List of Papers Browse by Subject Areas Author List
Abstract
Masked autoencoders (MAEs) have recently shown promise for self-supervised representation learning of resting-state brain functional connectivity (FC). However, a fundamental question remains unresolved: how should FC matrices be tokenized to align with the intrinsic modular organization of large-scale brain networks? Existing approaches typically adopt region-centric or graph-based schemes that treat FC as structurally homogeneous elements and overlook the large-scale network brain organization. We introduce NERVE (Network-Aware Representations of Brain Functional Connectivity via Bilinear Tokenization), a self-supervised learning framework that redefines FC tokenization by partitioning FC matrices into patches of intra- and inter-network connectivity blocks. Unlike image-based MAE, where fixed-size patches share a common tokenizer, FC patches defined by network pairs are heterogeneous in size and correspond to distinct functional roles. To resolve this problem, NERVE embeds FC patches through a novel structured bilinear factorization. This formulation preserves network identity and reduces parameter complexity from quadratic to linear scaling in the number of networks. We evaluate NERVE across three large-scale developmental cohorts (ABCD, PNC, and CCNP) for behavior and psychopathology prediction. Compared to structurally agnostic MAE variants and graph-based self-supervised baselines, the proposed network-aware formulation yields more stable and transferable representations, particularly in cross-cohort evaluation. Ablation studies confirm that the proposed bilinear network embedding and anatomically grounded parcellation are critical for performance. These findings highlight the importance of incorporating domain-specific structural priors into self-supervised learning for functional connectomics. Code is available at: https://github.com/leomlck/NERVE
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3895_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/leomlck/NERVE
Link to the Dataset(s)
ABCD dataset: https://nda.nih.gov/abcd
PNC dataset: https://www.med.upenn.edu/bbl/philadelphianeurodevelopmentalcohort.html
CCNP dataset: https://www.scidb.cn/en/detail?dataSetId=c81f0e90a51b4cfca348ce4da6ca734e
BibTex
@InProceedings{MilLeo_NetworkAware_MICCAI2026,
author = { Milecki, Leo AND Hu, Qingyu AND Jafrasteh, Bahram AND Sabuncu, Mert R. AND Zhao, Qingyu},
title = { { Network-Aware Bilinear Tokenization for Brain Functional Connectivity Representation Learning } },
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 introduces NERVE, a self-supervised MAE framework for brain functional connectivity (FC) representation learning. The central contribution is a network-aware bilinear tokenization scheme that partitions FC matrices into intra- and inter-network connectivity blocks defined by established large-scale functional brain networks, and embeds these heterogeneous patches via a structured Khatri-Rao (column-wise Kronecker) product of network-specific learnable embeddings. This replaces either a shared linear projection (which ignores network identity) or patch-specific projections (which scale quadratically in parameters) with a factorized bilinear form that scales linearly in the number of networks while preserving network specificity. The framework is evaluated on behavioral and psychopathology prediction across three developmental cohorts (ABCD, PNC, CCNP).
- 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.
1.The core technical idea is well-motivated and genuinely novel in this domain. The framing of the tokenization problem is crisp and compelling: image patches have natural spatial locality, but FC patches defined by network pairs are heterogeneous in size and semantically distinct. The three-way comparison between shared linear, patch-specific linear, and bilinear tokenization is a clean design space analysis. The use of the Khatri-Rao product to factorize patch-specific projections into network-level embeddings is technically elegant — it achieves parameter sharing across network pairs while preserving network identity through bilinear interactions. The parameter complexity analysis (O(R²/N²n) → O(R)) clearly formalizes the efficiency gain. 2.The ablation design is rigorous and well-controlled. The ablation studies are among the strongest aspects of the paper. Table 2 cleanly isolates the tokenization strategy (shared vs. patch-specific vs. bilinear) within the same NERVE framework. Table 3 goes further by testing: (a) random region permutation as a null model, (b) fixed 16×16 square patches removing functional network grounding, and (c) coarser 7-network parcellation. The permutation baseline is particularly valuable — it directly tests whether network structure matters rather than just the tokenization machinery — and the result (NERVE with 17-network parcellation significantly outperforms permuted baselines across most targets) is a strong, principled validation. 3.Evaluation protocol is methodologically sound. Several good practices are followed: bootstrap 95% confidence intervals, paired bootstrap significance testing between top methods, age and sex deconfounding fit only on training folds, and KRR as downstream model (justified by prior literature). The use of CCNP as a held-out out-of-domain cohort with a different age range and preprocessing pipeline is a meaningful test of cross-cohort generalizability that goes beyond most comparable papers. 4.The biological motivation is grounded in established neuroscience. Anchoring tokenization to the Yeo 17-network parcellation is a principled choice backed by decades of resting-state fMRI literature. The insight that network pairs are the natural analog of spatial patches in images is conceptually sound and clearly articulated. 5.Computational efficiency. Training on a single NVIDIA L40 GPU in 4,000 epochs with batch size 1,024 is practical. The bilinear formulation’s linear parameter scaling is not just theoretically appealing — it is practically important for larger parcellations.
