Abstract

The generation of connectional brain templates (CBTs) has recently garnered significant attention for its potential to identify unique connectivity patterns shared across individuals. However, existing methods for CBT learning such as conventional machine learning and graph neural networks (GNNs) are hindered by several limitations. These include: (i) poor interpretability due to their black-box nature, (ii) high computational cost, and iii) an exclusive focus on structure and topology, overlooking the cognitive capacity of the generated CBT. To address these challenges, we introduce \emph{mCOCO (multi-sensory COgnitive COmputing)}, a novel framework that leverages Reservoir Computing (RC) to \emph{learn} population-level functional CBT from BOLD (Blood-Oxygen-level-Dependent) signals. RC’s dynamic system properties allow for tracking state changes over time, enhancing interpretability and enabling the modeling of brain-like dynamics, as demonstrated in prior literature. By integrating multi-sensory inputs (e.g., text, audio, and visual data), \emph{mCOCO} captures not only structure and topology but also how brain regions process information and adapt to cognitive tasks such as sensory processing, all in a computationally efficient manner. Our \emph{mCOCO} framework consists of two phases: (1) mapping BOLD signals into the reservoir to derive individual functional connectomes, which are then aggregated into a group-level CBT—an approach, to the best of our knowledge, not previously explored in functional connectivity studies —and (2) incorporating \emph{multi-sensory} inputs through a cognitive reservoir, endowing the CBT with cognitive traits. Extensive evaluations show that our mCOCO-based template significantly outperforms GNN-based CBT in terms of centeredness, discriminativeness, topological soundness, and multi-sensory memory retention. Our source code is available at \url{https://github.com/basiralab/mCOCO}.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2633_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/basiralab/mCOCO

Link to the Dataset(s)

N/A

BibTex

@InProceedings{SouMay_MultiSensory_MICCAI2025,
        author = { Soussia, Mayssa and Mahjoub, Mohamed Ali and Rekik, Islem},
        title = { { Multi-Sensory Cognitive Computing for Learning Population-level Brain Connectivity } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15971},
        month = {September},
        page = {520 -- 529}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper introduces a multi-sensory COgnitive COmputing (mCOCO) framework to learn population-level CBT from BOLD signals endowed with cognitive traits. This model manly addressed three challenges in the process of generating CBT: (1) poor interpretability, (2) high computational cost, and (3) overlook the cognitive capacity of the generated CBT. The main contributions are as follows:

