Abstract

Post-traumatic epilepsy (PTE) is a severe complication of traumatic brain injury (TBI). Yet, early identification remains challenging due to the complex structural and functional alterations it induces in the brain. To address this, we propose a dynamic multimodal Mixture-of-Experts (MoE) framework that integrates functional and structural connectivity through time-aware functional-structural encoding and class-conditioned expert routing. Within this framework, modality-specific and cross-modal experts learn complementary representations, while a Modality-Class MoE (MCoE) module dynamically adjusts expert weights according to each classification objective. Experimental results across three binary classification tasks demonstrate that the framework consistently outperforms static fusion baselines, and interpretability analyses further reveal meaningful interactions among regions of interest (ROIs). This dynamic multimodal expert framework effectively captures class-dependent brain interaction patterns and provides an interpretable approach for PTE diagnosis and risk stratification.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4376_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{DinJun_DynFSMoE_MICCAI2026,
        author = { Ding, Jun-En AND Chen, Spencer AND Noren, Henry AND Valdivia, Daniel AND Yohn, Christine AND Patel, Suhina AND Zink, Taylor AND Sun, Hai AND Liu, Feng},
        title = { { DynFS-MoE: Dynamic Functional-Structural Mixture-of-Experts for Post-Traumatic Epilepsy Diagnosis } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16886},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    In this work, the authors propose a dynamic multimodal Mixture-of-Experts framework that integrates functional and structural MRI through time-aware functional-structural encoding and class-conditioned expert routing. By combining time-aware functional-structural encoding with class-conditioned expert routing, the model adaptively integrates fMRI and sMRI information and captures class-specific brain patterns.

  • 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.sMRI can leverage a temporal MoE mechanism to select the most informative patches within fMRI segments for cross-modal fusion and routing. 2.Modality-specific experts are employed to extract individual feature representations from each modality, producing dedicated embeddings that preserve modality-unique information. 3.A class-based MoE module adaptively selects the most class-relevant modality patches to facilitate more discriminative classification.

  • 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 number of subjects in the dataset is relatively small, with only around 20 samples for each type. Will this affect the validity of the results? Should a larger dataset be used to validate the findings? 2.More state-of-the-art methods should be included in the comparative experiments for model performance evaluation. 3.In addition to discussing binary classification problems, is it feasible to apply the model to a three-class classification task that distinguishes among HC, TBI and PTE? 4.The text in Figure 2 is too small and should be optimized for better readability. 5.A more comprehensive ablation study is recommended.

  • 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.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

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

    While the paper presents a novel DynFS-MoE framework, a dynamic multimodal Mixture-of-Experts approach for PTE diagnosis post-TBI, the associated experimental data, result analysis, and other key components remain to be further optimized and improved.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    This paper aims to address a highly challenging clinical task: diagnosing the risk of progression to post-traumatic epilepsy (PTE) following traumatic brain injury (TBI) using structural MRI (sMRI) and functional MRI (fMRI). To tackle the issues of information loss and modality competition present in traditional fusion strategies, the authors propose a dynamic multimodal Mixture-of-Experts framework, DynFS-MoE. This method includes a time-aware encoder, Modality-Specific Experts (MSoE), and a core Modality-Class MoE (MCoE) module, which dynamically allocates feature patches using a 3D class-conditioned routing mechanism. Additionally, the paper introduces a brain interpretability regularizer based on Conditional Mutual Information (CMI) to prevent routing collapse. Evaluated on a cohort of 69 subjects, DynFS-MoE significantly outperforms existing static graph network baselines across three binary classification tasks and provides highly compelling spatial-temporal interpretability for epileptogenesis.

  • 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 MCoE module is a significant innovation in this paper. Unlike standard MoEs that route solely based on input features, MCoE constructs the routing logits as a 3D tensor, explicitly introducing the conditional constraint of “disease class”. This allows the model to dynamically adjust its reliance on different modalities according to different disease phenotypes. 2.The interpretability analysis is exceptionally well-executed and aligns perfectly with clinical logic. The derived routing weights compellingly demonstrate that TBI diagnosis heavily relies on static structural connectivity (sMRI), whereas PTE diagnosis cleverly down-weights the reliance on sMRI and shifts focus to dynamic functional abnormal discharges in early/mid-temporal segments (S2-S4). Furthermore, the ROI interaction heatmaps reveal progressive “frontal-occipital” abnormal connectivity unique to PTE, which is highly consistent with existing neuropathological literature. 3.Utilizing scaled dot-product attention to inject global sMRI structural priors into local fMRI temporal patches via residual connections is a lightweight yet highly effective strategy. This cleverly mitigates modality competition prior to entering the MoE routing layer.

