List of Papers Browse by Subject Areas Author List
Abstract
Precise parcellation of functional networks (FNs) of early developing human brain is the fundamental basis for identifying biomarker of developmental disorders and understanding functional development. Resting-state fMRI (rs-fMRI) enables in vivo exploration of functional changes, but adult FN parcellations cannot be directly applied to the neonates due to incomplete network maturation. No standardized neonatal functional atlas is currently available. To solve this fundamental issue, we propose TReND, a novel and fully automated self-supervised transformer-autoencoder framework that integrates regularized non-negative matrix factorization (RNMF) to unveil the FNs in neonates. TReND effectively disentangles spatiotemporal features in voxel-wise rs-fMRI data. The framework integrates confidence-adaptive masks into transformer self-attention layers to mitigate noise influence. A self-supervised decoder acts as a regulator to refine the encoder’s latent embeddings, which serve as reliable temporal features. For spatial coherence, we incorporate brain surface-based geodesic distances as spatial encodings along with functional connectivity from temporal features. The TReND clustering approach processes these features under sparsity and smoothness constraints, producing robust and biologically plausible parcellations. We extensively validated our TReND framework on three different rs-fMRI datasets: simulated, dHCP and HCP-YA against comparable traditional feature extraction and clustering techniques. Our results demonstrated the superiority of the TReND framework in the delineation of neonate FNs with significantly better spatial contiguity and functional homogeneity. Collectively, we established TReND, a novel and robust frame-work, for neonatal FN delineation. TReND-derived neonatal FNs could serve as a neonatal functional atlas for perinatal populations in health and disease.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/4481_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/haohuanglab/TReND-MICCAI
Link to the Dataset(s)
The developing human connectome project (dHCP) dataset (second data release): https://biomedia.github.io/dHCP-release-notes
HCP Young Adult (HCP-YA) dataset: https://www.humanconnectome.org/study/hcp-young-adult/overview
BibTex
@InProceedings{MohSov_TReND_MICCAI2025,
author = { Mohapatra, Sovesh and Ouyang, Minhui and Tan, Shufang and Guo, Jianlin and Sun, Lianglong and He, Yong and Huang, Hao},
title = { { TReND: Transformer derived features and Regularized NMF for neonatal functional network Delineation } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15971},
month = {September},
}
Reviews
Review #1
- Please describe the contribution of the paper
The manuscript presents a framework for parcellating the brain into functional networks based on resting-state fMRI data. The framework is based on a transformer model encoding original voxelwise time series to a low dimensional embedding and then performing a R-NMF to map the embedding further to the NMF space. Finally a k-means clustering groups those signals to form functional networks. The framework was validated on both synthetic data and the dHCP dataset.
- 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 tackles a challenging problem of neonatal functional parcellation, an understudied yet important neuroscience problem.
-
The proposed method shows good reproducibility on the dHCP data and resembles the Yeo parcellation on HCP-YA data, suggesting the feasibility of the proposal.
-
- 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.
-
Clarity of the formulation needs to be improved. Fig. 1 seems to suggest that it produces a personalized parcellation, but the experiments seem to suggest a population-level parcellation. The construction of matrix V is described unclearly in the text and seems to be different from what Fig. 1 describes.
-
I failed to understand how the proposal fits into the literature. The design choice seems to be generic across all age span, not specifically tailored for infants, which means it should be compared to SOTA brain parcellation methods. The current comparison focuses on the most traditional methods, PCA/ICA/NMF.
-
It is also unclear what is considered the main novelty. I’d assume it is the new attention mechanism for embedding, but there are many existing works for brain time series encoding and should be compared to the proposed model. The introduction mainly focuses on the neuroscience side but did not specify what is the current SOTA fMRI embedding approach.
-
The only experiments with ground-truth validation are based on synthetic data, which was not clearly documented. What was the tool used? What was the parameters for simulation? I’d assume all parcellation methods should operate on the same spatial domain, but why nearly all methods in Fig. 2 result in networks defined outside the “brain”?
