Abstract

Functional magnetic resonance imaging (fMRI) measures task-evoked neural activity through blood-oxygen-level-dependent (BOLD) signals and enables activation mapping to localize task-related brain responses. Conventional activation mapping typically relies on the general linear model (GLM), which assumes linearity and a predefined hemodynamic response function (HRF). However, mounting evidence indicates that BOLD responses are nonlinear and state-dependent. Meanwhile, the high dimensionality and low signal-to-noise ratio of fMRI further challenge the linear assumption. These limitations motivate a neural-manifold view, in which high dimensional task dynamics evolve on a low-dimensional state space. We propose a geometry-aware manifold trajectory modeling framework that represents task-related activity as structured trajectories. The framework identifies a central baseline cluster, computes geodesic distances from each embedded frame to this cluster along the manifold trajectory graph, and converts them into vertex-wise activation maps via the proposed Manifold Geodesic-based Similarity (MGS) without HRF fitting. Using the Human Connectome Project (HCP) motor-task fMRI dataset, we uncover a petal-like looped trajectory, indicating a reproducible low-dimensional organization of task-evoked dynamics. At the individual-level, MGS robustly recovers canonical motor activations despite substantial single-subject fMRI noise. At the group-level, compared with HRF-based GLM, MGS yields more discriminative activation patterns and highlights task-related effects less emphasized by GLM, demonstrating the advantage of geometry-aware manifold trajectories for task fMRI activation mapping. Code is available at https://github.com/PCMLYR/ManifoldTraj_MGS

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0473_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/PCMLYR/ManifoldTraj_MGS

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LiYue_GeometryAware_MICCAI2026,
        author = { Li, Yueran AND Chang, Xinle AND Yang, Yang AND Zhao, Yu AND Su, Jingyong},
        title = { { Geometry-Aware Manifold Trajectory Modeling for Task fMRI Activation Mapping } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16894},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes a data-driven method to model temporal dynamics of task fMRI data using manifold clustering and trajectories, which, unlike the commonly-used GLM approach, does not depend on the assumption of linearity and selection of the hemodynamic response function, which varies across brain regions and subjects. The method averages the time series of 100 scans to produce a smoother and stronger response to the motor task, which makes the manifold projection and clustering more robust. The temporal clustering of the manifold projects clearly show that the different conditions can be clustered correctly when accounting for the delay of the hemodynamic response, which motivates the study in the first place. The condition-level group analysis shows activation maps that are comparable to the GLM baseline.

  • 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 geodesic distance to the baseline cluster is an interesting way of representing the group-level response to the different conditions in the task. 2.The proposed approach does not assume linearity in the BOLD response and doesn’t depend on defining a hemodynamic response to retrieve task-related activation maps.

  • 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 advantage provided by the Manifold Geodesic-based Similarity over averaging over the BOLD signal on the same folds (MGS without the weight vector w) is not shown. 2.No quantitative analysis was described. 3.The proposed approach uses the group-level geodesic distance to weight the importance of individual-level responses, which can bias the individual-level activation maps towards the group response.

  • 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.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    Major concerns:

    1.My main major concern is that the benefit of using UMAP projections, clustering in the manifold space and geodesic distances is not shown over a simple averaging over the BOLD frames within the condition blocks. The MGS equation only uses the geodesic distance in the weight vector w. The manuscript would highly benefit from showing a comparison with an “MGS” activation map that is calculated without the weight w. This would completely remove the group-level bias and would show how much the MGS activation maps depend on the process of projecting, clustering and calculating the distance at the group level. This “weightless” MGS could still take the resting-state blocks just like the proposed MGS does, and would make the comparison direct. 2.The manuscript is limited to a qualitative analysis. It would benefit from a quantitative analysis that, for instance, calculates the sensitivity, specificity and Dice of the MGS-recovered maps using the GLM activation maps as the ground truth. 3.The manuscript also lacks an individual-level trial study to see if the proposed MGS approach can recover trial-level maps better than the GLM or hemodynamic deconvolution methods such as Paradigm Free Mapping and Total Activation.

    Minor comments:

    1.Figure 2 would highly benefit from being consistent with the colours across UMAP plotting and clustering plotting. 2.It is not clear whether the resting state pseudo-blocks were picked randomly or in the same order as with the task data. 3.It is not clear if each frame (time point) is projected along the geodesic path. I think they are but authors should make that clearer. 4.I believe that Figure 3 shows $\Delta \text{MGS}$ and not MGS.