    1. First RC-based CBT Generation: Introduces a novel multi-sensory Cognitive COmputing (mCOCO) framework, leveraging Reservoir Computing (RC) to model nonlinear temporal dependencies in BOLD signals.
    2. Two-Phase Architecture: Proposes a unique two-stage approach: (1) a random reservoir generates population-level functional CBTs from BOLD signals by capturing dynamic brain activity, and (2) a cognitive reservoir integrates multi-sensory inputs (visual, auditory, textual) to endow the CBT with cross-modal memory retention and cognitive processing capabilities—a first in functional connectivity research.
    3. Cognitive CBT: Introduces the concept of a cognitively enhanced CBT that integrates multi-sensory inputs, enabling the template to reflect not only structural and topological properties but also dynamic adaptations of brain regions to cognitive tasks (e.g., sensory processing).
    4. Superior Performance: Validated on the ABIDE dataset (ASD vs. TD populations), the mCOCO-generated CBTs significantly outperform comparison method (e.g., DGN [10]).
  • 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. Coherent and Systematic Article Logic: The paper begins by clearly articulating the limitations of existing CBT methods (e.g., lack of interpretability, high computational cost, and neglect of cognitive traits) and logically progresses to propose mCOCO as a solution. Each section builds on the prior, from hypothesis formulation (RC’s suitability for brain-like dynamics) to framework design and clinical validation.
    2. Innovative Use of Reservoir Computing (RC) for Functional Connectivity Modeling: Unlike static graph-based methods (e.g., GNNs), RC’s recurrent dynamics explicitly model temporal evolution of neural activity through reservoir states (Eq. 1). This enables the framework to capture transient and long-range dependencies in BOLD signals, which are critical for robust functional connectivity estimation.
    3. Well-Designed Two-Phase Framework Architecture: Introduces a cognitive reservoir that processes multi-sensory inputs (text, audio, visual) through delayed prediction tasks. This hierarchical design ensures that the CBT first establishes robust functional connectivity before integrating higher-order cognitive traits.
  • 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. This paper highlights the limited interpretability of conventional models for constructing CBT and posits that reservoir computing (RC) offers a potential solution to enhance model transparency. However, this interpretability gain is not rigorously extended to the entire CBT generation pipeline. Need to analyze the interpretability of the model from a holistic perspective.
    2. The construction of CBT can be broadly classified into single-graph and multi-graph paradigms, depending on the heterogeneity of input modalities used to derive the population-level connectivity representation. This paper constructs CBT based on the functional connections of input, using a single view input method, while DGN[1] belongs to a multi view method. It is necessary to explain the reasons for doing so and how the model inputs are set.
    3. The comparison method chosen in this paper is not the latest. The DGN [1] selected in this article was published in 2020, and subsequent methods such as MGN Net [2] and Dual Net [3] were proposed.
    4. Lack of ablation experiments on the two-stage model, it is necessary to demonstrate the performance of removing the second stage and adding the second stage to DGN. [1] Gurbuz, M.B., Rekik, I.: Deep graph normalizer: A geometric deep learning approach for estimating connectional brain templates. Medical Image Computing and Computer Assisted Intervention (2020) 155–165. [2] M. Burak Gürbüz, I. Rekik, MGN-Net: a multi-view graph normalizer for integrating heterogeneous biological network populations, Med. Image Anal. 71 (2021) 102059, doi: 10.1016/j.media.2021.102059. [3] F. S. Duran, A. Beyaz, I. Rekik, Dual-HINet: dual hierarchical integration network of multigraphs for connectional brain template learning, in: L. Wang, Q. Dou, P. T. Fletcher, S. Speidel, S. Li (Eds.), Medical Image Computing and Computer Assisted Intervention – MICCAI 2022, Lecture Notes in Computer Science, vol. 13431, Springer Nature Switzerland, Cham, 2022, pp. 305–314. doi: 10.1007/978-3-03116431-6_29.
  • 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.

  • 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. In the Functional CBT generation section of the Methods, the author writes about “We apply this process to each subject, producing an individual functional connectivity matrix. These matrices are then aggregated to form a population-level CBT.” It is best to explain what technology was used to achieve it.
    2. The author needs to highlight the innovation of the work and demonstrate where improvements have been made.
    3. The paper mentions that the constructed model is low-cost and can be validated through additional experiments.
    4. The author may consider increasing the impact of cognitive knowledge on CBT discriminability, as well as the principles behind these impacts.
    5. The author may consider adding a description of the second stage of cognitive enhancement, highlighting the necessity and importance of this part.
    6. Reference [4] should be updated to its officially published version: Chaari N, Akdağ H C, Rekik I. Comparative survey of multi-graph integration methods for holistic brain connectivity mapping[J]. Medical Image Analysis, 2023, 85: 102741.
  • 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?
    1. Insufficient methodological novelty. While the paper demonstrates framework innovation by applying Reservoir Computing (RC) and Echo State Networks (ESNs) to connectional brain template (CBT) construction, the core methodology largely relies on established RC/ESN principles without introducing significant theoretical or algorithmic advancements.
    2. Outdated comparative baselines and insufficient justification for method selection. The chosen benchmark method (e.g., DGN ) lacks alignment with recent advancements in CBT generation. Additionally, the rationale for selecting specific comparison methods (e.g., exclusion of state-of-the-art multi-graph or single-graph approaches) is inadequately explained, weakening the validity of performance claims.
  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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.

