List of Papers Browse by Subject Areas Author List
Abstract
Resting-state functional MRI (rs-fMRI) provides crucial insights into dynamic brain network mechanisms underlying psychiatric disorders. While deep learning methods have advanced rs-fMRI-based disease recognition and dynamic brain network modeling, existing approaches are often limited by two key assumptions: (1) fixed subnetwork partitioning, where predefined functional templates or partition rules impose a fixed region-subnetwork mapping that, although interpretable, restricts the discovery of complex coordination patterns; and (2) synchronization assumption, where inter-subnetwork interactions are modeled as strictly time-aligned, overlooking physiologically plausible delays and weakening diagnostic signals.
To address these limitations, we propose BrainACU, an asynchronous coordination unit modeling framework for dynamic brain network analysis. BrainACU removes fixed partitioning by learning coordination units (CUs) directly from data, capturing both synchronous and asynchronous coordination. Using conventional sliding windows, we segment BOLD sequences into temporally consistent patches for subsequent CU learning and interaction modeling. BrainACU then learns stable ROI-to-CU mappings via data-driven clustering with cross-sample consensus, and projects variable-size CU signals into a unified latent space for dimensional alignment. Finally, an asynchronous interaction module models both time-aligned and time-lagged unit-level dependencies to form dynamic brain network representations. Experiments on ASD, MDD, and BD demonstrate that BrainACU outperforms existing methods and provides more discriminative and interpretable dynamic brain network evidence. Our code: https://github.com/guoguiliang111/BrainACU.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0936_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/BrainACU.
Link to the Dataset(s)
N/A
BibTex
@InProceedings{GuoGui_BrainACU_MICCAI2026,
author = { Guo, Guiliang AND Li, Lanting AND Liu, Qi 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 = { { BrainACU: Asynchronous Coordination Unit Modeling for Dynamic Brain Network Analysis } },
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 BrainACU, a framework that addresses two limitations in dynamic brain network analysis: fixed subnetwork partitioning and the assumption of strict temporal synchronization. Its key contributions include: (1) Data-driven CU Discovery: Learning Coordination Units directly from BOLD signals via consensus clustering; (2) Asynchronous Interaction Module: Using directed attention and lag-promoting regularization to capture time-shifted dependencies between brain units; and (3) Improved Performance and Interpretability: Achieving state-of-the-art results across three disorders and providing “lag-step statistics” to visualize disease-related temporal delays.
- 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.Relaxing the “zero-lag” synchrony assumption is physiologically well-motivated, as neural propagation involves inherent delays. 2.The consensus clustering approach for CU discovery allows the model to adapt to different disorders where canonical templates might be suboptimal. 3.The model was tested on multiple disorders (ASD, MDD, BD) and outperformed dynamic baselines like BrainDGT and ST-GCN.
- 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 proposed “Coordination Units” are essentially derived from k-means clustering over ROIs. This is conceptually similar to many existing parcellation or clustering-based approaches and does not fundamentally differ from prior functional grouping methods. 2.The authors fix the number of Coordination Units at K=3.While the paper argues that fixed templates restrict the discovery of complex patterns, the proposed data-driven CU discovery via consensus clustering essentially functions as a clustering-based parcellation. Setting K=3 is biologically oversimplified, as the human brain is typically organized into more granular networks (e. g. , Yeo 7 or 17 networks). The authors have not demonstrated how these learned CUs correspond to known functional systems. Additionally, without visualizing these ROI-to-CU mappings or providing a biological mapping, it is difficult to determine if the “asynchronous delays” learned between these units are neuroscientifically meaningful or merely data artifacts. 3.The framework maps variable-sized CU signals to a shared latent dimension d using CU-specific linear projections. However, if the number of ROIs within each CU varies significantly, projecting them into the same low-dimensional space (d=36) may lead to severe information bottlenecks for large CUs or representation bias for smaller ones. A more detailed justification or sensitivity analysis regarding this dimension alignment is required. 4.Regarding Equation 8, the L_lag penalty successfully forces the model to capture time-shifted dependencies. However, since the brain is partitioned into only three broad units, the granularity of the “dynamics” being modeled is quite coarse. If a single CU aggregates regions with distinct functions, the resulting time-lagged signals might represent an uninterpretable average of multiple competing processes. The authors should clarify what these units represent to justify the analysis of their temporal delays. 5.The evaluation on ABIDE is restricted to the NYU site. While this reduces multi-center variability, it limits the assessment of the model’s robustness and generalization. Please clarify if this selection was due to sample size constraints at other sites or concerns regarding overfitting, and discuss the model’s potential performance in a multi-site transfer learning setting. 