Abstract

Abdominal aortic aneurysm (AAA) rupture risk prediction remains a complex and clinically important task, largely due to substantial variability in morphology, texture characteristics, and the frequent presence of incomplete clinical information. Many existing approaches rely on single-modality representations or static fusion strategies, which can limit robustness when data are missing or partially observed. To address these limitations, we propose an adaptive multi-expert multimodal fusion framework for AAA rupture classification. The proposed approach includes geometric measurements, region-based radiomics texture features, deep semantic representations extracted from mask-guided CTA slices, and structured clinical variables within an uncertainty-guided routing architecture. A missingness-aware spherical clinical embedding module jointly models clinical feature values and their corresponding missingness patterns, enabling stable representation learning under incomplete observations. In addition, a unified cross-modal expert is introduced to capture higher-order interactions between imaging and clinical representations. An adaptive routing mechanism further refines the fusion process by dynamically weighing expert contributions according to confidence-driven descriptors, allowing case-specific integration under heterogeneous structural and clinical conditions. We evaluated XMoE on a private AAA dataset and the publicly available OASIS dataset. XMoE achieves 91.25% accuracy on AAA, outperforming state-of-the-art multimodal fusion approaches, and 88.08% accuracy on OASIS. In addition, it maintains 86.25% accuracy on AAA with 30% simulated missing modality conditions.



Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3622_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/anmspro/xmoe-miccai

Link to the Dataset(s)

Private AAA dataset: Not publicly available due to Institutional Review Board (IRB) approval and patient privacy restrictions. OASIS-1 dataset: https://www.oasis-brains.org/

BibTex

@InProceedings{RobMer_Spherical_MICCAI2026,
        author = { Roby, Merjulah AND Sakib, Abu Noman Md AND Zhang, Zijie AND Muluk, Satish C. AND Eskandari, Mark K. AND Kashyap, Vikram S. AND Finol, Ender A.},
        title = { { Spherical Clinical Embedding with Adaptive Multi-expert Fusion for AAA Rupture Classification } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16885},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The main contribution is a multimodal fusion framework for AAA rupture prediction that handles missing clinical data gracefully by treating missingness itself as a meaningful signal through a spherical embedding module. This is paired with a mixture-of-experts router that dynamically decides how much to trust each modality on a per-patient basis, rather than applying a fixed fusion strategy. Together these two ideas make the framework more robust and flexible than prior approaches in this 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.

    The spherical clinical embedding treating missingness as a structured signal rather than noise is well-motivated and practically valuable for retrospective clinical data. The confidence-driven adaptive router is a principled approach to case-specific fusion, and the robustness experiments under progressively increasing missing rates are rigorous and clinically realistic. Evaluation on OASIS-1 alongside the private AAA dataset provides reasonable evidence of domain-agnostic generalizability.

  • 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.

    No analysis is provided of whether the learned routing weights vary meaningfully across patients or collapse to near-uniform distributions during training. The missing modality simulation protocol is underspecified: it is unclear whether modality branches are zeroed out or dropped entirely, and how baseline methods are adapted to handle missing inputs, which affects the validity of the robustness comparison.

  • 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?

    The paper addresses a clinically meaningful problem and proposes technically sound contributions, particularly the missingness-aware spherical clinical embedding and the uncertainty-guided adaptive routing mechanism. The robustness experiments under missing data and missing modality conditions are well-designed and the results are compelling. Evaluation on a second public dataset adds credibility to the generalizability claims. The primary concerns holding the score back are the small private dataset of only 95 patients with no external validation, the absence of statistical significance testing, and the underspecified ablation study which makes it difficult to attribute performance gains to specific design choices. These are meaningful limitations for a clinical application paper, but they do not fundamentally undermine the contribution. With stronger experimental validation the work would be a clear accept.