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

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

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

    The decisive factor in this recommendation is the novelty of representing group-level task responses as manifold trajectories without assuming a specific hemodynamic response function or linearity in the BOLD signal. There is clear potential in using the proposed UMAP embedding space to study individual- and group-level responses to different types of stimuli, and the geodesic distance to the baseline cluster is an interesting and principled way of encoding task-evoked dynamics.

    That said, the contribution is weakened by the absence of a quantitative analysis and by a missing ablation that is critical for interpreting the results. The activation maps shown are qualitatively compelling, but the manuscript would need to report sensitivity, specificity and Dice against a GLM reference to substantiate the claims more rigorously. More importantly, the benefit of the manifold projection and geodesic weighting over a simpler averaging of the BOLD frames within condition blocks is not demonstrated. Without a comparison against a weightless version of MGS, it is difficult to assess how much of the observed performance comes from the manifold geometry itself rather than from the block-averaging procedure. These are addressable weaknesses, but they do keep this submission at the borderline.

  • 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.

    The authors addressed all three of my major concerns and my four minor clarifications in their response. On the weightless-MGS ablation (Q1), they argued that the geodesic weighting is, by construction, the component that separates MGS from a normalized block-averaged BOLD summary. The argument is internally consistent but does not constitute the empirical evidence I requested, and the question of whether the manifold geometry adds information beyond block averaging therefore remains open in the submitted manuscript. On the quantitative comparison with GLM (Q2), the authors reported Dice and sensitivity values in the rebuttal, but per the MICCAI guidelines these cannot enter the camera-ready and I do not weight them in this assessment. On the individual-level trial study (Q3), the authors acknowledged the value of a comparison against single-trial GLM or hemodynamic deconvolution methods but did not offer it. The four minor clarifications were addressed correctly, and the authors committed to releasing the code through an anonymous repository.

    In line with the AC guidance that meta-reviewers should not penalize papers for failing to provide new experimental results in the rebuttal, I do not downgrade my recommendation on those grounds. My initial Weak Accept was based on the manuscript as submitted, and the rebuttal did not surface anything that materially changes that assessment. The novelty of the manifold-trajectory framing and the principled use of geodesic distance to baseline as a data-driven, HRF-free signature of task engagement outweigh the rigor gap, and I maintain my positive recommendation.



Review #2

  • Please describe the contribution of the paper

    A task fMRI activation mapping framework based on low-dimensional manifold trajectory and geodesic distance modeling is proposed; the MGS index is designed to achieve the transformation from geometric dynamics to vertex-level activation maps; and the potential of this method in restoring classical motion activation patterns and improving the compactness of population activation maps is verified on HCP motor-task data.

  • 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 topic is novel, attempting to introduce low-dimensional manifold trajectory geometry into task fMRI activation mapping; the methodological framework is relatively complete, and the MGS index has a certain degree of interpretability; on HCP motor-task data, it can recover the topology of motor body parts with neurophysiological rationality and shows a more compact group activation pattern compared to traditional GLM.

  • 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 paper points out that compared to GLM, MGS yields “more compact and discriminative” group activation maps. However, the current support for this point in the paper mainly relies on qualitative visualization comparisons, lacking sufficient quantitative results for comparison. For example, methods such as IOU comparison of activation regions lack quantitative comparison results.

    2.When comparing different modules, the paper lacks comparisons of their effectiveness. For example, is the performance derived from manifold modeling itself, shared templates, or geodesic distance?

    3.Some parameter settings in the paper are based on empirical values, lacking sensitivity analysis of these parameters, such as sensitivity analysis of different n_neighbors, min_dist, and distance metrics.

    4.The paper uses a fixed 6-frame delay alignment, calling it coarse latency alignment rather than HRF fitting. The authors should provide the rationale for using this parameter, whether it is based on experience or experimental conclusions.

  • 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 provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

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

    (2) Reject — should be rejected, independent of rebuttal

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

    The method presented in this paper is novel, but it lacks sufficient experimental evidence to demonstrate its effectiveness.