  • 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.Critically Small Dataset and High Overfitting Risk: The entire study is based on only 69 subjects (26 HC, 24 TBI, 19 PTE). Training an MoE architecture containing complex routing mechanisms and regularizations on such a tiny dataset is highly susceptible to severe overfitting. 2.Severe Disconnect Between Figures and Text: There is a confusing disconnect between the framework diagram and the textual description. For instance, Section 2.4 specifies that the input to MCoE incorporates both cross-modal fusion and modality-specific processing from prior layers, but in Figure 1, the input to MCoE appears to solely come from the cross-modal fusion path, demonstrating a clear inconsistency. 3.Lack of Explicit Mathematical Constraints for “Modality-Specific” MSoE: In Section 2.3, the authors claim that MSoE extracts “modality-specific features”. However, structurally, MSoE is merely a standard MoE layer. The paper lacks any explicit loss function or physical isolation to force a certain expert to “only process fMRI” or “only process sMRI”; it relies entirely on data-driven learning. 4.Missing Unimodal Baselines and Key Ablation Details: Since one of the paper’s claimed innovations is that “dynamic multimodal fusion is superior to unimodal methods,” comparisons with unimodal baseline models must be included. Moreover, the current ablation study only verifies the routing strategy, lacking a systematic component-wise ablation of the core modules.

  • 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 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.

    (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?

    Please refer to the major weaknesses in Section 7.

  • 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.

    N/A



Review #3

  • Please describe the contribution of the paper

    The paper introduces a DynFS-MoE framework that utilizes a class-conditioned joint distribution to dynamically route modality patches to specialized experts. This is a significant step beyond static fusion 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.

    Develops a time-aware functional-structural encoder that partitions fMRI into temporal segments and uses scaled dot-product attention to integrate structural connectivity features; MoE-based dynamic modality fusion; strong performance on a clinical dataset; A Brain Interpretability Regularizer based on conditional mutual information is used to reveal disease-specific brain interaction patterns.

  • 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 study relies on a very small cohort of 69 subjects, which raises significant concerns about the robustness and generalizability. The paper lacks a comparison against a standard multimodal fusion baseline, such as simple feature concatenation, to isolate the specific benefit of the MoE routing; he paper does not report $p$-values or formal statistical significance tests to confirm that the improvements over baselines are statistically robust.

  • 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

    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?

    noval framework, effective multimodal fusion; limited data used in experiment, generalizability concerns.

  • 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



Author Feedback

General Response[R1,R2,R3]: We sincerely thank the AC and reviewers for their constructive comments.