  • Reviewer confidence

    Somewhat confident (2)

  • [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 adequately addresses my two primary concerns. The missing-modality simulation protocol is now sufficiently specified (patient-level MCAR, 5–30% dropout, zero-imputed baselines). The slice-correlation leakage concern raised by Meta/R2 is directly refuted via patient-level mean pooling with fold-consistent assignment. Indirect evidence for routing diversity (86.25% accuracy under 30% missing) is acceptable at this stage. Remaining limitations — single-center cohort of 95 patients, coarse ablation, absence of standard deviations, and the OASIS inclusion — are real but do not fundamentally invalidate the core methodological contribution. I maintain my score of 4 (Weak Accept) and defer the boundary decision to the Area Chair.



Review #2

  • Please describe the contribution of the paper

    The main contribution is a method for multimodal integration (at least that is what it appears to me).

  • 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.Multi-modal integration framework with many bells and whistles 2.Generally well written

  • 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.

    Weaknesses: 1.The motivation is not clear: the authors state that physiological models for AAA rupture prediction are “highly dependent on segmentation, meshing, and material modeling assumptions, making them less scalable for practical clinical use.” But the presented method too depends upon segmentations being available. There are also implicit assumptions about CT acquisitions (radiomics fails without these assumptions). The primary contribution seems instead to be a method for multi-modal integration, but it is not really framed that way. 2.Inappropriate experiments: Most of the paper (including the introduction and methods) is framed around AAA rupture prediction, but one of the main experiments is on the OASIS dataset and the label used is a binary dementia label! This has no relation to the AAA task. 3.Inadequate details: Many key details are missing: a. For the AAA rupture prediction task, all relevant details are missing: for example, how many patients had a rupture, what was the time horizon for capturing ruptures, presence of interventions, what clinical factors were used, etc. 4.Basic baselines are missing: For example, what is the performance without multi-modal integration? What if only the CT was present, or only clinical factors? What about replacing the adaptive routing and fusion module with simple averaging? Possibly bad evaluation methodology: It seems that there were 4,550 unruptured and 4,627 ruptured AAA images from 95 patients, and that the evaluation was done on a per-slice basis (hard to tell as it is not clarified, and no confidence intervals are provided). However that neglects the fact that adjacent slices from the same patient are highly correlated and as a result, the metrics are likely to be inflated. One has to control for the correlations to get accurate estimates.

  • 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.

    (1) Strong Reject — must be rejected due to major flaws

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

    Not well motivated, no clear benefit to the proposed method, inappropriate methodological and experimental design.

  • 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



Review #3

  • Please describe the contribution of the paper

    The paper proposes XMoE, an adaptive multimodal expert-fusion model for AAA rupture classification.

    Author correctly identifies two practical pain points: multimodal heterogeneity and missing clinical information and makes a reasonable design choice in separating modality-specific experts from a cross-modal expert

  • 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.

    AAA rupture prediction matters, and moving beyond diameter-only heuristics toward multimodal modeling is a valid and important direction. The four-expert decomposition is sensible: handcrafted structure/radiomics for explicit descriptors, ConvNeXt-based semantic image features for learned morphology, a clinical expert for tabular variables under missingness, and a unified expert for higher-order interactions. This is one of the paper’s good aspects: the architecture matches the problem. The paper explicitly feeds the missingness mask into the clinical pathway and the unified expert, which is more principled than naïve imputation alone.

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

    1.The paper provides a reasonable high-level method description, but several implementation details that materially affect outcomes are missing or unclear: exact train/test splitting unit, how slice-level features are aggregated and how many slices are used per patient, hyperparameters for α and τ, imputation objective weighting, definitions of simulated modality missingness, preprocessing details for clinical variables, and whether segmentation masks are ground truth or auto-generated for all experiments. That is not fatal, but it is not strong reproducibility.

    2.Right now, the ablation is too coarse. It shows two modules matter, but does not isolate whether the gain is from routing, from the unified branch, from mask usage, or from simply having more capacity.

    3.The paper simulates missing data/modality from 5% to 30%, but does not specify the missingness mechanism in enough detail. Missing completely at random is much easier than clinically structured missingness. For example, sicker patients may have richer labs, and missingness can be highly informative. Since the clinical expert explicitly uses missingness masks, the exact corruption process matters a lot. Without this, the robustness claim is weaker than presented.