- 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.
1.Effect sizes are modest and confidence intervals are wide, particularly for OOD evaluation. While the paper correctly contextualizes performance within the known difficulty of behavioral prediction from FC (r < 0.15 is common), the confidence intervals in Table 1 are very wide — particularly for CCNP (e. g. , Ext. r=. 33±. 14). This means the performance differences between methods are often not statistically significant, and the headline result for CCNP rests on a single target with a very wide CI. With N=178 for CCNP, this is expected, but it limits the strength of conclusions drawn from OOD evaluation. The paper would benefit from explicitly reporting which pairwise comparisons are and are not significant across all targets, rather than only noting gray shading for the top-1 vs. top-2 comparison. 2.The bilinear tokenization idea, while novel in the FC context, has precedents in the broader ML literature that are not discussed. Bilinear factorizations and Kronecker-structured projections are well-known in multi-task learning, tensor decomposition, and vision-language models. For instance, bilinear pooling in visual question answering (e. g. , MLB, BLOCK) and Kronecker-factored linear layers in efficient transformers share structural similarities. The paper does not situate its bilinear tokenization within this broader literature, which weakens the novelty framing for a technically sophisticated reviewer. A brief discussion acknowledging these connections and clarifying what is specifically novel in the FC/neuroscience context would strengthen the paper. 3.The MAE reconstruction objective and masking strategy are entirely standard. Beyond the tokenization, the MAE framework is a straightforward application of He et al. (2022) — same masking ratio tuning, same encoder-decoder structure, same MSE reconstruction loss. This is not a criticism per se, since the paper’s claim is that tokenization is the key design variable, but it does mean the contribution is entirely concentrated in the tokenization scheme. If the tokenization advantage diminishes with larger datasets or different downstream tasks, there is no complementary architectural contribution to fall back on. 4.Comparison with BrainMass is potentially unfair. BrainMass is described as a “foundation-style model combining temporal masking and contrastive learning” — implying it operates on temporal BOLD signals rather than static FC matrices. The paper states it restricts comparisons to “approaches operating on FC matrices,” but then includes BrainMass which appears to operate differently. It is unclear whether BrainMass is being applied in its intended modality here or adapted to FC matrices, and if adapted, whether the adaptation is faithful to the original design. This needs clarification. 5.Only one parcellation resolution is tested for the bilinear scheme. The paper tests 7-network vs. 17-network parcellations, but only with the bilinear tokenization. It would be informative to know whether the bilinear advantage is specific to the 17-network granularity or holds across resolutions. The finding that “when network structure is disrupted, patch-specific linear tokenization tends to outperform bilinear” (Table 3) is interesting but raises a question: is the bilinear factorization only beneficial when the parcellation is meaningful? If so, the method’s performance may be more sensitive to parcellation quality than currently acknowledged. 6.No interpretability analysis is provided. The paper mentions future analysis of learned network-specific weights and attention patterns, but no such analysis is included in the current work. For a method whose central claim is that network-aware tokenization captures biologically meaningful structure, even a basic visualization of the learned network embeddings (e. g. , similarity structure of Ul matrices across networks) or attention weights between network-pair tokens would substantially strengthen the biological plausibility argument. 7.Preprocessing heterogeneity between cohorts is not fully addressed. ABCD FC matrices come from the CBIG release (pre-processed externally), while PNC and CCNP are processed using DeepPrep. The paper does not discuss whether this preprocessing difference could confound cross-cohort comparisons, particularly for the OOD CCNP evaluation.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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 recommendation of weak accept reflects that NERVE addresses a well-defined and previously underexplored problem (principled FC tokenization for MAE), proposes a technically sound and parameter-efficient solution, and supports its claims with rigorous ablations including permutation null models and OOD evaluation. The weaknesses — modest effect sizes, wide CIs especially in OOD, missing interpretability, and incomplete literature contextualization — are real but do not invalidate the contribution. This is a solid, focused paper that makes a clear and verifiable methodological point.
