, and the interpretability results are genuinely interesting. However, the novelty positioning against existing hierarchical brain network methods is insufficient (W1), and the hierarchical consistency loss has an unresolved structural problem that the authors need to either justify theoretically or redesign (W2). W3 and W4 are addressable in rebuttal with additional analysis. The writing quality requires careful proofreading before resubmission.

  • 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

    This paper proposes the Brain Hierarchical Organization Learning (BrainHO) to learn inherently hierarchical brain network dependencies based on their intrinsic features rather than predefined sub-network labels.

  • 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 study introduces the Brain Hierarchical Organization Learning (BrainHO) framework, which aims to capture the inherently hierarchical dependencies within brain networks by leveraging their intrinsic features instead of relying on predefined sub-network labels. The proposed approach incorporates a hierarchical attention mechanism that enables the model to aggregate nodes into a structured hierarchical organization. The analysis reveals cross-network interaction patterns characterized by strong Pearson correlations, which are insufficiently represented by conventional rigid, prior-based organizational models. In contrast, some existing methods predominantly depend on predefined functional sub-networks to define relationships between sub-networks.

  • 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 major weaknesses of the paper: 1.There are numerous studies that have utilized the static PCC connectivity matrix (or brain functional networks) for classifying brain disorders. However, the paper compares only two such methods; including a comparison with a broader range of representative approaches would enhance the evaluation and more comprehensively demonstrate the efficacy of the proposed method. 2.I agree that cross-network interaction patterns with high Pearson correlation are significant. However, recent studies classifying brain functional networks—especially those employing graph neural networks based on static Pearson correlation coefficient (PCC) connectivity matrices—often do not depend on predefined functional subnetworks, although some methods do. To present a more comprehensive overview of the current research, the paper should clarify this distinction.

  • 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 analysis reveals cross-network interaction patterns characterized by strong Pearson correlations, which are insufficiently represented by conventional rigid, prior-based organizational models.

  • 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



Author Feedback

N/A




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.

    I recommend an accept for this paper. Reviewers highly praised the clinically motivated BrainHO framework and its strong interpretability in discovering cross-network biomarkers. Although one reviewer gave a borderline score (3) citing concerns over the breadth of baseline comparisons and cross-validation protocols, these do not undermine the core methodological novelty, solid experimental results, and strong clinical relevance of the work.



back to top