Abstract

Various neuroimaging studies suffer from small sample size problem which often limit their reliability. Meta-analysis addresses this challenge by aggregating findings from different studies to identify consistent patterns of brain activity. However, traditional approaches based on keyword retrieval or linear mappings often overlook the rich hierarchical structure in the brain. In this work, we propose a novel framework that leverages hyperbolic geometry to bridge the gap between neuroscience literature and brain activation maps. By embedding text from research articles and corresponding brain images into a shared hyperbolic space via the Lorentz model, our method captures both semantic similarity and hierarchical organization inherent in neuroimaging data. In the hyperbolic space, our method performs multi-level neuroimaging meta-analysis (MNM) by 1) aligning brain and text embeddings for semantic correspondence, 2) guiding hierarchy between text and brain activations, and 3) preserving the hierarchical relationships within brain activation patterns. Experimental results demonstrate that our model outperforms baselines, offering a robust and interpretable paradigm of multi-level neuroimaging meta-analysis via hyperbolic brain-text representation.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0365_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{BaeSeu_MNM_MICCAI2025,
        author = { Baek, Seunghun and Lee, Jaejin and Sim, Jaeyoon and Jeong, Minjae and Kim, Won Hwa},
        title = { { MNM: Multi-level Neuroimaging Meta-analysis with Hyperbolic Brain-Text Representations } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15960},
        month = {September},
        page = {400 -- 409}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This work proposes a novel framework that leverages hyperbolic geometry to bridge the gap between neuroscience literature and brain activation maps. The proposed method uses hyperbolic space, which expands exponentially under constant negative curvature. In neuroimaging meta-analysis, it seeks to identify the most relevant brain regions from a given textual description or retrieve corresponding brain activations. This paper’s goal is to embed brain activation data and LLM-representation of its corresponding neuroscientific text into a shared space while preserving both the brain-text hierarchy as well as the structural hierarchy of the brain. This tries to connect brain activation and text information in feature space.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    Hyperbolic space is effectively utilized to represent brain activation data and LLM-representation of its corresponding neuroscientific text. Brain activation map is well estimated by the proposed method. Mathematical point is clearly described.

  • 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.

    Although the mathematical points are presented well, it is hard to see the entire flow of the proposed method. This will be improved by augmenting Implementation Details. This paper is neuroimaging metaanalysis paper and it does not handle real images. I have no idea this paper fits to MICCAI. But it may be ok.

  • 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.

    (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?

    This paper bridges the gap between textual information and the brain activation map.

  • Reviewer confidence

    Somewhat confident (2)

  • [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

    The authors introduce a hyperbolic embedding framework for neuroimaging meta-analysis, aiming to bridge the gap between textual descriptions of neuroscience articles and corresponding brain activation maps. By leveraging hyperbolic space, the framework is able to capture semantic associations and hierarchical structures inherent in both brain data and text. This approach is more suitable for representing the tree-like hierarchical organization of the brain compared to traditional Euclidean embeddings.

  • 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 paper proposes a novel brain structural hierarchy guidance within the hyperbolic space. This method ensures that broader, more general brain activation patterns (e.g., entire hemispheres) are positioned closer to the origin of the hyperbolic space, while specific, localized brain regions (e.g., individual gyri) are placed further away, better capturing the hierarchical nature of brain organization. This enhances the interpretability of the learned embeddings and facilitates multi-level neuroimaging meta-analysis.

  • 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 framework described in the paper is quite complex, involving the integration of multiple hierarchical layers, contrastive learning objectives, and a specialized Lorentzian model. While these innovations are promising, the paper could be critiqued for potentially making the model difficult to interpret, especially for clinicians or researchers not familiar with hyperbolic geometry.

    While the theoretical background of hyperbolic geometry is well-detailed, it might overshadow the practical implications and real-world applications of the model. The paper provides robust theoretical justifications, but the empirical validation primarily focuses on cross-modal retrieval and brain activation map prediction, which may not be fully representative of the model’s potential in clinical or broader neuroscience applications.

    The model heavily depends on large-scale pretrained language models (Mistral-7B) and specific hierarchical regularization, which could lead to overfitting if not carefully handled, especially in cases where the data used for training might have biases or limited coverage of the full spectrum of neuroimaging data.

    The experimental results are based on cross-modal retrieval and brain activation prediction, but the evaluation metrics used (such as recall@K) may not be sufficient to evaluate the full capabilities of the model in more complex or multi-faceted tasks. Moreover, no statistical significance tests or confidence intervals are provided to support the observed improvements.

  • 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.

    (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?

    As listed above

  • Reviewer confidence

    Not confident (1)

  • [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 manuscript proposes a framework, MNM, that introduces a hyperbolic embedding approach to improve neuroimaging meta-analysis by capturing the hierarchical relationships between neuroscience text and corresponding brain activation maps.

  • 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.

    By embedding both modalities in a shared Lorentzian hyperbolic space, the approach aligns brain-text pairs in the same representational space while preserving structural hierarchy in the brain. Overall, the method is well-motivated and technically sound.

