Abstract

Dynamic functional connectivity (dFC) analysis has revealed that functional connectivity fluctuates over short timescales, reflecting the intrinsic transitions of brain among multiple states. However, dFC data typically exhibit the characteristics of high dimensionality and noise, making it difficult to extract stable and accurate states. Furthermore, accurately identifying model order (i.e., number of states) is challenging due to lack of prior knowledge. To address the above issues, we propose a model order-free method for extracting stable states. Our method can simultaneously capture multi-scale state information and improve the stability of the state. Furthermore, our method esti-mates the number of states adaptively based on data-driven methods. Based on synthetic data, we evaluated the effectiveness of our method. The results showed that, compared to traditional methods, our method not only accurately estimated the number of states but also extracted states with greater robustness and precision. Additionally, we evaluated the effectiveness and sta-bility of the method using fMRI data from 602 healthy controls and 519 schizophrenia patients. Results demonstrated that our method exhibited significant consistency among the states extracted by multiple runs. Moreover, we identified reliable biomarkers for schizophrenia. In conclusion, we propose a novel state extraction method that does not rely on predefined state numbers, while accurately and stably identifying states.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/5388_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{FanSon_AModel_MICCAI2025,
        author = { Fang, Songke and Calhoun, Vince D. and Pearlson, Godfrey and Kochunov, Peter and van Erp, Theo G.M. and Du, Yuhui},
        title = { { A Model Order-Free Method for Stable States Extraction in Dynamic Functional Connectivity } },
        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

    This paper proposes a model order-free method for extracting stable states from dynamic functional connectivity (dFC) data in resting-state fMRI. The method preserves multi-scale clustering information by aggregating stable clusters across multiple model orders and refines them via inter-cluster similarity propagation using a random walk-based graph. The resulting meta-states are shown to be stable across multiple runs and are used to identify schizophrenia-related connectivity biomarkers.

  • 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.
    • Model Order-Free State Inference: The paper presents a novel mechanism that avoids pre-defining the number of clusters in dFC analysis, addressing a longstanding issue in the field.
    • Multi-Scale Stability Aggregation: Leveraging clustering ensembles and random walks across a graph of cluster similarities is original and grounded in solid theoretical motivations (e.g., modularity optimization and spectral refinement).
    • Robustness and Stability Validation: Results on both synthetic and real-world datasets (n=1121) clearly show high inter-run consistency and accurate state recovery even under varying noise levels.
    • Clinical Relevance: The method successfully identifies reliable FC alterations in the subcortical-cerebellar circuits of schizophrenia, which is biologically plausible and well-motivated.
  • 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.
    • Theoretical Justification of Stability Score Threshold: The paper adopts a threshold (e.g., 0.7) for cluster stability without systematic justification. Is this a tunable hyperparameter? How does performance change across different thresholds?
    • Limited Comparison with Recent Deep Models: While traditional clustering methods are compared, deep clustering or representation learning approaches are not considered.
    • Lack of Ablation Analysis: There is no ablation study on key components like the random walk step size, voting mechanism, or use of Louvain community detection. Their individual contributions to performance remain unclear.
    • Model Generalization: The paper evaluates only schizophrenia-related datasets. Testing the method in other psychiatric conditions (e.g., MDD, ASD) could help establish generalizability.
  • 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
    • Consider reporting runtime and computational complexity—how scalable is the proposed ensemble clustering and graph-based propagation?
    • The biomarker analysis is compelling but could be strengthened by visualizing group-level meta-states or reporting classification performance (AUC, accuracy) if used in downstream tasks.
    • Please clarify how subject-level windows are used for voting—do all subject windows vote equally? Are longer subjects weighted more?
  • 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 novelty and robustness of the method are clear, especially the model order-free design and the multi-scale graph construction. However, missing comparisons with deep clustering models and lack of ablation weaken the empirical evaluation. A strong rebuttal that addresses these points could shift the score upward.

