List of Papers Browse by Subject Areas Author List
Abstract
Functional parcellation of the cerebral cortex from resting-state functional magnetic resonance imaging (rs-fMRI) is a fundamental step for large-scale brain network analysis. Learning functional parcellation from rs-fMRI is essentially a dimensionality reduction of vertex-wise functional connectivity (FC), where previous studies replied linear dimensionality reduction techniques, failing to capture the nonlinearity nature of brain functional organization. In this work, we propose a self-supervised representation learning framework for whole-cortical, group-level parcellation to release the linear constraints in feature extraction. Specifically, we first decouple FC strength from spatial proximity via distance-based residualization, enabling the learning of functionally meaningful features beyond geodesic constraints. A variational autoencoder (VAE) is then employed to learn compact and uncertainty-aware FC representations, augmented with a neighborhood-based contrastive objective to explicitly promote local functional coherence. Spatial continuity is further encouraged through Laplacian positional encoding prior to clustering. Experiments demonstrate that the proposed method consistently achieves better performance than widely used cortical parcellations, achieving improvement in functional homogeneity, functional boundary alignment, clustering quality, and consistency with task-evoked activation patterns.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0399_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZhuJia_Representation_MICCAI2026,
author = { Zhu, Jianfei AND Wei, Baichun AND Liu, Shaohui AND Zhu, Haiqi AND Jiang, Feng AND Yi, Chunzhi},
title = { { Representation Learning with Distance-Residualized Functional Connectivity for Cortical Parcellation from rs-fMRI } },
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
This paper proposes a self-supervised representation learning framework for cortical parcellation using (rs-fMRI). The primary contribution lies in the introduction of neighborhood contrastive learning and Laplacian positional encoding, which achieves superior cortical segmentation results.
- 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.
Successfully utilized functional connectivity (FC) features to construct fine-grained cortical parcellations, demonstrating effectiveness in capturing inter-individual variability in functional boundaries.
- 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 directly concatenate the VAE latent variables with the Laplacian eigenvectors prior to K-means clustering. Given that these two features set possess entirely different numerical scales and physical meanings, and considering that K-means is highly sensitive to absolute distances, direct concatenation may cause the clustering results to be dominated by the features with the larger variance. Please clarify whether any re-normalization or weighting strategies were applied to the features before clustering.
- It is well known that the K-means algorithm is highly sensitive to initial cluster centers. Even with the k-means++ initialization strategy, the final clustering results are inevitably subject to random fluctuations. The current study lacks a systematic evaluation of the stability of the resulting parcellations. It is recommended that the authors include a consistency analysis across multiple runs.
- Table 1 reveals that DRFC does not achieve the best performance in almost any single metric. Its top ranking relies solely on the arithmetic mean of the standardized scores across the four metrics. Framing a “superior overall average” as “consistently producing superior parcellations” is overstated. The authors should soften their claims.
- The current study exclusively trains and tests the model on the HCP dataset, which features extremely high temporal resolution (TR=0.72s) and an excellent signal-to-noise ratio. Can this model generalize to resting-state fMRI data acquired from routine clinical 3T scanners, which typically have lower temporal resolution and higher noise levels? The lack of cross-dataset validation significantly limits the practical utility of this method as a universal parcellation template.
- In Table 2, removing the neighborhood contrastive learning module (w/o NC) causes the DCBC metric to plummet from 2.79 to 0.20.This indicates that relying solely on the VAE with a Gaussian decoder for feature reconstruction fails completely to capture functional boundary information.
- At the beginning of Section 2.1, it is stated that the maximum geodesic distance is truncated at 50 mm for computational efficiency. However, in the subsequent definition of the residualization formula, G_max is set to 45 mm. Please verify and correct such inconsistencies.
- 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
There are several typographical errors in the manuscript. For example, in the Abstract, “. . . previous studies replied linear dimensionality. . . “. In the Conclusion, “. . . fuctional boundary alignment. . . “.