6.The asynchronous interaction module models directed cross-CU attention across a token timeline. The authors should provide a quantitative discussion on the computational overhead (FLOPS, inference time) compared to standard self-attention or static GNN baselines to justify the added architectural complexity.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses an important problem and proposes a conceptually interesting framework. However, the core idea of coordination units is not sufficiently novel and is implemented using a relatively simple clustering approach. The biological interpretability is weak, and the experimental design has limitations (e.g., dataset selection, lack of cross-site validation). The performance improvements are also modest. Overall, the paper is below the acceptance threshold.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
While the authors provided detailed clarifications for several concerns in their rebuttal, their response to Concern 3 (Data Inconsistency between Table 1 and Table 2) has exposed a fatal flaw in both the methodological robustness of the proposed framework and the rigor of their experimental evaluation. Consequently, I am downgrading my rating and recommending a Reject. 1.The admission that a simple change in the cross-validation random seed caused a massive 4.2% performance drop (81.0% vs. 76.8%) on the MDD task (N=397) proves that the proposed framework is highly volatile and overfitted, as the diagnostic outcomes of 16–17 patients flip purely based on how the data is split. 2.Furthermore, utilizing unaligned random seeds across Table 1 and Table 2 violates standard benchmarking protocols and raises serious concerns regarding selective reporting (cherry-picking) to showcase optimized numbers in the main ablation table. 3.This extreme seed-dependency strongly implies that the learned coarse-grained (K=3) coordination units and asynchronous delays are merely overfitting to localized data noise rather than capturing reproducible, clinically reliable neuroscientific signatures, making the manuscript unsuitable for publication at MICCAI.
Review #2
- Please describe the contribution of the paper
BrainACU proposes a data-driven coordination unit framework that jointly models synchronous and asynchronous inter-unit dependencies in rs-fMRI, demonstrating consistent diagnostic improvements over fixed-partitioning and zero-lag baselines across multiple psychiatric disorders.
- 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 two core limitations addressed — fixed subnetwork partitioning and the synchronization assumption — are both well-motivated by neuroscientific evidence and clearly distinguished from prior work. The use of directed cross-CU attention to jointly model time-aligned and time-lagged coordination is a principled and elegant design that directly addresses the synchrony bias in existing dynamic connectivity methods. The ablation study systematically removes each component and reports consistent performance drops across all three diagnostic tasks, providing clear evidence for the contribution of CU discovery, latent alignment, asynchronous interaction, and each regularization term. The lag-step statistics further offer mechanistically interpretable evidence that disease groups exhibit larger temporal delays in cross-CU coordination, connecting the model’s behavior to known neurophysiology.
- 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.
A critical inconsistency exists between Table 1 and Table 2.The methodology states that K=3 is used throughout, yet the performance reported for K=3 in Table 2 (e.g., 81.0%/80.1% ACC/AUC for HC vs. MDD) is substantially higher than the full model reported in Table 1 (76.8%/77.9%). These two tables appear to report results under different experimental conditions without explanation, raising serious concerns about the reliability of the reported numbers. The model is constrained to K=3 coordination units across all tasks and disorders, which is a very coarse granularity relative to established brain network parcellations. The rationale for fixing K=3 is based solely on a grid search showing K=3 outperforms K=2, 4, and 5, but a broader range and task-specific sensitivity analysis would be needed to justify this as a general choice. The primary dataset (MDD and BD cohort) is private, with institution and ethics details anonymized. Combined with a relatively small sample size (246 HC, 151 MDD, 126 BD), this limits both reproducibility and the generalizability of the reported performance. The public evaluation is restricted to a single ABIDE site (NYU, 172 subjects), which is insufficient to establish broad applicability across datasets.
- 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 inconsistency between Table 1 and Table 2 undermines the reliability of the reported results, and while the core idea of asynchronous coordination unit modeling is well-motivated, the limited public evaluation and small private dataset prevent a confident assessment of generalizability.
- 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.
I appreciate the authors’ thorough responses to my comments. They have provided reasonable explanations regarding the inconsistencies between tables, the dataset limitations, and the choice of K=3.Overall, I believe this work would be of interest to the MICCAI community.