  • 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 paper presents an method for adaptively extracting meta-states from dynamic functional connectivity. It makes two key contributions:1) By employing varying model orders and multiple initializations, the method obtains clustering solutions across multiple scales, then selects stable clusters as the basis for subsequent meta-state identification. This approach not only leverages rich multi-scale information but also reduces the impact of unstable clusters on meta-state detection, resulting in more obvious cluster structures and robust extraction. 2) The use of community detection algorithms enables adaptive estimation of the number of meta-states without requiring prior knowledge, thereby mitigating potential biases introduced by manual selection.

  • 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) Multi-scale information utilization: Unlike approaches that rely on a single scale, the method employs multiple model orders to derive clustering solutions, thereby leveraging multi-scale information for meta-state extraction. 2) Robust cluster selection: By exclusively utilizing highly stable clusters for subsequent meta-state identification, the approach naturally mitigates the influence of unstable clusters. This enhances the distinctness of the resulting meta-state structures and improves the reliability of the extraction process. 3) Adaptive meta-state estimation: The method adaptively determines the optimal number of meta-states in a data-driven manner, eliminating the need for manual specification and reducing potential biases introduced by subjective selection.

  • 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) Redundancy in the Abstract: The phrase “uncovering biomarkers for schizophrenia” is unnecessarily repeated (e.g., “while also uncovering biomarkers for schizophrenia… Moreover, we identify reliable biomarkers for schizophrenia”). 2) Repetitive descriptions in Section 2.2: The sentence “Each window is assigned to the meta-states with the highest voting score” appears twice with only minor variations (e.g., “…highest voting score: …” vs. “…highest vote count”). 3) Equation error in Section 2.2: Equation 2 contains symbol inconsistencies that need correction. 4) Punctuation error in Section 2.2: The sentence “As shown in Fig. 1(B), we propagate inter-cluster similarities through random walk. to capture potential relationships…” incorrectly splits the thought with a period mid-sentence. Additionally, the methodological description in Section 2.1 should be elaborated further to avoid potential ambiguity in interpretation.

  • 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.

    (4) Weak Accept — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    My recommendation falls slightly below ‘weak accept’ for two reasons: 1) The manuscript contains some writing issues, including repetitive phrasing and punctuation errors that need correction. 2) Code availability was not mentioned, though this would substantially improve reproducibility and enable follow-up studies.

  • 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 a model order-free method for extracting stable states in dynamic functional connectivity (dFC) from resting-state fMRI. Unlike traditional clustering approaches that require predefined model orders, the proposed method leverages an ensemble of clustering results across multiple model orders, evaluates cluster stability, and integrates multi-scale information via a novel inter-cluster similarity graph followed by community detection. This approach enables adaptive estimation of the number of connectivity states and achieves robust, reproducible extraction of meta-states. The method is validated on both synthetic and large-scale clinical fMRI datasets, and successfully identifies disease-relevant biomarkers in schizophrenia.

  • 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 proposed approach addresses a significant limitation in dFC state analysis, i.e., the reliance on a fixed and often arbitrary number of clusters, by adaptively determining the number of states in a data-driven manner. 2) The introduction of a multi-scale random walk on a cluster similarity graph is an innovation that improves the identification of coherent meta-states across scales. 3) The method is comprehensively evaluated on synthetic datasets with known ground truth and a large multi-site schizophrenia cohort, showing superior performance.

  • 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 involves several hyperparameters (e.g., number of model orders, clustering runs, stability thresholds, random walk steps). Although some values are drawn from literature, an ablation study analyzing the sensitivity to these parameters would strengthen the paper. 2) Lack of comparison with state-of-the-art methods. The paper only did comparison with traditional clustering, no baseline involving deep learning-based dFC modeling is included, which could limit the perceived impact in cutting-edge methodological contexts. 3) The performance improvement over K-means with correlation distance is small. Is the added complexity of multi-scale stability estimation and graph-based meta-state refinement justified? 4) The evaluation relies solely on NMI for synthetic data, which, while informative, is insufficient on its own; a more comprehensive assessment including stability, temporal dynamics, and interpretability metrics would strengthen the validation of the proposed method. 5) The paper would benefit from an ablation study to disentangle the contributions of key components, such as stability-based cluster selection, multi-scale random walk refinement, and community detection, to the overall performance.

