List of Papers Browse by Subject Areas Author List
Abstract
Dynamic functional connectivity captures time-varying brain states for better neuropsychiatric diagnosis and spatio-temporal interpretability, i.e., identifying when discriminative disease signatures emerge and where they reside in the connectivity topology. Reliable interpretability faces major challenges: diagnostic signals are often subtle and sparsely distributed across both time and topology, while nuisance fluctuations and non-diagnostic connectivities are pervasive. To address these issues, we propose BrainSTR, a spatio-temporal contrastive learning framework for interpretable dynamic brain network modeling. BrainSTR learns state-consistent phase boundaries via a data-driven Adaptive Phase Partition module, identifies diagnostically critical phases with attention, and extracts disease-related connectivity within each phase using an Incremental Graph Structure Generator regularized by binarization, temporal smoothness, and sparsity. Then, we introduce a spatio-temporal supervised contrastive learning approach that leverages diagnosis-relevant spatio-temporal patterns to refine inter-sample similarity and capture discriminative spatio-temporal features, constructing a well-structured semantic space for interpretable representations. Experiments on ASD, BD, and MDD validate the effectiveness of BrainSTR, and the discovered critical phases and subnetworks provide interpretable evidence. Our code: https://github.com/guoguiliang111/BrainSTR.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0931_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/guoguiliang111/BrainSTR
Link to the Dataset(s)
N/A
BibTex
@InProceedings{GuoGui_BrainSTR_MICCAI2026,
author = { Guo, Guiliang AND Wen, Guangqi AND Liu, Lingwen AND Song, Ruoxian AND Cao, Peng AND Yang, Jinzhu AND Wang, Fei AND Liu, Xiaoli AND Zaiane, Osmar R.},
title = { { BrainSTR: Spatio-Temporal Contrastive Learning for Interpretable Dynamic Brain Network Modeling } },
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 addresses a significant challenge in dynamic functional-connectivity-based brain disorder diagnosis: improving spatio-temporal interpretability by pinpointing when diagnostically relevant signals emerge and where they reside in the connectivity topology. The authors propose BrainSTR, a spatio-temporal contrastive learning framework. The method infers state-consistent phase boundaries via an Adaptive Phase Partition (APP) module. It then separates disease-related and irrelevant connectivities within each phase using an Incremental Graph Structure Generator, and refines subject representations through spatio-temporal supervised contrastive learning. Experiments on MDD, BD, and ASD datasets demonstrate consistent performance improvements, and the identified critical phases and subnetworks provide interpretable evidence consistent with prior neurobiological findings.
- 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 addresses a highly meaningful problem. Explicitly targeting spatio-temporal interpretability—identifying both critical time phases and disease-relevant topological patterns—is a valuable and less explored direction in dynamic functional connectivity. (2) The overall design is technically sound and logically flows toward the stated objective. The modules are not disconnected; rather, the APP identifies informative temporal segments, the incremental structure generator extracts disease-related connectivities, and the contrastive objective constructs a discriminative semantic space. (3) The paper reports strong results across three distinct diagnostic tasks (MDD, BD, and ASD). BrainSTR achieves the highest reported ACC and AUC metrics across all tasks in Table 1, such as a notable +2.9% / +6.5% (ACC/AUC) improvement on the MDD dataset over the strongest dynamic baseline. (3) The ablation experiments successfully justify the proposed architecture. Removing the APP, the incremental structure update ($\Delta S$), the contrastive objectives, or the structure regularization terms all lead to performance degradation. (4) The paper goes beyond pure classification metrics. The case-level and group-level visualizations of important phases and retained subnetworks (e. g. , highlighting DMN-related patterns) make the contribution much more compelling than a standard performance-driven dynamic graph paper.