- 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 primary reasons for this rating are the limited novelty of the methodology and the lack of demonstrated model robustness.
- 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
This paper presents an novel self-supervised representation learning framework for group-level cortical parcellation. Several well-conceived designs, including the decoupling of FC patterns from spatial proximity, the use of a variational autoencoder with a Gaussian decoder, and the introduction of Neighborhood Contrastive Learning, collectively enable the resulting parcellation to achieve higher functional homogeneity, better alignment of functional boundaries, and improved clustering quality.
- 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 work makes a valuable contribution to the field of nonlinear representation learning for resting-state functional connectivity. In my view, three key strengths underpin its success: 1) Decoupling functional connectivity patterns from spatial proximity enhances sensitivity to functional similarities between vertices without the confounding influence of spatial location, which facilitates the identification of functional boundaries during parcellation; 2) The use of a VAE framework with a Gaussian decoder improves robustness by effectively accounting for noise in functional connectivity data; 3) The design of Neighborhood Contrastive Learning further increases the discriminability of features, thereby benefiting subsequent clustering tasks. Together, these well-conceived strategies effectively address key challenges in group-level cortical parcellation based on functional connectivity, particularly regarding spatial proximity, noise sensitivity and stability.
- 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)When computing distances between samples using K-means, is weighting between features and positional encodings not necessary? Would directly calculating Euclidean distances on the concatenated features risk either positional information or feature information being neglected? 2)I agree with the authors’ approach of comparing each method’s results against a null model and calculating Z-scores to ensure fair evaluation. However, to more intuitively demonstrate parcellation performance, I suggest also presenting the raw metrics of the parcellations. 3)It is suggested that the authors provide more explanatory context for the ablation experiments, as they currently read more as a mere description of results. 4)It is recommended that the authors consider using GCN instead of MLP in future work, as GCN inherently account for graph structure and are more robust to noise in the data.
- 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?
The multiple well-conceived designs in this paper are not overly complex yet align well with the requirements of the task, making the overall approach both concise and effective.
- 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
The paper proposes a self-supervised framework for cortical parcellation that learns nonlinear representations from rs-fMRI instead of relying on linear methods. It removes spatial distance bias from functional connectivity data, learns compact features using a VAE with contrastive learning, and then reintroduces spatial information to produce better parcellations.
- 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 paper has a nice idea of removing spatial distance bias from functional connectivity before learning representations, which is important because FC is usually confounded by proximity. This makes the learned features more functionally meaningful.
2.The use of a Gaussian decoder combined with contrastive learning is good, since it captures nonlinear structure while also enforcing local functional consistency in the latent space.
3.The overall pipeline is good, especially the idea of decoupling spatial information first and then reintroducing it using Laplacian positional encoding is a clean approach.
4.The evaluation is comprehensive, with multiple metrics (functional homogeneity, DCBC, silhouette, task alignment) and comparisons against many baselines, which makes the improvements more convincing.
- 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 method is evaluated only at the group level using HCP data but lots of existing works [1] focus on individual-specific parcellations and generalization across subjects, which is a more clinically relevant setting . The authors do not test if the learned representations transfer to subject-level parcellation tasks.
2.The baseline comparisons are mostly against classical atlases like Schaefer and Glasser, but they miss more recent geometric deep learning approaches. It would be important to compare against GNN-based methods for fMRI (e.g., masked graph convolution models like [2]) to better show the technical advantage.
3.The contrastive setup depends on several hyperparameters (top 15% positives, bottom 25% negatives, fixed neighborhood radius), but there is no sensitivity analysis. So it’s difficult to judge how robust the method is to these hyperparameters.
[1] Kong et al. Individual-Specific Areal-Level Parcellations Improve Functional Connectivity Prediction of Behavior, Cerebral Cortex 2021.[2] Wang et al. Unrevealing Reliable Cortical Parcellation of Individual Brains Using Resting-State Functional Magnetic Resonance Imaging and Masked Graph Convolutions, Frontiers in Neuroscience, 2022
- 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.
