Abstract

We propose a graph information compression framework, called Behavior-Informed Subgroup-consistent Connectome Template (BISCoT), that learns interpretable functional subnetworks from restingstate fMRI (rs-fMRI) connectivity, which simultaneously capture the heterogeneity of a diverse patient cohort. BISCoT uses multidimensional behavioral profiles to guide the decomposition of a rs-fMRI connectivity matrices into sparse yet representative subnetworks that are consistent within behavioral sub-groups. In particular, our framework adopts a graph convolution network to capture local connectivity features and applies a subgroup-informed pooling process to extract the final subnetworks. We evaluate BISCoT on an in-house dataset of individuals with autism spectrum disorder and demonstrate that the learned subnetworks improve the performance of multiple downstream prediction tasks. In addition, BISCoT extracts consistent connectivity “templates” at the subgroup level, which allows for interpretable biomarker identification.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/4637_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/zijianch/biscot

Link to the Dataset(s)

N/A

BibTex

@InProceedings{CheZij_BiSCoT_MICCAI2025,
        author = { Chen, Zijian and Beeler-Duden, Stefen and Lawson, Sophie and Jacokes, Zachary and Van Horn, John Darrell and Pelphrey, Kevin A. and Venkataraman, Archana},
        title = { { BiSCoT: Behavior-Informed Subgroup-Consistent Connectome Template for Interpretable Brain Network Analysis } },
        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

    Authors propose a novel graph information compression framework that uses multidimensional behavioral data to enforce subgroup consistency in its learned sparse representation.

  • 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 incorporation of behavior profile seems useful.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    The dataset used in this study is relatively small, while the number of learnable parameters—particularly within the GIN layer—is quite large. However, the authors did not describe any strategies employed to mitigate the risk of overfitting in this context. The purpose of incorporating geometric information gijg_{ij}gij and guvg_{uv}guv is also unclear and should be clarified. Additionally, the methodology for evaluating reconstruction performance is not well explained. It is unclear whether reconstruction loss was also applied to the other ablation baselines. Moreover, since top-K pooling and similar operations are not designed for reconstructing the connectome, it is questionable how these methods are meaningfully compared in terms of reconstruction performance. The use of “multi-task” in the title is potentially misleading, as the paper does not involve multi-task prediction simultaneously. It is testing the model on different tasks.

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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?

    One major concern is that the experiments were conducted on a single dataset comprising only 105 subjects. Given the large number of learnable parameters associated with the various pooling operations, the dataset size is relatively limited and may not be sufficient to support reliable generalization.

  • 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.

    The authors have addressed my concerns regarding the limited dataset and have provided detailed explanations and motivations for the different modules. I recommend a final decision of Accept.



Review #2

  • Please describe the contribution of the paper

    The main contribution of the paper is the development of BISCoT (Behavior-Informed Subgroup-consistent Connectome Template), a novel graph-based framework that learns interpretable and subgroup-consistent functional brain subnetworks from resting-state fMRI data. By integrating multidimensional behavioral profiles, BISCoT guides the decomposition of functional connectivity matrices into sparse, representative subnetworks that reflect neurobehavioral heterogeneity across individuals. The framework introduces a subgroup-informed pooling mechanism and uses graph convolutional networks to capture local connectivity patterns. It not only improves performance on downstream prediction tasks but also facilitates the identification of interpretable connectivity templates at the behavioral subgroup level.

  • 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 a novel framework that integrates behavioral information into brain network modeling, enabling the extraction of subgroup-consistent functional subnetworks from resting-state fMRI data. The use of behavioral profiles to guide graph decomposition is an original and impactful method, allowing the model to capture neurobehavioral heterogeneity in a clinically meaningful way. By combining graph convolutional networks with a subgroup-informed pooling mechanism, the method effectively learns sparse and interpretable connectivity templates. Its application to autism spectrum disorder demonstrates strong clinical relevance, improving prediction performance while facilitating the discovery of subgroup-level biomarkers.

  • 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.