Minor: it is not clear why V can be considered non-negative.
-
- 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 provide sufficient information for reproducibility.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The main experiments that have ground-truth evaluation are the synthetic experiments, which are not clearly reported. The literature review is inadequate, so it is hard to judge the merit of the proposal.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This study proposes a novel method that integrates functional connectivity with spatial information. It begins by constructing a comprehensive data matrix combining functional connectivity matrices and brain surface geodesic distances, then employs regularized non-negative matrix factorization initialized via non-negative double singular value decomposition for decomposition. Finally, it transforms the continuous feature representations into discrete functional brain parcellations through cluster analysis, achieving stable and biologically plausible modeling of neonatal brain functional partitioning.
- 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.
- Multimodal Integration:Uniquely combines functional connectivity (temporal dynamics) with geodesic distances (spatial constraints) in a single framework, capturing both functional and anatomical relationships.
- Robust Decomposition:Uses NNDSVDa initialization for NMF to avoid unstable random starts, ensuring reproducible and biologically interpretable components.
- Hierarchical Parsing:First extracts continuous functional features (RNMF), then refines them into discrete parcels (KMeans), balancing granularity and practicality.
- Neonatal-Specific Design:Addresses unique challenges of infant brain data (low SNR, rapid development) through spatial regularization and stability-focused initialization.
- Computational Efficiency:Maintains scalability by decoupling heavy matrix decomposition (RNMF) from lightweight clustering (KMeans), suitable for large-scale neuroimaging datasets.
- 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.
- Sensitivity to Predefined Number of Regions:The K-means clustering requires a predefined number of brain parcels (K), but there is no gold standard for the optimal K in neonatal functional parcellation. Incorrect K may lead to over-segmentation or under-segmentation.
- Fixed Trade-off Between Spatial and Functional Information:The fusion weights between functional connectivity (FC) and geodesic distance are usually manually set, which may not adapt to individual differences or developmental stages.
- Limitations of NMF’s Linear Assumption:Non-negative matrix factorization (NMF) assumes linear combinations, which may fail to capture nonlinear interactions in brain functional networks.
- Dependence on Initialization:Although NNDSVDa improves stability, results may still vary slightly across different runs due to initialization sensitivity.
- Ambiguity in Cluster Boundaries:K-means enforces hard clustering, whereas real brain regions may exhibit functional gradients or overlapping boundaries, leading to unnatural parcellations.
- 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.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The introductory part of the article is concise, the methodology part is clearly described, and the model structure is correlated with the model diagram. The model design has its own ingenuity, especially the attention module, which is well worth learning. This is an article that makes you applaud without looking at the experimental data.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
This is a solid piece of work. Although some parts are not clearly explained, the level of innovation is sufficient, and there are no obvious issues in terms of methodological design.
Review #3
- Please describe the contribution of the paper
This paper describes a new framework for delineating functional brain networks, which is applied to neonatal fMRI data. The framework uses a transformer-based autoencoder to extract features from functional timecourses, which are then decomposed into clusters using a regularised NMF. The resultant functional parcellation of the neonatal brain is robust and exhibits distinct differences from the adult parcellation in networks that are not fully developed during this period, such as the default mode network.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
This paper presents novel methods for delineating functional networks; as far as I’m aware this is the first time that transformer-derived features have been combined with NMF for this task. The main methodological contribution comes from the confidence-adaptive mask applied during the self-attention operation, which down-weights unreliable tokens.
Evaluation in simulated data helps to show advantage of each component of the framework over other methods. The authors also demonstrate a thorough evaluation of different aspects of the results in real datasets. The results on the neonatal data indicate that this method could provide robust functional atlases of the developing brain.
The manuscript is very well written and presented.
- 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 methodological details are missing, particularly the training procedure and selection of hyperparameters. I also found it unclear how the spatial encodings are obtained and at what point are they incorporated into the functional information.