- 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 contribution appears stronger at the level of framework integration than at the level of sharply novel technical principles. Individual components—such as adaptive temporal partitioning, graph structure masking, and contrastive learning—are closely related to existing ideas in the literature. (2) While the authors note that the Incremental Graph Structure Generator is designed to capture progressive topological changes, they do not empirically justify why this recursive formulation is superior to simpler, independent phase-wise structure learning strategies. Similarly, the APP module’s temporal modeling claims would be much more convincing if compared against simpler adaptive partitioning baselines. (3) The interpretability analysis is promising but predominantly qualitative. While the paper provides retained-connectivity ratios for representative subjects, the group-level findings lack rigorous statistical support. The paper would be significantly stronger with quantitative validation, such as the stability of discovered subnetworks across folds, reproducibility metrics, or statistical significance tests validating the differences between “important” and “non-important” phases. (4) The experimental section reports mean +- std for 5-fold cross-validation, but no statistical significance analysis (e. g. , p-values from t-tests) is provided. This is necessary to conclusively support the reported gains over strong baselines. (5) Broad claims regarding “robust cross-disorder generalization” should be phrased more cautiously, as the ASD experiments are conducted exclusively on the NYU site of the ABIDE dataset (74 ASD, 98 controls) to avoid multi-center variability. (6) The major performance claims on MDD and BD rely entirely on a private cohort of 246 controls, 151 MDD, and 126 BD subjects. Without public access to this dataset, the broader research community cannot independently verify the results or benchmark against them.
- 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
This is a well-motivated paper addressing an important challenge in dynamic brain network analysis. I particularly appreciate the dual focus on improving diagnostic performance and spatio-temporal interpretability. The framework is logically constructed to support these goals, and the empirical results across multiple tasks are encouraging. To further strengthen the manuscript, I encourage the authors to: (1) Publicly release the private MDD and BD datasets: If restricted by ethics constraints, please provide an anonymized, de-identified, or synthetic subset. Releasing the dataset is critical to ensure the community can fully reproduce your findings and build upon your work. (1) Clarify explicitly what constitutes the main methodological novelty relative to existing dynamic graph and contrastive learning approaches. (2) Provide statistical significance analysis (e. g. , p-values) for the main performance comparisons in Table 1 to robustly validate the improvements. (3) Include quantitative statistical validation for the interpretability claims, rather than relying solely on qualitative visual comparisons of the subnetworks. (4) Temper broader claims around robustness and generalization to better align with the current experimental scope (e. g. , single-site ASD testing).
- 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?
Overall, this paper represents a meaningful and technically coherent contribution. The problem setting is highly relevant, the framework aligns well with the objective of identifying diagnosis-relevant signals in time and topology, and the experimental results are consistently positive. While the technical novelty of the individual components is moderate, the overall integration is well-executed, and the interpretability analysis adds genuine value. I lean toward a weak accept, but I strongly believe the paper must be improved by releasing the underlying data for verification, clarifying its novelty, adding required statistical significance tests, and providing quantitative backing for its interpretability claims.
- 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
1.This paper proposes an interpretable spatiotemporal contrastive learning framework for dynamic brain network modeling, which extracts sparsely distributed diagnostic signals across spatial and temporal dimensions while filtering out irrelevant noise. 2.This paper jointly selects critical temporal phases and identifies disease-related connectivity patterns in the topology to construct a well-structured semantic space.
- 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.This paper addresses the issue that disease-related signals are sparsely distributed in both time and topology, whereas irrelevant fluctuations and connectivity are pervasive. 2.To improve the interpretability of dynamic functional connectivity (FC), this paper employs a data-driven approach to partition phase boundaries with consistent states and identifies compact, diagnostically critical phases. 3.An incremental graph generator is proposed, which integrates multiple constraints to achieve stable and interpretable connectivity selection within each phase.
- 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 introduction mentions that traditional contrastive learning methods are disturbed by irrelevant signals, which impairs semantic relevance. However, this discussion should be placed in the limitations section rather than in the context of proposing the solution, and should only be briefly addressed there. 2.The authors claim to learn natural boundaries of brain states through the model, yet they still use a sliding window approach for partitioning. As a result, the subsequent data-driven adaptation appears more like a secondary encoding of coarse-grained windows. Does this truly correspond to the switching points of real neural events? 3.The design logic of the proposed model is to first determine temporal boundaries and then select spatial topology within those boundaries. This raises a question: if the temporal phase selection is incorrect, how can the subsequent capture of connectivity patterns be corrected? 4.Adaptive phase partitioning essentially forces a unified timeline for all brain regions. However, brain network activity is not only synchronous but also asynchronous, generating signals at different times. How is this aspect accounted for? 5.How is the time-invariant component defined? The orthogonal approach mentioned can make two vectors uncorrelated, but this is a high-dimensional constraint, not a physiological constraint that ensures they truly represent stable brain states. 6.If the goal is merely to find abnormalities in connection density, such as in the default mode network (DMN), can the proposed method identify more specific patterns compared to ordinary change point detection? 7.Regarding experimental details, the data preprocessing is not clearly described—for example, whether any processing has been applied to the raw data.
