Abstract

Alzheimer’s disease (AD) is the most prevalent neurodegenerative disorder, and current treatments cannot reverse its progression, making early diagnosis and intervention essential. Deep learning methods have shown promise for diagnosing AD from neuroimaging data, but these models are often affected by both visible and hidden sources of noise. Visible noise typically reflects systematic imaging differences, such as scanner-related biases and acquisition artifacts, whereas hidden noise is linked to individual variation that can lead models to depend on spurious patterns that do not generalize well. Most existing approaches emphasize statistical associations and do not adequately account for these confounding influences. Motivated by these limitations, we propose a multimodal, confounder-aware framework for AD diagnosis inspired by causal reasoning. The model includes a region-masking component (Mask Decoupling Module) that probes the image by making small, targeted changes to localized areas, then highlights the regions that most strongly affect the prediction, encouraging the network to prioritize structurally informative features. We also introduce a covariate-guided feature modulation mechanism that uses clinical and demographic variables to adjust local image features, thereby reducing variation attributable to these factors. Although the framework does not estimate causal effects directly, it is designed to learn more stable representations by limiting reliance on shortcuts, improving robustness and generalization in AD diagnostic models. Our code is available at \url{https://anonymous.4open.science/r/Causal_fusion-4E6E}.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1600_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/redtea-code/Causal_fusion

Link to the Dataset(s)

N/A

BibTex

@InProceedings{CheYua_AConfounderaware_MICCAI2026,
        author = { Chen, Yuanhao AND Chen, Jingwen AND Xu, Jinyi AND Yu, Haoqi AND Wang, Bin AND Li, Chunzhong AND Zhang, Yongquan AND Fan, Fenglei AND Elazab, Ahmed AND Wan, Xiang AND Wang, Changmiao},
        title = { { A Confounder-aware Representation Learning Framework for Alzheimer’s Disease Classification } },
        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

    The main contribution is a causally inspired confounder-aware AD diagnosis framework that improves model robustness by explicitly reducing sensitivity to imaging artifacts and demographic biases through masking-based region analysis and covariate-conditioned feature modulation.

  • 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 is clearly written and well structured, making the proposed method easy to follow. The experimental evaluation is comprehensive and well designed, with appropriate comparisons that support the claims. In addition, the proposed modules are well motivated, with a clear connection to the problem of confounding factors in Alzheimer’s disease diagnosis, which strengthens the overall methodological design.

  • 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.

    A major weakness of the paper is that the ablation study suggests the main performance gain comes from the CAPM component rather than the MDM module, which is more prominently emphasized in the manuscript. In the comparison with state-of-the-art methods, it is also observed that two tabular-information-based models achieve strong performance, This creates a potential mismatch between the narrative emphasis and the empirical contributions. In particular, the specific advantage of CAPM over existing confounder-aware approaches is not clearly articulated, and the role of MDM appears less impactful than suggested in the main text. The paper would benefit from a clearer explanation of the distinct contribution of CAPM and a more balanced discussion of the actual empirical contributions of each module.

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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 is generally well written, clearly structured, and easy to follow. The proposed method is well motivated, and the experimental setup is comprehensive with meaningful comparisons. However, my concerns regarding the relative contributions of the proposed modules influenced my final evaluation. I rated the paper as a weak accept rather than a stronger accept.

  • 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

    1.A prediction-sensitivity-based mask learning strategy for spatially robust representations. The paper proposes the Mask Decoupling Module (MDM), which estimates the importance of local image regions by masking individual patches and measuring the resulting change in model output. This provides a structured, perturbation-driven mechanism to separate predictive regions from spurious ones, encouraging the network to allocate representational capacity toward disease-relevant brain structures rather than incidental imaging patterns. 2.A covariate-conditioned feature modulation module for confounder suppression. The paper introduces the Confounder-Aware Patch Modulation (CAPM) module, which integrates tabular clinical and demographic variables to explicitly adjust patch-level visual features. By learning a per-patch scaling factor conditioned on covariates such as age and sex, the module provides a direct mechanism to attenuate variation in image representations that is attributable to confounding factors rather than disease status, promoting more robust multimodal fusion. 3.A unified, end-to-end multimodal framework inspired by causal reasoning for AD diagnosis. Rather than relying on multi-stage pipelines or post-hoc correction, the paper combines MDM and CAPM into a single jointly trained framework with a composite loss function that simultaneously encourages reliance on causal features, penalizes dependence on non-causal image regions, and prevents the tabular modality from dominating the prediction. This design offers a structurally simple yet principled approach to learning confounder-aware representations from heterogeneous data sources.