    In BISCoT, the entire process of subnetwork selection, node scoring, and edge reconstruction is primarily guided by the behavioral embedding. While this behavior-driven method enhances interpretability, it may lead to overfitting to behavioral labels, potentially limiting the model’s ability to capture intrinsic neural variations in the fMRI data and reducing generalizability. In addition, the proportion of top-k nodes retained significantly impacts the final subnetwork representation, but the paper does not provide a detailed sensitivity analysis. Furthermore, the method is relatively complex, incurring high training costs, and its interpretability relies on multiple opaque components. The experimental dataset has a relatively small sample size, which raises the risk of overfitting.

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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

    (1) The method involves multiple components, including an AutoEncoder, GIN, attention mechanisms, and pooling layers. However, this paper does not clearly state whether these modules are trained in an end-to-end manner. It also remains unclear whether there are any gradient conflicts or convergence issues arising from jointly optimizing these components. (2) The core objective of BISCoT is to identify behavior-relevant subgraphs from brain connectivity data. However, the method is only compared against Transformer and GNN-based methods, without including dedicated subgraph discovery methods. This limits the ability to demonstrate BISCoT’s effectiveness specifically in the context of subgraph learning tasks. (3) The experiment involves only 102 subjects, which is a relatively small sample size. Given the high complexity of the model, there is a potential risk of overfitting. (4) A key limitation of the experiments is the lack of evaluation on additional clinical datasets or publicly available benchmarks such as the ABIDE autism dataset. This limits the generalizability and external validation of the proposed method. (5) The formatting of the references is inconsistent, some conference names are written in all uppercase letters while others are in all lowercase. Please check and revise them for consistency.

  • 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 overall recommendation is based on a balanced consideration of the paper’s strengths and limitations. The proposed BISCoT framework introduces a behavior-informed approach to brain network modeling, which is conceptually meaningful and clinically relevant. It effectively integrates behavioral profiles into subgraph discovery and demonstrates improved performance and interpretability in a clinical context. However, the paper lacks methodological novelty, as it mainly combines existing components such as AutoEncoder, GNN, and attention-based pooling. Moreover, key issues such as limited dataset size, potential overfitting, absence of comparisons with subgraph discovery baselines, and lack of validation on external datasets like ABIDE reduce the strength of the empirical claims. These factors collectively influenced my score.

  • 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

    The authors propose a novel method that integrates behavioral profiles into the learning process of functional brain subnetworks, guiding the selection of the most relevant brain regions and connections. They applied their approach to a cohort of autistic subjects, demonstrating higher accuracy than state-of-the-art models. Moreover, the method extracts consistent connectivity “templates” at the subgroup level, enabling interpretable biomarker identification

  • 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.

    A major strength of this paper is the incorporation of behavioral profiles through a subgroup-informed mechanism, where the authors use these profiles to guide the selection of nodes and edges. This approach is innovative compared to conventional pooling methods, which typically rely solely on graph structure.structure alone.

  • 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 authors introduce their method in the abstract without first establishing a clear rationale, motivation, or gap in the existing literature. As a result, the “why” behind the proposed approach is not well justified.

    2) While the method yields subgroup-level connectivity templates and identifies autism-related biomarkers, the paper lacks a clear explanation of how this contributes to interpretability. Many existing studies also identify biomarkers, so it would be important to clarify what distinguishes this approach in terms of interpretability compared to previous methods. biomarkers, in other words, they did identify templates and biomarked related autism but how this adds value in terms of interpretability compared to other methods ?

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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

    1) The abstract would benefit from including the motivation behind the method. Currently, the authors jump directly into the solution without explaining the “why” — the gap or challenge the method aims to address.

    2) A key question that remains unanswered: How are different behavioral profiles treated? Behavioral data can vary widely — is the embedding process adapted based on the profile type, or is a uniform embedding approach used regardless of the input variation?

    3) The authors did not provide any details on how the parameters, such as the profile embedding dimension (de = 16), temperature (τ = 0.2), and autoencoder hidden dimensions (128, 64), were tuned. It would be helpful to include information on the tuning process, whether it was done through experimentation, cross-validation, or other optimization techniques

    4) Paper Formality: the manuscript contains several typos and inconsistent reference formatting. Both should be corrected for clarity and professionalism.