    4.The related-work positioning around multimodal mixture-of-experts under missing or heterogeneous modalities is currently incomplete. In particular, the authors should discuss Flex-MoE: Modeling Arbitrary Modality Combination via the Flexible Mixture-of-Experts (NeurIPS 2024 Spotlight), which explicitly studies arbitrary modality combinations and missing-modality robustness; FuseMoE: Mixture-of-Experts Transformers for Fleximodal Fusion (NeurIPS 2024), which develops MoE-based fleximodal fusion for irregular/missing multimodal inputs; and Dynamic Modeling of Patients, Modalities and Tasks via Multi-modal Multi-task Mixture of Experts (ICLR 2025), which is especially relevant for author’s design of modality-specific and shared learning. These papers would help better contextualize what is new here versus prior MoE-based multimodal fusion and missingness-aware routing.

  • 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.

  • 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?

    I am positive on the problem importance and on the overall modeling intuition. The paper’s empirical evidence supports that the proposed system is promising, but not yet that it constitutes a robustly validated methodological advance for clinical translation. I would wait for the rebuttal and see if authors can address my concerns.

  • 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.

    authors have addressed my comments fully, thus I recommend for accept



Author Feedback

We sincerely thank the reviewers and Meta-reviewer for constructive feedback and address the priority issues below.

(Meta/R2, evaluation protocol) We clarify that all evaluations are performed strictly at the patient level. Slice features are aggregated through mean pooling into a single patient-level prediction (Section 2), and all slices from the same patient remain in the same fold under 5-fold stratified CV. So, no slice-correlation leakage occurs. Section 3.1 will be reworded to state this explicitly.

(R2, segmentation dependency) Our discussion of biomechanical models intended to highlight the complexity of finite-element pipelines, which require precise segmentation, meshing, and material parameter estimation. Our framework instead uses automatically generated masks from published tool only as a soft attention prior and for radiomics extraction. No biomechanical simulation or meshing is performed. Segmentation serves as a lightweight localization prior, not a physics-dependent requirement. The introduction will be revised accordingly.

(Meta/R2, OASIS justification) Our intention was not to claim clinical transferability. OASIS was included to evaluate generalizability under a different setting involving MRI and structured clinical variables. Gains on OASIS (Table 2: 0.8808 ACC vs. 0.8485 best baseline; MCC 0.7408 vs. 0.6854) support that the routing and spherical embedding remain effective across heterogeneous modalities and incomplete data. We will make this motivation explicit in the revision.

(Meta/R2/R4, reproducibility and clinical details) The 95-patient cohort was collected retrospectively under IRB approval with a waiver of informed consent. Clinical variables cover demographics, comorbidities (hypertension, COPD, diabetes, hyperlipidemia, heart disease), labs (creatinine, BUN, BP), and risk factors (smoking, steroid use, family history, renal impairment, prior CT). Key hyperparameters are temperature 0.5, sphere dimension 64, sphere weight 0.1, contrastive weight 0.1, and imputation weight 0.1.To further support reproducibility, our full source code and training scripts will be released upon acceptance. The revision will add rupture prevalence, intervention status, time horizon, and preprocessing details.

(R1/R4, missing-modality simulation) We simulated missing modalities at the patient level under a missing-completely-at-random protocol. Each non-imaging modality was independently dropped with probability between 5% and 30%, while CT inputs were retained. The missingness mask is passed to the clinical and unified experts, and baselines receive the same zero-imputed inputs for fair comparison. Informative missingness is more clinically realistic and will be noted as a limitation.

(R1/R4 ablation; R2 basic baselines; R1 routing) The ablations in Tables 3–4 isolate the spherical and unified experts while preserving backbone capacity, so the drops reflect mechanism rather than added capacity. Removing the unified expert collapses cross-modal interaction and quantifies the value of multimodal integration over uni-modal pathways. The learned routing weights vary across patients and do not collapse, consistent with the 86.25% accuracy retained under 30% missing modality. The revision will isolate routing and mask-guided attention and add a per-patient routing-entropy analysis.