- 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
The paper introduces a self-supervised framework for learning network-aware functional connectivity representations from rs-fMRI. Its leverage bilinear tokenization strategy that encodes inter-network connectivity efficiently, preserving network specificity while reducing parameters. Experiments show that NERVE produces robust embeddings that achieve strong behavioral prediction in limited datasets.
- 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.
1.The authors effectively combine prior knowledge of functional networks with a self-supervised learning strategy 2.The authors detailed analysis of the experimental results.
- 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.
1.The technical novelty and contribution are limited, as the main innovation is a redesign of the tokenizer, while the rest of the method closely follows traditional MAE without introducing additional self-supervised training strategies. 2.The amount of data used for pretraining to extract embeddings is relatively small compared to other methods (e.g., BrainMass, BrainJEPA). I want to know if increasing the pretraining data could improve representation quality. 3.The authors argue that patch-specific linear projections are less practical due to higher computational cost, but it remains unclear whether, with larger datasets, a patch-specific linear approach could outperform the proposed bilinear method. 4.The experimental evaluation focuses only on behavioral prediction, which is limited. Including additional tasks (e.g., age prediction) could provide a better assessment of the robustness and stability of the learned embeddings.
- 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.
- 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?
The methodological novelty is limited. Since the results are obtained from pretraining on relatively small, specific datasets, the proposed method shows some advantage under limited data. However, it is unclear whether, with larger datasets, approaches such as patch-specific linear projections or other more detailed strategies could outperform it. Therefore, more extensive experiments are needed to fully demonstrate the method’s superiority.
- 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 #3
- Please describe the contribution of the paper
This paper proposes NERVE, a network-aware self-supervised framework for functional connectivity (FC) representation learning. It addresses the limitation of existing methods that ignore the modular structure of brain networks by introducing a new tokenization strategy based on intra- and inter-network patches. The authors further design a bilinear embedding mechanism that models interactions between networks while reducing parameter complexity. Experiments on multiple large-scale datasets show that the method achieves more stable and transferable representations compared to existing approaches, especially in cross-cohort settings.
- 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 paper presents a novel and well-designed framework for FC tokenization, addressing the limitations of conventional approaches that overlook the functional organization of brain connectivity.
The paper is well-structured, and the methodology is clearly described.
The extensive experimental results provide strong evidence supporting the effectiveness of the proposed method.
- 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.
Some descriptions in the paper are not entirely accurate. For example, in the Introduction, FC is described as having high dimensionality and low signal-to-noise ratio. In practice, the dimensionality of FC largely depends on the chosen parcellation scheme, and it is the fMRI signal itself rather than FC that typically suffers from low signal-to-noise ratio.
In the Method section, the authors claim to “learn structured and transferable representations,” but the mechanism by which transferability is achieved is not clearly demonstrated or justified.
The comparison with related work is incomplete. In particular, the authors overlook recent MAE-based approaches for fMRI that employ temporal masking strategies, such as TARDL: Task-Aware Reconstruction for Dynamic Representation Learning of fMRI.
In Table 1, the proposed method performs significantly worse than existing approaches on the CCNP dataset. The paper does not provide an explanation for this discrepancy.
The authors do not provide code, which limits the reproducibility of the work.
- 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.
- 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.
(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?
Based on the aforementioned strengths and weaknesses
- 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.
The paper addresses an important and timely problem in self-supervised functional connectivity representation learning. Although reviewers noted modest effect sizes, incomplete comparisons, and limited downstream task diversity, they agreed that the network-aware bilinear tokenization strategy is well motivated and technically sound. The method is clearly presented, supported by strong ablations and cross-cohort evaluation, and shows practical impact in behavioral and psychopathology prediction.