  • 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 overall score reflects a balance between the paper’s novelty and its current limitations. The integration of behavioral profiles into the graph pooling mechanism is an innovative and promising direction, particularly for personalized brain network analysis in autism. The experimental results are strong and demonstrate improvements over state-of-the-art models. However, the paper lacks a clear motivation in the abstract and introduction, and the interpretability claims are not sufficiently explained or compared to existing work.

  • 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 thank all the reviewers for their constructive assessment of our paper. Below, we respond to the major concerns noted by the reviewers. Text edits (e.g., removing the phrase “multi-task” and correcting the references) will be addressed in the final version.

  1. Small Sample Size (Revs 1 & 2): The innovation of BISCoT is to use fine-grained behavioral characterizations to inform rs-fMRI subnetwork identification. Thus, we rely on a modest yet publicly available dataset of N=105 autism patients that contains both a comprehensive Sensory Profile and traditional diagnostic evaluations. The sensory profiles query a broad spectrum of sensory-related behaviors and are essential for demonstrating the utility of our method. Such behavioral information is not available in larger public datasets like ABIDE. We note that our sample size is comparable to recent GNN‑based rs‑fMRI works on ASD (e.g., BrainGNN [4], MGCN [32]).

  2. Risk of Overfitting (Revs 1 & 2): We use careful modeling choices to mitigate overfitting. First, we keep the architecture intentionally light. The GIN module relies on small MLPs rather than dense weight matrices, and the pooling operation is re-parameterized with shared weights. These choices yield far fewer free parameters than a standard GCN of comparable depth. Second, as noted in Section 2.2, our training employs L2 regularization and early stopping. Third, training follows a two-stage schedule – self-supervised training of the learned subnetworks followed by a separate training of the downstream predictor modules to showcase generalizability.

  3. Profile-Embedding Module (Revs 2 & 3): The embedding module of BISCoT uses an autoencoder to transform the discrete behavioral profiles into compact continuous embeddings. We avoid overfitting to specific behavioral traits by only using the embedding information at the pooling stage. Generalizability of the learned profiles in BISCoT is confirmed by our results on multiple downstream prediction tasks.

  4. Redistribution of Removed Edge Signals (Rev 1): The purpose of this step is to transfer information from discarded edges to “adjacent” edges (defined geometrically by the edge coordinates g_{ij}) so that the total retained information is maximized.

  5. Experimental Procedures (All Revs): Our pipeline is multi-stage and not end-to-end. First, the profile-embedding module is trained and frozen. The connectivity model is trained on the same splits with two self-supervised objectives: reconstruction and group separation. Downstream tasks, which showcase the amount of information in the subgraphs, also operate on the same splits and take the learned subgraph as input. This multistage procedure isolates the subgraph information from any task-specific tuning to demonstrate generalizability. Hyperparameters are selected based on the average validation loss across different split assignments. The pooling ratio is set to a commonly used value of 0.5 and is not tuned based on performance.

  6. Baselines Comparisons (Rev 1): Our baseline methods use identical reconstruction and classification heads for a fair comparison. We note that top-K pooling is not used for reconstruction, but rather to obtain a subgraph. Current subgraph-mining frameworks prioritize graph-inherent structures, rather than external information about behavior that is central to BISCoT. Thus we benchmarked against SOTA methods for rs-fMRI to provide an informative comparison of the value-add from considering behavioral data.

  7. Unclear Interpretability (Rev 3): The BISCoT pooling layer is Lipschitz-continuous wrt. the sensory-profile embedding, which ensures that subjects with similar behavioral profiles are assigned identical pooling masks. This strategy leads to behaviorally consistent subnetworks. For example the sensory-seeking subgraph (Fig 2) appears in 85% of participants. In contrast, standard pooling methods yield inconsistent node sets even among “similar” subjects (BrainGNN [4], MGCN [32]).




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