  • 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.

    (4) Weak Accept — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The core idea is good and the clinical application is valuable, but the evaluation and analysis need further depth to meet a higher bar.

  • 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.

    Some of my concerns have been addressed. This paper could be accepted with a clearer explanation of its distinctions from conventional methods, along with more detailed descriptions of the experimental setup and methodology




Author Feedback

We appreciate all the constructive comments and suggestions. We will meticulously incorporate the valuable feedback into the camera-ready version. Code: We will make the well-documented code publicly available. Response to Common Questions: [R#3&R#4] Comparison with deep models: Previous deep learning-based dFC analysis methods typically utilize neural networks for feature extraction and then employ traditional K-means for state extraction (Qiao, C. et al. 2023, https://doi.org/10.1016/j.media.2023.102941). These methods primarily focus on obtaining embedding features for clustering, whereas our method aims to extract reliable states based on predefined features. Indeed, our method can be applied to features extracted by deep learning. We will clarify this in the final paper. [R#3&R#4] Ablation analysis: 1) Community detection and voting. Community detection aims to adaptively estimate the meta-states. The voting is designed to compute the meta-state label to which each window belongs. They are integral components of our method, making it challenging to completely decouple them for ablation experiments. 2) Stability clusters selection and random walk. Our ablation experiments on synthetic data demonstrate that removing the stable cluster selection process results in instability in the number of states, and discarding the random walk leads to reduced accuracy. Due to space constraints, we did not include these results in the paper. [R#3&R#4] Hyperparameters: 1) Clustering runs. Increasing the number of clustering runs can enhance result stability. For computational efficiency, we set the parameter to M = 40. Experimental results demonstrate that this choice yields highly stable outcomes (Table 2, std=0.01). 2) Stability threshold. The parameter was set to 0.7 to retain stable base states for subsequent graph construction for random walk-based community detection, thereby enhancing the stability of the resulting meta-states. Response to R#1: (Q1-Q4) Redundancy in the abstract; Repetitive descriptions in section 2.2; Equation error in section 2.2; Punctuation error in section 2.2. We will remove redundant content and correct the mistakes in the camera-ready version. (Q5) The methodological description in section 2.1. We will incorporate more descriptions of extraction of stable cluster in the camera-ready version. Response to R#2: (Q1) Performance improvement with K-means. Different from K-means that depends on a fixed and often arbitrary the number of clusters, our method enables adaptive estimation of cluster number and higher stability of the cluster centers. On synthetic data, our method accurately identifies the number of states and yields significantly stable results (std = 0.01), whereas K-means exhibits instability even when given the optimal number of states. (Q2) More indicators for synthetic data. We calculated the standard deviation (across 20 runs) in the synthetic dataset to evaluate the stability (Table 2). Due to space constraints, we did not include the relevant dynamic indicators (such as fractional occupancy). Response to R#3: (Q1) Model generalization. We will evaluate the method on other psychiatric conditions in the future. In fact, we have already applied the method to the extraction of co-activation patterns and achieved promising results. (Q2) Runtime. Runtime is mainly limited by the generation of basic states. For extremely large-scale datasets, this limitation can be mitigated through parallel computing or by adopting the mini-batch K-means. (Q3) Visualizing group-level meta-states. We visualized the extracted group-level meta-states in Figure 4A. (Q4) Voting. The voting score for each window of each subject is calculated independently, based solely on whether the window belongs to a specific meta-state. Therefore, it is unrelated to the total number of windows for each subject.




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



back to top