(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 proposes a technically sound representation learning framework that shows consistent gains across multiple metrics. However, the evaluation is limited to group-level HCP data, lacks comparisons to modern GNN-based methods, and does not analyze sensitivity to key hyperparameters, which makes it harder to fully assess robustness of the framework.
- 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
We sincerely thank the reviewers and meta-reviewer for their constructive comments and for recognizing the value of our work. Since the paper has been provisionally/early accepted, we focused the revision on issues that could be addressed within the strict page limit, while regarding the remaining suggestions as important future directions. First, regarding the concerns from Reviewers 1 and 2 about concatenating latent representations with Laplacian positional encodings before K-means clustering, we clarified in Sec. 2.1 that each Laplacian feature dimension is z-score normalized across vertices before concatenation. This reduces the risk that clustering is dominated by positional dimensions due to scale differences. We agree that weighting between functional embeddings and positional encodings is important, and will explore more systematic weighting strategies in future work. Second, in response to Reviewer 1’s comment that the original claims were overstated, we softened the wording in the Abstract, comparison results, and Conclusion. The revised manuscript now states that DRFC achieves favorable average standardized scores and competitive trade-offs across complementary metrics, rather than claiming superiority on every individual metric. Third, we would like to clarify Reviewer 1’s concern about the apparent inconsistency between truncating geodesic distances at 50 mm and using Gmax = 45 mm in the residualization formula. These two values serve different purposes. The geodesic distance matrix is computed separately for each subject and then averaged to obtain the group-level distance matrix. We used a slightly larger computational cutoff of 50 mm to avoid potential truncation bias around the analysis threshold. For example, a vertex pair may have a distance of 44.5 mm in some subjects and 45.5 mm in others. If distances were truncated at exactly 45 mm during subject-level computation, values above 45 mm would be clipped and could bias the group-level average near the residualization threshold. Thus, 50 mm is used as a safer computational cutoff, whereas Gmax = 45 mm defines the neighborhood range used in residualization. We regret that this distinction was not sufficiently clear in the manuscript. Fourth, regarding Reviewer 1’s observation that removing neighborhood contrastive learning causes a large DCBC decrease, we agree with this interpretation and would like to clarify that this result is consistent with our method motivation. As stated in the manuscript, optimizing reconstruction alone primarily enforces similarity between FC patterns, without explicitly encouraging functional homogeneity among nearby vertices. The neighborhood-based contrastive loss was introduced to address this limitation. Therefore, the degradation in the w/o NC ablation is expected and supports the necessity of this component for capturing local functional organization and boundary-sensitive information. We also appreciate the suggestions on additional analyses, including multi-run K-means/parcellation stability, reporting raw metrics, more extensive ablation explanations, cross-dataset validation on lower-SNR clinical 3T data, individual-level transfer, comparison with recent GNN-based parcellation methods, and hyperparameter sensitivity analysis. Due to the strict page limit and the scope of the current group-level HCP study, we were not able to add all these experiments in the revised manuscript. We agree that they are valuable directions. Future work will investigate robustness across random initializations, generalization to routine clinical datasets, extension to individual-specific parcellation, and graph-based encoders such as GCNs. Finally, we corrected typographical and minor grammatical errors throughout the manuscript. We thank the reviewers again for their helpful feedback, which improved the clarity of our method description and the precision of our claims.
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.
All reviewers are generally positive and lean toward acceptance, noting the paper is technically sound, with strengths in spatial decoupling, contrastive learning, and strong empirical evaluation. However, reviewers highlight limited novelty, lack of robustness and sensitivity analysis, and evaluation restricted to HCP group-level data. Overall, it is considered effective but somewhat incremental, with concerns about generalization and validation rigor.