  • Sample Size Issue:We spent 4 years collecting this dataset because PTE is a relatively rare neurological disorder requiring long-term follow-up for definitive diagnosis. Acquiring both fMRI and DTI data is also costly, while preprocessing and manual curation are highly time-consuming. Existing studies used similarly limited cohorts: [1] included 24 HC and 22 TBI subjects [2] included HC: n=13, TBI: n=16, and TBI converters/seizure group: n=17.To our paper survey, no prior study has simultaneously included larger TBI and PTE cohorts with both fMRI and DTI modalities. Moreover, our work is the first to jointly investigate fMRI and DTI for TBI/PTE-related epilepsy using a deep learning framework. [1] Kang et al. (2025), Relationship Between Structural and Functional Network Connectivity Changes for Patients With Traumatic Brain Injury and Chronic Health Symptoms. [2] Raizman et al. (2020), Traumatic Brain Injury Severity in a Network Perspective: A Diffusion MRI–Based Connectome Study.
  • Overfitting in Small Samples [R2]:To mitigate overfitting in a relatively small dataset, we adopt a lightweight MoE with class-conditioned dynamic routing instead of a large dense architecture, activating only selected experts per sample to control effective capacity. We further apply conditional mutual information regularization to constrain modality–expert routing, improving stability. DynFS-MoE consistently improved performance, while ablation studies confirmed the importance of class-conditioned routing.
  • Comparison with Unimodal Methods and Statistical Significance [R2,R3]: Following the reviewer’s suggestion, we conducted Welch’s t-test between DynFS-MoE and the best baseline methods. For HC vs. PTE, DynFS-MoE improved ACC, AUC, and F1-score by 5.0%, 7.7%, and 9.5%, respectively. Improvements in AUC and F1-score were highly significant (p<0.001), while ACC also achieved significance (p<0.05). Compared with the best unimodal models (fMRI-FC MLP or sMRI MLP), DynFS-MoE further improved AUC and F1-score by 15.1% and 11.0%. For TBI vs. PTE, DynFS-MoE improved ACC by 6.9% and AUC by 22.4% over the best baseline, both with high significance (p<0.001). We also compared our approach with recent SOTA models, including MultiViT (2024), IBrainGNN (2025), and AGIIBM (2025). DynFS-MoE consistently outperformed competing methods across all cohorts, achieving ACC gains of 2.7%, 7.7%, and 10.0%, respectively. [R1-2]In addition, DynFS-MoE (AUC=0.67) can flexibly support 3-class classification and achieves an average improvement of 2.41% across 3-class classification tasks.
  • Figure–Text Inconsistency[R2]: [R2-2] We thank the reviewer for identifying the inconsistency between Fig. 1 and Section~2.4, and we will revise the arrow connection to MCoE in Fig.1.Section~2.4 states that the input to the MCoE is the refined modality patch embeddings $Z={z_i}{i=1}^{|P|}$ after cross-modal fusion and modality-specific processing. Each $z_i$ incorporates both the cross-modal fusion outputs from Section~2.2 ($\tilde{f}^{(k)}$, $z{\text{fused}}^{(k)}$) and the MSoE outputs from Section~2.3.Therefore, MCoE operates on joint representations containing both cross-modal interactions and modality-specific specialization. [R2-3] In the MSoE module, fMRI (color blocks) and sMRI (red-highlighted blocks) are already represented as modality-specific inputs. Specialization arises from the modality-defined patch space, routing mechanism, and class-conditioned dispatch. Specifically, DynFS-MoE constructs modality patches $P={f^{(1)},\dots,f^{(K)},s}$ and routes them through $P_\theta(E=e_j\mid P=p_i)$, enabling experts to specialize to different modality distributions. In addition, the class-conditioned routing and conditional mutual information regularizer $I(P;E\mid Y)$ encourage modality- and class-dependent expert associations instead of uniform expert sharing.




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.

    The three reviewers offer a mixed overall recommendation, with two weak accepts and one weak reject. They confirmed the novel DynFS-MoE framework for its dynamic, class-conditioned routing and clinically meaningful interpretability, but unanimously raise serious concerns about the extremely small cohort (only 69 subjects) that heightens overfitting risk and limits generalizability. Further shared weaknesses include the absence of unimodal and simple fusion baselines, a lack of statistical significance testing, inconsistencies between text and figures, and insufficient ablation studies. The authors should prepare the rebuttal to address those concerns.

  • 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.

    After considering the initial reviews, the authors’ rebuttal, and the updated reviewer assessments, the paper is recommended for acceptance. The reviewers acknowledged the novelty of the DynFS-MoE framework and its clinically meaningful interpretability. The authors also addressed the reviewers’ concerns in the rebuttal by providing statistical tests that confirmed significant improvements over baselines, clarifying the framework’s design and correcting the figure inconsistency, and justifying the dataset size by the clinical rarity of post-traumatic epilepsy. The work’s dynamic class-conditioned routing and multimodal fusion, despite the data limitations, are seen as sufficiently valuable for publication.



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.

    Two reviewers (Weak Reject, Weak Accept) did not submit post-rebuttal assessments. The third reviewer recommends to accept this paper. I share this assessment.

    Reviewers appreciated particularly the methodological contribution of this work paired with the conducted (interpretability) analyses. A shared main criticism centred around the small sample size of the dataset. However, I agree with the authors that acquiring a suitable, large dataset is challenging, given the application. I consider the provided evidence as sufficient and would down-weight this critique.



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.

    This paper contributes a clinically motivated dynamic functional-structural mixture-of-experts framework for post-traumatic epilepsy diagnosis, with a sensible attempt to combine heterogeneous signals in a difficult small-sample setting. The rebuttal seems to have clarified the design motivation and the overfitting-mitigation perspective, which keeps the paper acceptable after rebuttal.



back to top