- 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?
The problem of dynamic spatiotemporal interpretability addressed in this paper is meaningful. However, the aspects considered and the methods chosen are not entirely comprehensive. Nevertheless, the experimental validation on a large number of datasets sufficiently supports the conclusions drawn.
- 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 BrainSTR, a spatio-temporal contrastive learning framework for interpretable dynamic brain network modeling from rs-fMRI data. The method combines an Adaptive Phase Partition (APP) module for data-driven temporal segmentation, an incremental graph structure generator for selecting disease-relevant connectivity, and a contrastive learning objective to learn discriminative representations. Experiments on ASD, BD, and MDD datasets demonstrate improved classification performance and provide qualitative interpretability analysis.
- 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.Addresses an important problem of spatio-temporal interpretability in dynamic functional connectivity, going beyond standard classification-focused approaches. 2.Proposes a coherent framework that integrates temporal segmentation (APP), graph structure learning, and contrastive learning into a unified pipeline. 3.Introduces a data-driven temporal partitioning (APP) that moves beyond fixed sliding windows, aligning with recent trends in adaptive dFC modeling. 4.Explicitly models disease-relevant vs. irrelevant connectivity, providing a more interpretable and structured representation of brain networks. 5.Demonstrates consistent performance improvements across multiple datasets (MDD, BD, ASD), including strong baselines. 6.Provides both temporal and spatial interpretability, linking identified phases and subnetworks to known brain systems.
- 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.Insufficient positioning with respect to prior work. The introduction lacks a clear discussion of existing methods and their limitations, and the relationship between the proposed components and prior work is not well articulated. In particular, the APP module is closely related to adaptive temporal partitioning approaches (e.g., adaptive windowing in [5]), but this connection is not clearly discussed.
2.Unclear novelty of the APP module. Conceptually, APP performs data-driven temporal segmentation via change detection on learned representations, which is similar to existing adaptive segmentation approaches in dFC. The paper does not clearly explain what is fundamentally new beyond implementation details.
3.Design limitations of the incremental structure generator. The structure $S_t$ evolves with strong smoothness constraints and small fixed updates (e.g., $\alpha=0.0$), which may lead to near time-invariant masks and limit the ability to capture phase-specific connectivity. In addition, updates are based on a low-dimensional descriptor rather than the full connectivity $A_t$, potentially restricting expressiveness.
4.Heuristic and insufficiently justified contrastive formulation. The selection of important phases (e.g., threshold $\alpha_t >1/W$) is heuristic, and the contrastive objective introduces a non-standard formulation (e.g., subtracting similarity from $H^0$) without sufficient justification or ablation.
5.Limited experimental validation and analysis. While ablations are provided, they do not fully analyze key design choices (e.g., APP sensitivity, structure update formulation, contrastive components). In addition, experiments are restricted to a single ABIDE site (NYU), raising concerns about generalizability across sites.
- 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 paper presents a coherent framework for spatio-temporal modeling of dynamic brain networks and shows consistent performance improvements across multiple datasets, which supports its potential impact. However, the novelty and positioning with respect to prior work—particularly for the APP module—are not clearly established, and several design choices are insufficiently justified or analyzed. In addition, the experimental evaluation could be strengthened in terms of deeper ablation and generalizability. Overall, while the method is promising, these issues limit the strength of the contribution, leading to a weak accept recommendation.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We sincerely thank all reviewers and the AC for their constructive feedback and positive assessment. We will carefully revise the manuscript accordingly in the camera-ready version.
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 authors address the main critiques raised during the review process in the final version (1) the spatio-temporal interpretation of the detected changes in brain state (2) justification for why standard change-point detection fails at this task and positioning against related work (3) design and parameter selection choices (beyond heuristics) and the main methodological innovation (4) lack of significance testing