(R4, related work) We will extend the related work to include Flex-MoE, FuseMoE (NeurIPS 2024), and the multi-task MoE (ICLR 2025). Prior methods focus on generic modality dropout through learned gating, whereas our approach incorporates missingness-aware spherical clinical embedding and confidence-driven routing derived from expert logits. It enabled routing under test-time modality absence without retraining.

(R1/R2, statistics and external validation) Cross-fold std will be added in the revision. Tables 1-4 report per-fold means. The 95-patient single-cohort scale and absence of external validation are acknowledged limitations.




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.

    This work proposes, XMoE, a multimodal MoE framework for AAA rupture classification with spherical clinical embedding for handling missing data and confidence-driven adaptive routing. The clinical problem is relevant and the architecture design is sensible. However, the biggest problem pointed out by reviewers is reproducibility.

    R2 raised issues that need direct answers. First, the evaluation appears to be done on a per-slice basis across 95 patients (roughly 9,000 slices), and if so, the metrics are inflated because adjacent slices from the same patient are highly correlated. The authors need to confirm whether the evaluation unit is patient-level or slice-level. Second, R2 found the OASIS dementia experiment inappropriate since it has no clinical relation to AAA rupture. Third, R2 noted that the paper criticizes physiological models for depending on segmentation, but the proposed method itself requires segmentation, which is contradictory. R4 also flagged that the ablation is too coarse to isolate what drives the improvement (routing, unified branch, mask usage, or just added capacity), and that recent MoE work on missing modalities (Flex-MoE, FuseMoE, NeurIPS 2024; ICLR 2025) is not discussed.

    For the rebuttal, please pay more attention to the following aspects: 1.clarify whether evaluation is per-patient or per-slice and address the correlation concern; 2.justify the inclusion of the OASIS experiment; 3.discuss the related MoE papers R4 cited; 4.provide the missing clinical details (rupture prevalence, time horizon, intervention status).

    Reproducibility issue is critical because it affects the validity of all reported numbers.

  • 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 issue is solved. R2 pointed out per-slice scoring inflated every number, but the submitted method already mean-pools to patient level with each patient kept in one fold, so it can’t happen. R1 and R4 both confirmed this.

    However, the real gaps stay open. No significance tests, no standard deviations (R1, R2), single-center 95 patients with no external test (R1, authors concede), and no CT-only, clinical-only, or simple-average baselines (R2). These are either barred new work or pushed to revision, so the core “fusion helps” claim is under-supported.



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.

    While the initial review process highlighted a stark divergence between the panel, the authors’ comprehensive rebuttal successfully defused the major methodological concerns. Crucially, the authors clarified that their evaluation was executed strictly at the patient level via cross-validation, completely refuting the threat of slice-correlation data leakage. They also clarified that their pipeline avoids expensive, unscalable biomechanical meshing, using segmentation solely as a soft prior. Although the single-center nature of the 95-patient cohort and the inclusion of the unrelated OASIS dataset for out-of-domain evaluation remain minor experimental constraints, the model’s performance stability under simulated missingness (up to 30%) provides compelling validation of its architectural robustness. The reviewers are satisfied with the rebuttal, and the paper is recommended for acceptance.



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.

    After considering the reviews and rebuttal, I recommend acceptance. The paper addresses a clinically important AAA rupture classification problem and proposes a missingness-aware multimodal MoE framework with spherical clinical embedding and adaptive expert fusion.

    The rebuttal clarifies the most critical concern: evaluation is performed at the patient level, with slice features pooled into one patient-level prediction and all slices from the same patient kept within the same fold. This substantially mitigates the concern about slice-level leakage.

    My acceptance recommendation relies on the rebuttal clarification that the AAA evaluation was performed strictly at the patient level. The final version should explicitly state the patient-level aggregation, fold assignment, clinical cohort details, and cross-fold variability.



back to top