    I appreciate the authors’ efforts in responding to the reviewers’ concerns. However, I find the rebuttal to be somewhat limited in addressing key methodological and experimental issues raised earlier.  In terms of methodological novelty, the authors state that the innovation lies in applying RC to construct CBTs and integrating this into a cognition-driven generation process. While this integration is conceptually interesting, it gives the impression that the proposed approach mainly represents a recombination of existing components, with novelty primarily stemming from applying RC in a new context. From the perspective of MICCAI standards, where stronger methodological contributions are often expected, this level of innovation may be considered incremental.  Furthermore, the proposed method is built upon a single-view framework. Since 2020, there has been increasing interest in multi-view approaches for brain connectivity modeling, due to their ability to capture more comprehensive representations. Constructing FC directly from BOLD signals offers only a static and limited view, whereas incorporating the dynamic nature of FC can provide richer, multi-perspective information. The decision to remain within a single-view setting could benefit from further justification, especially in light of these developments.  Regarding interpretability, the authors argue that RC enhances the interpretability of FC. However, as the resulting FC is still fed into a deep learning model that may be considered a “black box,” it is not entirely clear how the learned CBTs remain interpretable within this framework. A more thorough discussion or empirical evidence supporting this claim would strengthen the authors’ argument.  As for the choice of baseline models, the authors select DGN based on its performance reported in Reference [1], published in 2023. However, that work does not compare DGN with more recent and relevant methods such as MGN-Net or Dual-HINet. Additionally, the input data used in Reference [1] differ from those in the current study. As a result, relying solely on DGN as a benchmark may not provide a sufficiently comprehensive evaluation. The justification for excluding comparisons with newer or topologically-aware models—on the basis that they include additional regularization losses—also raises questions. Does this imply that the role of topological structure was not considered important in the proposed framework? Or was it deliberately excluded due to specific assumptions in RC-based CBT construction? Clarification here would help address potential concerns regarding the completeness of the methodological design.  With respect to the ablation study, while the authors emphasize that the main goal is to propose a new idea (i.e., RC-based CBT generation) rather than a modular architecture, the framework itself consists of two distinct stages. It would therefore be natural to assess the individual contributions of each stage. Such an analysis is important for validating the effectiveness of different components and would make the experimental results more convincing.  In conclusion, while the proposed direction is interesting and potentially valuable, the current version of the paper may benefit from (1) a clearer articulation of the methodological novelty beyond architectural recombination, (2) broader and more up-to-date comparative experiments, and (3) more comprehensive analysis to support claims of interpretability and design choices. Addressing these points could significantly improve the overall contribution of the work to the MICCAI community.

    [1] Chaari N, Akdağ H C, Rekik I. Comparative survey of multi-graph integration methods for holistic brain connectivity mapping[J]. Medical Image Analysis, 2023, 85: 102741.



Review #2

  • Please describe the contribution of the paper

    The paper introduces mCOCO, a novel framework that uses Reservoir Computing to generate functional population-level Connectional Brain Templates from BOLD signals. By integrating multi-sensory inputs, it adds cognitive processing capacity to the CBT, enabling both structural and functional characterization of brain networks in a computationally efficient and interpretable way.

  • 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 use of RC for modeling temporal dynamics in BOLD signals is novel in the context of CBT generation, and it offers better interpretability compared to traditional deep learning models like GNNs. 2.The paper introduces a second-stage cognitive reservoir that models memory and recall across sensory modalities, which is a creative and underexplored direction in brain network analysis. 3.Experiments show that mCOCO outperforms a strong GNN-based baseline (DGN) in terms of centeredness, topological alignment, and cognitive memory capacity.

  • 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 is evaluated only on the ABIDE dataset, which limits the assessment of its generalizability. Testing on additional datasets would better support the robustness and broader applicability of the proposed framework. 2.The evaluation is limited to DGN, without comparison to earlier non-GNN approaches such as kernel-based or manifold learning methods, which would provide a more complete performance context. 3.While results on ASD vs. TD subjects are presented, the paper offers minimal clinical insight or neurobiological discussion about the observed group differences.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The 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.

    (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?

    Methodological contribution

  • 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.

    My concerns have been addressed.



Review #3

  • Please describe the contribution of the paper

    The novel approach to this research is development of mCoco framework to learn about Connectional Brain Templates (CBT). In this framework, they also incorporate multi-sensory inputs to enrich the CBT with higher order cognitive traits. The authors approach to mCoco CBT has been evaluated to outperform GNN based CBTs. In their design of mCoco CBT, the authors also argue that their approach would approximate a solution to the black box nature of the DNN models applied to understand the temporal relationships in individual and population level connectivity.

  • 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 exploits the strengths of RCs to answer very relevant problems in functional connectomics.