  • 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 rebuttal partially addresses my concerns. In particular, the authors provide additional quantitative comparison with GLM using Dice and sensitivity, and they give a more reasonable explanation for the fixed 6-frame delay based on the HCP TR and plausible BOLD timing.

    However, my main concerns are not fully resolved. The response still does not provide ablation studies that separate the contributions of manifold modeling, shared templates, and geodesic distance. The explanation of replacing the geodesic weight with an all-one vector is useful, but it remains conceptual rather than experimental. Similarly, the authors acknowledge that broader sensitivity analysis would be useful, but they do not provide actual sensitivity results for n_neighbors, min_dist, or distance metrics.

    Therefore, while the rebuttal improves the clarity of some points, I do not think it fully resolves the methodological validation concerns.



Review #3

  • Please describe the contribution of the paper

    This manuscript proposed a geometry-aware manifold trajectory modeling framework that characterizes task-fMRI activation as structured trajectories. The main idea is to calculate the geodesic distances to a baseline cluster on the manifold trajectory, and convert them into vertex-wise activation maps via the proposed Manifold Geodesic-based Similarity (MGS) without requiring HRF fitting. The HCP motor-task dataset was used to evaluate the effectiveness of the proposed method and showed superior performance compared to conventional HRF-based GLM methods.

  • 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 background and motivation is well justified. 2.The formation of brain task activation mapping using manifold geodesics is novel. 3.Both qualitative and quantitative evaluations are performed to demonstrate the effectiveness of the proposed method.

  • 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.Lacking independent reproducibility data. Although the authors used HCP motor task data, it is unclear whether the identified pattern and mapping characteristics is reproducible in other motor-related data, even in other task fMRI data.

    2.Lacking details for subject selection. They merely used a subset of 100 subjects from the HCP dataset. However, it is unclear why not using all subjects (~1000) in this dataset.

    3.Lacking comprehensive comparisons. The authors should perform comprehensive comparisons between the proposed method with conventional GLM-based methods to show its superiority in identifying meaningful activation patterns. Moreover, the authors should compared the proposed method with other manifold-based methods.

    4.Lacking justification of parameter-setting. Some of the parameter selection is unclear. For example, different embeddings and different latency shifts.

  • 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.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

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

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

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

    The formation of brain task activation mapping using manifold geodesics is novel.

  • 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.

    The authors have addressed my concerns about data inclusion, etc. I appreciate the novelty of this study in fMRI analysis.



Author Feedback

Thank all reviewers for valuable comments. Our responses are presented as follows.

1.Quantitative comparison vs GLM (MR-Q1, R1-Q2, R2-Q1, R3-Q3) We thank the reviewers for requesting quantitative evidence. We had computed Dice and sensitivity between MGS and GLM group maps for the five motor conditions (Dice: 0.61, 0.66, 0.82, 0.72, and 0.79 for l/r hand, l/r foot, and tongue; sensitivity: 1 for all). Still, we think using GLM as absolute ground truth is not fully appropriate, since our goal is to examine whether task-activation information encoded in looped manifold trajectories can be translated into vertex-wise activation statistics under possible HRF mismatch. Thus, the experiment is not merely a qualitative comparison with GLM, but an evaluation of the activation information carried by the looped manifold trajectory. The GLM comparison further confirms that MGS recovers canonical motor activation patterns.

2.Effectiveness of individual components (MR-Q1, R1-Q1, R2-Q2) We agree that a weightless-MGS control is a meaningful ablation. In Eq. (2), w(dG) transfers manifold-trajectory information into the vertex-wise activation statistic. Replacing it with an all-one vector would reduce MGS to a normalized block-average BOLD summary and discard the baseline-departure/return profile learned through UMAP, DBM clustering, and geodesic distance estimation. Thus, the geodesic weighting is the component that separates MGS from simple block averaging. Task condition blocks were used to provide the same design basis as GLM, where condition blocks are convolved with an HRF, enabling direct comparison while avoiding HRF fitting in MGS. Future work can further reduce task-label dependence by defining task-related segments from the unsupervised loop trajectory.