Review #3
- Please describe the contribution of the paper
This paper introduces a Transformer-based framework (BrainACU) for dynamic fMRI diagnosis. BrainACU learns data-driven Coordination Units (CUs) that eliminate the need for fixed subnetwork parcellations and predefined functional templates. Additionally, it integrates both synchronous and asynchronous coordination modeling within a unified attention-based architecture. Two method-aligned regularizers encourages different CUs to encode complementary coordination patterns and lag-aware cross-CU coupling. The framework demonstrates superior performance on one public dataset (ABIDE) and a private cohort for three diagnoses (MDD, BD, ASD), outperforming multiple state-of-the-art baselines.
- 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.
It is interesting and novel to model both time-aligned and time-lagged dependencies across data-driven coordination units via directed cross-CU attention, it enables the architecture to integrate dynamic brain network information more effectively by modeling asynchronous interactions among subnetworks.
- 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.
a) For the Asynchronous Interaction Module, the lagged interactions are modeled at the CU level, where CUs are obtained via data-driven clustering (with K=3 in experiments), raising concerns regarding interpretability. Since these CUs represent abstract groupings rather than predefined neurobiological units, and because interactions are primarily learned at this coarse level, the model’s biological interpretability remains inherently constrained. A deeper exploration of the learned interactions (e.g., attention patterns) align with established neurobiological principles would strengthen the paper. b) The paper states that hyperparameters were tuned on an internal validation set using a five-fold cross-validation scheme. However, the results reported in Table 1 (e.g., NYU: ACC =72.3%, AUC=74.1%) exceed those those in Table 2 for K=3 (ACC =71.0%, AUC =71.8%), which is quite counterintuitive. To ensure reproducibility and methodological transparency, the authors should explicitly clarify whether these results are obtained under identical data splits, train/validation/test boundaries, and evaluation protocols. Moreover, due to the limited sample size, performance variance across splits may be significant. Independent validation on larger datasets would strengthen the robustness and generalizability of the reported conclusions.
Minor points: a) The notation in Equation 3 lacks sufficient definition; please explicitly define all symbols. b) The architectural logic depicted in Figure 1 is unclear. The relationships between the modules are not clearly defined, leading to confusion. c) In the “Interpretability Analysis of Asynchronous Coordination” section, the disease cohort used for lag-step statistics is not specified.
- 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?
Although the proposed framework is innovative, its empirical validation is constrained by limited sample size and several methodological inconsistencies, which collectively undermine the reliability and persuasiveness of the reported findings.
- 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.
After reviewing the rebuttal, I find that the authors have convincingly addressed most of my concerns. The core methodological innovation of modeling both time-aligned and time-lagged functional dependencies across data-driven coordination units (CUs) via directed cross-CU attention, is conceptually sound and substantively novel. Importantly, the rebuttal strengthens the paper’s scientific grounding by clarifying the design rationale and elaborating on the biological interpretability of the learned CUs. Overall, I consider the rebuttal sufficient. The paper is reasonably motivated and presents a novel framework for modeling asynchronous brain interactions. My final decision is Accept.
Author Feedback
We thank all reviewers for constructive feedback. 1.CLARIFICATION ON THE METHODOLOGICAL CONTRIBUTION OF CU DISCOVERY (R1) (a) Our contribution is not k-means itself, but the complete pipeline integrating data-driven CU discovery with asynchronous modeling. Three distinctions from standard parcellation: (i) per-subject clustering of brain regions, aggregated via a co-association matrix, reflecting population coherence while accommodating individual variability; (ii) CUs serve as the organizational basis for lag-aware directed attention, and no prior work combines data-driven macro-unit discovery with explicit time-lagged interaction modeling; (iii) ablation (Table 1, “w/o CU Discovery”) shows fixed partitions degrade all tasks. We acknowledge k-means is standard; the novelty lies in how and why it is deployed, not the algorithm itself. (R1-Q1) 2.JUSTIFICATION FOR K=3 & BIOLOGICAL INTERPRETABILITY (R1, R2, R4) (a) We mapped CU assignments against functional networks (BD cohort). The learned CUs capture functionally coherent circuits: CU0 aggregates subcortical/limbic structures (caudate, putamen, thalamus; 55.6%) with temporal