    1. The multi-modal fusion obtained by using the inherently multi-modal RC is really clever.
    2. The idea of using MC as a proxy to cognitive capacity is really appreciated.
    3. The interpretation provided for different MCs obtained for different tasks is compelling
  • 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 paper uses RCs but does not do a good job of explaining them. I believe most people are not acquainted with the terminology of RCs enough to appreciate the neat innovation in this paper. For instance, 1) why do RCs only train one layer whilst holding W_in and W_res random but constant? This subtlety is lost if not explained clearly.
    2) Are the input and reservoir weights the same for all subjects?

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

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

    On balance, I believe this paper has shown enough merit to be published in MICCAI 2025.

  • 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.

    sufficient novelty merits publication.




Author Feedback

The compelling aspects of our work were highlighted by all reviewers (R1, R3, R4): (1) Novelty: the work proposed a novel framework (R1, R4) which is a first RC-based CBT generation (R3). (2) Architecture: It is a unique two-stage approach and well-designed (R3), the second-stage of cognitive reservoir is creative and underexplored direction (R4), a first in functional connectivity research (R3), and the use of multi-modal RC is really clever (R1) and innovative (R3), also the use of MC as a proxy is really appreciated(R1). (3) Superior performance: the mCOCO-generated CBTs significantly outperform comparison method (R3), outperforms a strong GNN-based baseline (R4). (4) Interpretation: the interpretation provided for different MCs is compelling (R1). (5) Logic: coherent and systematic article logic (R3).

Clarifying concerns *Use of RC (R1) and interpretability (R3): RC uses a fixed reservoir to project input data into a high-dimensional space, training only the readout layer. It also mimics how the brain processes information [13]. Its simplicity and biological grounding enhance interpretability, making it a strong fit for our CBT generation framework. While our design promotes transparency, we acknowledge that a more comprehensive interpretability analysis across the full pipeline remains an important direction for future work.

CBT construction (R3): Previous methods mostly worked with morphological networks, starting with single-graph approaches before moving to multi-graph models. Since we are the first to *directly derive CBTs from functional networks, we began with a single-graph. Extending to multi-graph is a natural next step. For aggregation, after obtaining subject-specific functional connectivity matrices (first stage), we use mean aggregation to produce the final CBT. We also tested median aggregation (as in DGN) but the mean gave better centeredness.

*Benchmark method (R3&R4): We selected DGN as our benchmark as it outperformed seven integration models for generating CBTs (including non-GNN methods and a Nature one) [4]. Although DGN was proposed in 2020, it remains state-of-the-art as shown in the 2023 review. Its effectiveness and wide adoption make it a strong baseline. We appreciate the reviewer’s suggestion, however comparing against MGN-Net and Dual-HI Net (which have additional topological and modularity-preservation losses respectively) may not be fair since our method does not implicitly integrate such traits. Hence our benchmark against “the raw” DGN is more appropriate. We will explore adding such learning traits to mCOCO in future work, where such comparison becomes plausible.

*Ablation study (R3): our work primarily explores a novel idea—leveraging RC for CBT generation—and demonstrates its potential by evaluating the resulting templates through memory capacity using multi-sensory inputs. Our goal is to demonstrate feasibility and open a new research direction, rather than proposing a modular architecture. As such, the framework is not designed as a composite model with separable components for targeted optimization, and a traditional ablation study does not directly apply. That said, we recognize the value of further empirical analysis and extensions in future work.

*Methodology novelty (R3): While our framework builds on established RC/ESN principles, we respectfully highlight that the novelty of our work lies not in redefining the core RC algorithm, but in its original application to functional brain connectivity modeling and more importantly its integration into a cognition-driven CBT generation pipeline. This represents, to our knowledge, the first use of RC for CBT generation. In line with MICCAI’s scope, our contribution emphasizes a novel application, a unique architectural formulation, and demonstrated clinical relevance through memory capacity (MC) differences between ASD and TD groups—suggesting its potential as a biomarker for cognitive decline.




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.

    Reject

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    The reviewers note the interesting application of reservoir computing to learn connectional brain templates and interesting empirical analysis using memory capacity. However, there were concerns regarding unclear methods descriptions, level of methodological novelty, and insufficient experiments. While the rebuttal helped clarify questions regarding methods and I can appreciate the new application of reservoir computing to CBT, the concerns regarding experimental comparisons (against one baseline, not including newer more sophisticated approaches or traditional baselines) remain.



back to top