  • 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.Principled problem formulation grounded in causal reasoning. The paper clearly articulates the confounding problem in AD diagnosis using a causal dependency diagram (Z → X, Z → Y), and designs each module to address a specific edge in the causal graph. This structured motivation, with MDM targeting spurious image regions and CAPM targeting covariate-linked variation, provides a coherent and interpretable framework, even if the implementation does not perform strict causal inference. 2.Joint handling of both visible and hidden confounders in a single end-to-end framework. Unlike prior methods that address either imaging artifacts or covariate confounding in isolation, this framework tackles both within a unified, end-to-end trainable pipeline. The adversarial losses on the non-causal branch and the tabular branch (maximizing their classification errors) provide an elegant mechanism to simultaneously suppress two distinct sources of spurious correlation without requiring multi-stage training. 3.Strong empirical performance with consistent gains across two independent datasets. The model achieves substantial improvements over the strongest baseline on ADNI (+10.4% ACC, +16.5% REC, +18.3% F1) and also reaches state-of-the-art on the external NACC dataset. Consistent improvement across both in-distribution and cross-cohort settings suggests that the framework does learn more generalizable representations, despite the architectural concerns raised above.

  • 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.Architectural mismatch between CNN and patch-level causal reasoning. The MDM is designed to assess the contribution of individual patches to the prediction. However, the framework uses a CNN (ResNet) as the backbone, where deeper-layer features aggregate information across the entire spatial extent through hierarchical convolutions. This means that masking a single patch in the input does not cleanly isolate that patch’s contribution at the feature level, as the effect propagates non-linearly across spatially overlapping receptive fields. A Vision Transformer (ViT), which operates on explicit patch tokens with self-attention, would be a more architecturally consistent choice for patch-level importance estimation.

    2.Questionable validity of the mean-image prior. The ranking-consistency loss uses patch scores from the training set’s voxel-wise mean image as a structural prior. However, averaging across all subjects simultaneously smooths out both disease-relevant anatomical variations (e.g., heterogeneous hippocampal atrophy patterns) and confounding noise. It is unclear whether the resulting patch ranking reflects diagnostically meaningful regions or merely morphologically stable areas across the population. Using an ambiguous supervisory signal to guide mask learning may misdirect the model.

    3.Per-patch modulation in CAPM ignores global spatial dependencies. The modulation coefficient γ is computed independently for each patch, assuming that confounders affect patches in isolation. In reality, confounding factors such as aging-related brain atrophy have spatially correlated, global effects across the brain. Independent per-patch adjustment fails to capture these inter-region dependencies, potentially producing spatially incoherent adjusted features. Scale of γ is unconstrained, and subsequent Pool3D undermines the modulation. The paper does not explicitly constrain the range of γ. If γ approaches or exceeds 1, the adjusted features (1−γ)⋅X(1 - \gamma) \cdot X (1−γ)⋅X can collapse to near-zero or become negative, causing numerical instability. More critically, the adjusted features are immediately passed through a Pool3D layer, which aggregates across the spatial dimensions indiscriminately. This pooling blends heavily suppressed patches with minimally affected ones, effectively diluting the selective suppression that CAPM is designed to achieve. The modulation and the subsequent aggregation work against each other.

    4.Lack of clarity on the multi-class AUC computation. The task is a three-class classification (NC vs. MCI vs. AD), yet the paper reports AUC without specifying the computation strategy (e.g., one-vs-rest macro or weighted averaging, or one-vs-one). Given the substantial class imbalance in both datasets (e.g., ADNI has NC:MCI:AD = 355:303:569), the choice of averaging scheme can significantly affect the reported numbers. This omission limits reproducibility and fair comparison.

  • Please rate the clarity and organization of this paper

    Satisfactory

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

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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?

    The paper addresses a meaningful problem, learning confounder-aware representations for AD diagnosis, and the causal framing provides a coherent motivation. The dual-module design (MDM + CAPM) and the composite adversarial loss are conceptually appealing, and the empirical results show consistent improvements over baselines on two public datasets. However, several concerns regarding the core module designs prevent me from recommending acceptance at this stage.

    1.The MDM relies on patch-level perturbation to estimate regional importance, yet the backbone is a CNN (ResNet) whose hierarchical convolutions cause deep features to aggregate information well beyond individual patch boundaries. Masking a single patch at the input level does not produce a spatially isolated effect at the feature level, which undermines the theoretical justification for treating patch-wise scores as meaningful estimates of local causal relevance. A patch-native architecture such as a Vision Transformer would be more consistent with this design.