regions, consistent with cortico-subcortical loops for emotion regulation and reward processing [32]; CU1 combines prefrontal DMN (37.2%), CEN (25.6%), and visual areas (18.6%), mirroring the self-referential/executive control axis [17]; CU2 groups sensorimotor cortex (41.4%) with limbic structures (hippocampus, amygdala; 20.7%), reflecting sensory-limbic integration pathways disrupted in affective disorders [5]. No CU maps to a single canonical network, which is by design, consistent with Mitra et al. [20] who showed lag structure is orthogonal to canonical RSNs. Regarding lag interpretability (R1-Q4): if lag signals were within-CU averaging artifacts, they would affect disease and control groups equally; Fig. 2 shows disease groups exhibit systematically larger lag steps [20, 19], and this consistent disease-control divergence serves as functional validation. (R1-Q2/R1-Q4/R4-Q1) (b) K=3 is empirically optimal (Table 2) and consistent with the Triple Network Model [17], which posits three macro-level functional systems underlying psychopathology. Our data-driven CUs recover a compatible three-way granularity without imposing network priors, supporting K=3 as biologically plausible. (R1-Q2/R2-Q2/R4-Q1) 3.DATA INCONSISTENCY BETWEEN TABLE 1 & TABLE 2 (R2, R4) (a) We thank the reviewers for identifying this issue. The inconsistency arose from different random seeds for CV fold partitions between Table 1 and Table 2, which with moderate sample size (N=397) can shift CV means by several percentage points (e.g., K=3 MDD: 81.0% vs. 76.8%). Crucially, K=3 ranks first in all tasks; the relative ordering is unchanged. After unifying fold partitions, the ranking remains identical with K=3 matching Table 1.We will update the camera-ready and open-source all configurations. (R2-Q1/R4-Q2) 4.DIMENSION ALIGNMENT BOTTLENECK (R1) (a) This concern is mitigated by design: our balanced k-means variant (Sec. 2.2) constrains CU sizes to be approximately equal, and each CU has a dedicated projection W_k in R^{d x |V_k|} whose capacity scales with CU size. Sensitivity over d∈{16–64} further confirms stable performance (variance < 0.6%). (R1-Q3) 5.GENERALIZATION & DATASET LIMITATIONS (R1, R2) (a) Evaluated on full multi-site ABIDE (871 subjects, 17 sites): BrainACU achieves 69.7% ACC, outperforming BrainDGT (66.8%) and ST-GCN (57.2%), demonstrating cross-site transferability. (R1-Q5/R2-Q3) 6.COMPUTATIONAL OVERHEAD (R1) (a) Async module adds ~2.5ms inference and ~4.2% FLOPs over a static GNN baseline, operating on M patches with K=3 units. (R1-Q6) 7.NOTATION & FIGURES (R4) (a) Eq.3: A_ij=co-clustering frequency, N=training subjects, z_i^(n)=ROI i cluster label for subject n, C=total ROIs. Will redraw Fig.1 and annotate Fig.2 with cohort details. (R4-Q3/Q4/Q5) Ref: [32] Haber & Knutson, Neuropsychopharmacology 35(1), 4-26 (2010)
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- 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.
This paper presents BrainACU, a data-driven framework for dynamic brain network analysis through the discovery of data-driven coordination units and an asynchronous interaction module. While aligning asynchrony in the latent space has novelty, all reviewers noted that the small, fixed clustering parameter for the coordination unit and its influence on the projection. Several critiques were raised about the limited evaluations.
- 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.
This paper presents BrainACU, a data-driven framework for dynamic brain network analysis through the discovery of data-driven coordination units and an asynchronous interaction module. While aligning asynchrony in the latent space has novelty, all reviewers noted that the small, fixed clustering parameter for the coordination unit and limited evaluations. The rebuttal mostly addressed reviewers’ questions regarding the novelty of the paper over existing clustering-based parcellation methods, and the discussion on data and experimental results.
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.
The authors seem to have addressed the concerns raised in the original reviews and have resolved the discrepancy in reported results across tables. If accepted, I would urge the authors to maintain consistent experimental protocols as is standard practice in cross validation and make sure the biological interpretability and justification for parameter selection choice features into the explanation in the paper.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
This paper’s contribution is an asynchronous coordination framework for dynamic brain network analysis, and the idea is interesting because it aims to model temporal coordination in a less rigid way than synchronous alternatives. Although the rebuttal helped clarify parts of the coordination-unit design, the post-rebuttal signal still appears divided rather than convincingly positive, suggesting that the response improved presentation more than core confidence. On balance, the conceptual interest is not enough to overcome the remaining uncertainty about methodological necessity and validation strength.