3.Parameter setting (MR-Q2, R2-Q3&4, R3-Q4) The 6-frame shift equals 4.32 s at the HCP TR of 0.72 s and falls within a physiologically plausible range for motor/sensorimotor BOLD timing. Chen et al. (2023), “BOLD Response is more than just magnitude: Improving detection sensitivity through capturing hemodynamic profiles”, reported motor-cortex HRF peaks around 3–4 s, while canonical HRF models peak around 5 s. Fig. 2A also supports this choice, as the 6-frame shift gives clearer condition-coherent petals than 0-frame labels. The shift was fixed before MGS computation. For UMAP, n_neighbors=40, 2D embedding, and standardized Euclidean distance were fixed before activation inference as moderate settings to preserve the smooth group-level trajectory and reduce vertex-wise scale effects. We agree that broader sensitivity analysis would be useful, but these parameters were not tuned to optimize the final activation maps.

4.Dataset scope and selection (R1-Q3, R3-Q1&2) Fig. 3 reports individual-level ΔMGS maps, showing subject-level motor patterns. We agree that direct individual-level comparisons with GLM or deconvolution methods would further strengthen the evaluation. We also tested HCP Emotion and Relational tasks and observed consistent MGS behavior, but omitted them due to the MICCAI page limit. For subject selection, we used the official HCP 100-unrelated subset to reduce family-structure dependence and evaluate the method on a diverse independent cohort. We expect the same conclusion to hold on the full HCP-YA 1113-subject dataset.

5.Code availability (MR-Q3) We apologize for not preparing open-source code at submission. We have prepared an Anonymous GitHub repository named /r/MGS-0B07/. If accepted, the repository will be moved to GitHub and linked in the final paper.

6.Minor clarifications (R1) We thank R1 for the careful reading and detailed suggestions. We agree on Fig. 2 color consistency. Rest pseudo-blocks followed the same order as task blocks. Each frame was projected along the estimated geodesic path. Fig. 3 indeed shows ΔMGS. We will clarify these points if accepted.

We hope these responses clarify key points. Thanks again for all the thoughtful evaluations.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    Perhaps the most pervasive approach to perform segregational analysis in neuroimaging, inc. fMRI, is by using a characterization of the response function as a factor in the general linear model. Although less popular due to their inherent complexity, topological and geometrical approaches can also carry out such analysis with some collateral benefits over the plain statistical counterparts (although of course, strictly speaking the GLM is a manifold, but we hardly think of it as such). This paper is one good example of topological and/or geometrical approaches. The manuscript presents a framework to characterize brain activation at group level that exploit the geometry of the underlying manifold, where the actual modelling occurs, to make statements about task-related effects that would otherwise be missed by the GLM.

    All reviewers coincide in highlightin the novelty of the approach, and whilst perhaps for me this is less so (I happen to work very closely to this field), but I do coincide that there is novelty here.

    Among the weaknesses, the lack of quantitative evidence (R1/R2/R3) and the arbitrary decisions over the parameterization (R2/R3) have dominated the reviews. Coupled with the lack of direct access to the code (R1/R3), hence limiting the reproducibility, these shortcomings have somewhat limit the potential that the drat could have had.

    Nevertheless, given the two more senior reviewers, and I concurr, in supporting acceptance, but the split decision, I recommend inviting a rebuttal. Specifically, the rebutall must address:

    • Affording quantitative evidence inc. comprehensive comparison (R3) e.g. vs classical GLM, and about the effectiveness of the individual components (R2).
    • Provide rationale and/or empirical evidence to justify parameter setting.
    • Disclose code details
  • 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.

    The pre-rebuttal meta-review identified three major issues: (1) lack of quantitative evidence and comparison against classical methods such as the GLM; (2) unjustified parameter choices; (3) unavailability of source code. The rebuttal only partially addressed these points with Reviewer #2 explicitly maintaining the Reject score after the rebuttal. Quantitative benchmarking and ablation studies remain absent, and advantage over the gold standard (GLM) has been attempted but not clearly evidenced. Despite genuine novelty, the evidentiary basis for acceptance is insufficient at this time.



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.

    Despite some lingering concerns, the authors have done a good job addressing the main critiques to extent allowed by MICCAI (i.e., no new experiments). This work would be of interest to the community.



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.

    All reviewers are satisfied with the quality of the rebuttal responses and have voted to accept the paper. I think the authors have done a good job addressing all concerns raised during the review and plan to address the comment on reproducibility upon acceptance by releasing their Github repository.



back to top