  • 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.

    Motivation of the study: while the manuscript does explain the motivation for introducing the hyperbolic space to better discern the brain vs text representations, I still don’t see the intuitive advantage of doing so. Perhaps Fig 1 can be better used to clarify this point of why having text and brain in 2 different points in the representational space would offer an advantage.

    1. Model assumption: it would be great if the authors can elaborate on the assumption or the constraint that brain activation maps should lie “higher” along the xtime axis. Would the results change when a reversed constraint is placed, i.e. text embedding is higher than bran embedding along the xtime.
    2. Experimental validation: for the text-brain retrieval result in Table 1, would it be more consistent with the setups in the original Text2Brain and NeuroQuery papers to use Pearson’s correlation with a masked brain to match a predicted brain image with candidate brain images.
    3. Results: Figure 3a and b: please include the the hyperbolic origin in the two histograms a) and b) to make interpretation easier Figure 3c: please include the origin or arrows to guide the interpretation of the figure. Figure 4: “While NeuroConText [12] fails to capture activation in the temporal lobes using Yeo17 networks, MNM assigns high similarity scores to them as in the ground truth.” This is a rather odd result and assertion. The temporal activation in Figure 4 doesn’t typically correspond to visuospatial attention and not match the ground truth reference.

    Minor point: Terminology such as “x_time” and “external angle” may confuse readers unfamiliar with Lorentzian geometry. Please consider adding intuitive explanations or footnotes.

  • 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.

    (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?

    The overall technical and writing quality of the paper have met the bar for acceptance

  • 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 reviewers for their insightful feedback and for recognizing our work’s contributions, especially acknowledging the effectiveness of hyperbolic space for brain-text alignment in neuroimaging meta-analysis. Below, we address the main concerns:

[Q1] Motivation & Flow (R1, R4, R8) [A1] Hyperbolic space is inherently well-suited for modeling hierarchical relationships due to its geometry, where volume expands exponentially with radius. Specifically, in the Lorentz model (visualized in Fig. 2), moving along the time-like axis away from the hyperboloid origin increases the surface area, allowing angular neighborhoods to encompass more points. This geometry naturally represents broader, more general concepts closer to the hyperboloid origin (lower points), while finer, more specific details are positioned further away (higher points). In the context of neuroimaging meta-analysis, our approach leverages two key hierarchical intuitions: (1) Structural hierarchy of the brain: As depicted in Fig. 1 (left), large brain regions (e.g., hemispheres) can be hierarchically subdivided into finer subregions (e.g., ROIs). We capture this by positioning embeddings of brain regions with more widespread activations closer to the hyperboloid origin using a hierarchical loss. This is empirically validated by the ‘MNM w/o L_hier’ ablation study in Table 1 and Fig. 3. (2) Semantic hierarchy between brain regions and text: A single brain region can be associated with multiple specific semantic concepts expressed in text (Fig. 1, right). Thus, a brain region can be seen as a more general concept relative to the specific textual descriptions. To reflect this, we enforce that brain embeddings are positioned closer to the hyperboloid origin than their corresponding, more specific text embeddings. The effectiveness of this approach is confirmed by the ‘MNM-Reverse’ ablation study in Table 1.

[Q2] Figure Clarification (R1) [A2] We appreciate the suggestion to enhance figure interpretability. For Fig. 3 (a-b), we will replace axis indicators with one-directional arrows. Although the center of the Poincaré disk is the hyperboloid origin, our current depiction includes two disjoint semicircles; thus, we will revise Fig. 3(c) for improved clarity. We apologize for the mislabeling of “temporal lobes” in Fig. 4, the correct region is the “intraparietal sulcus”, and this will be corrected in the revision.

[Q3] Potential Applications & Scope (R8) [A3] A key strength of our approach is its flexibility and ability to model hierarchical relationships. While we adopted the data settings of NeuroConText [12] in this version to clearly isolate and demonstrate the specific benefits of our proposed methodology, our framework can be readily extended to other brain atlases or LLM-derived textual embeddings. This adaptability, combined with its hierarchical modeling capacity, allows for broader applicability across neuroimaging studies with varying spatial resolutions. Crucially, when multiple studies address similar topics but annotate non-identical yet hierarchically related brain regions (e.g., one study annotates a broad area, another a specific sub-region within it), our method can naturally bridge these discrepancies. This addresses a significant limitation of prior fixed-point alignment approaches, which struggle with such variations.

[Q4] Evaluation Metrics (R1, R8) [A4] We usedRecall@K for fair comparison with the NeuroConText [12]. We agree broader metrics, such as Pearson correlation on masked brain regions, would enrich analysis and plan these for an extended journal version where space permits.

[Q5] Reproducibility (R4, R8) [A5] While our data settings and training strategies closely follow those of NeuroConText [12] (which has already released its code), we will also publicly release our code for full reproducibility, as noted in our implementation details section.




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • 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



back to top