The authors state that “No standardized neonatal functional atlas is currently available.”, however there seems to be at least one example already in the literature: https://onlinelibrary.wiley.com/doi/10.1002/hbm.26718,
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
Since the confidence module is a key innovation of the proposed framework, it would be interesting in future to compare results with and without this component to see how influential it is on the parcellation. It would also be good to validate the framework in other neonatal datasets to test the generalisability of the neonatal parcellation specifically.
- 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?
This paper presents a novel framework that generates reliable and plausible functional parcellations for the neonatal brain. The paper is very well written and the results are well evaluated.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I am happy with the paper to be accepted at MICCAI, due to the strengths outlined in my original review.
Author Feedback
We thank the reviewers for their constructive suggestions. We summarized all the main critiques into 3 categories and provided point-to-point responses, with Rx-y referring to our response to comment #y from reviewer #x.
Category1: Novelty and method comparison
R2-2: We acknowledge that our method design can be applied to all age spans. We consider such method design as a strength for rigor, as we could validate our method with well-established adult brain parcellation to demonstrate the method’s robustness (Fig.5). We then applied the validated method to neonates which are much less studied for generating a high-fidelity functional neonate atlas. The core novelty is developing, to our knowledge, the first temporal BOLD fMRI signal feature extraction using a transformer-based autoencoder for brain parcellation. DL-based embeddings have so far only been used in “supervised” fMRI-based predictions, but not in any “unsupervised” SOTA brain parcellation yet. That is why we could only compare our temporal feature extraction method against PCA, UMAP, and TD, and clustering against k-PCA, ICA, and standard NMF (Fig.2).
R2-3: We will revise the introduction to clarify our main novelties including:1) first self-supervised DL-based time course feature extraction with a novel attention mechanism tailored for brain parcellation, 2) improved regularized NMF framework, and 3) novel fusion of geodesic spatial constraints with average functional connectivity. We will also expand the introduction of current SOTA fMRI embedding approaches, particularly those used in brain parcellation and prediction tasks, to better position our method within the existing literature.
Category2: Evaluation and robustness
R1-1: To address the sensitivity to predefined number of parcels (K), we tested K values from 2 to 20 (Fig.3A) and selected K=7 & 19 based on the lowest instability scores. R1-2: Since our method was designed to create a population-level parcellation for a single developmental stage (neonates), rather than individual-level parcellation, the fusion weights are unlikely to significantly affect the results. R1-3: We acknowledge the limitation and plan to explore kernelized RNMF in future work to better capture nonlinear interactions in brain networks. R1-4: To address initialization sensitivity, we fixed the random seed across runs during RNMF initialization, ensuring consistent and reproducible parcellations. R1-5: We used silhouette values-based confidence maps (Figs.3C&4B) to reveal boundary ambiguity.
Category3: Clarity and implementation details
R2-1: We’ll revise Fig.1’s caption to clarify that the panels are illustrative and the study produces a population‐level parcellation. We’ll also clarify that the V matrix, constructed by taking the absolute average functional connectivity across subjects, is non-negative by design.
R2-3: We used the SimTB toolbox (Ref15) to generate simulated fMRI data for parcellation validation. Specifically, we simulated 200 subjects with 1600 2D images (100×100 voxels) and 15 distinct functional networks. Rician noise was added to the images with contrast-to-noise ratios ranging from 0.1 to 0.3. All methods operated on the same spatial domain. Our method generated networks closer to the ground truth, while other methods produced noise-affected networks with spurious voxels appearing outside the ground-truth networks. We will update this in the manuscript.
R3-1: We trained the DL model for 1500 epochs on 4 A100 (80 GB) GPUs using a batch size of 1, Adam optimizer, and lr-scheduler starting at 1e-4 to extract 512 features. After computing the group‐averaged FC matrix across all subjects, we fused the spatial geodesic distance with the temporal FC-matrix.
R3-2: Our statement refers to absence of a coarse‐scale neonatal parcellation comparable to the widely used Yeo atlas with 7 or 17 networks in adults (Ref3). We acknowledge the fine-grained neonatal parcellation and will cite it in the revised manuscript.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A