    2.The mean-image prior used to supervise mask learning is not well justified. Averaging across all training subjects smooths out both disease-related variation and confounding noise, making it unclear whether the resulting ranking signal reflects diagnostically informative regions or simply morphologically stable areas.

    3.The CAPM module computes a per-patch modulation coefficient γ independently, ignoring the fact that confounders like age-related atrophy have spatially correlated global effects. Moreover, γ is not explicitly constrained in range, and the modulated features are immediately aggregated by Pool3D, which indiscriminately blends heavily suppressed and minimally affected patches. This effectively dilutes the selective suppression that CAPM is intended to achieve, creating a tension between the modulation step and the subsequent aggregation.

    4.The paper reports AUC for a three-class task without specifying the computation strategy (one-vs-rest, one-vs-one, macro or weighted averaging), which limits reproducibility given the notable class imbalance in both datasets.

    In summary, while the motivation is sound and the results are promising, the gap between the causal reasoning framework and the architectural choices that implement it raises fundamental questions about whether the modules function as claimed. I would be open to reconsidering if the authors can provide convincing explanations or additional experiments addressing these concerns in the rebuttal.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    Reject

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

    The paper presents a genuinely interesting causal framing for AD diagnosis and demonstrates consistent empirical gains across two independent datasets, which are meaningful contributions that work in the authors’ favor; however, the reviewer’s core concern about the architectural mismatch between the CNN backbone and the patch-level causal reasoning was only partially addressed, and while the ViT ablation result was a welcome addition, it does not fully resolve the theoretical inconsistency embedded in the MDM design; the rebuttal on CAPM was also somewhat underdeveloped, particularly regarding the Pool3D aggregation issue and the unconstrained range of γ, where the authors’ explanation relied heavily on the claim of a misunderstanding without providing sufficient clarifying evidence; on the positive side, the strong cross-cohort generalization results on NACC and the largely satisfactory responses to the prior loss and AUC concerns do leave a realistic window for acceptance; on balance however, the unresolved architectural tensions in the two core modules remain the primary obstacle, and the rebuttal did not provide enough theoretical or empirical grounding to fully close the gap between the causal reasoning motivation and the actual implementation choices made in the paper.



Review #3

  • Please describe the contribution of the paper

    This paper presents a multimodal, confounder-aware representation learning framework for AD classification. The MDM module identifies and prioritizes disease prediction relevant patches and CAPM weights patches based on tabular covariates, which are useful for achieving high performance.

  • 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 motivation is well-formulated in the context of localized disease information, spurious correlations and confounder effects in brain images for disease diagnosis. The results are strong over baselines, the ablation is included, and the attention maps are biologically reasonable.

  • 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.

    First, the paper is a little over-emphasing the causal terminologies, while the method is more related to perturbation-based feature importance and prediction sensitivity, not causal effects. Second, many baselines are vision-only methods (CausalMixNet, CAAM, MAD-Former), so they may get disadvantages without some key tabular covariates like clinical assessments that highly correlate with diagnosis. Third, in the CAPM module, it might be simplistic to decrease weights for particular patches uniformly across all dimensions, since different features could have different correlations with confounders. Fourth, MDM module doesn’t seem to contribute much to the performance in the ablation (wo MDM’s F1 and AUC are close to full model). Fifth, the population voxel mean image as a prior may discourage the identification of individual-level disease-related variations.

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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?

    The paper presents a causality-inspired, technically sound, sophisticated multimodal framework for AD classification. However, it’s unclear whether it’s the tabular features themselves that contribute more to high performance or the CAPM module. Ablation on MDM also does not fully demonstrate importance of the module.

  • 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.

    Although the justifications for CAPM and MDM modules are still not convincing, the paper has strong empirical evidence and interpretability presentations.



Author Feedback

We thank all reviewers for their comprehensive and constructive feedback on our paper, particularly regarding “module contributions” (R1 and R3), “implementation of CAPM” (R2 and R3), and “prior score loss” (R2 and R3). We first address three general issues and then provide detailed responses to each reviewer.

G1: Module Contributions MDM and CAPM address distinct issues: MDM penalizes unstable local dependencies via spatial regularization; CAPM adjusts for covariate-induced variation. When combined, MDM raises ACC from 0.819→0.833 (ADNI) and 0.802→0.831 (NACC); macro F1 and AUC also improve modestly (ADNI: F1 +0.012, AUC +0.009; NACC: F1 +0.010, AUC +0.008). Gains are modest but consistent, indicating CAPM is the primary performance driver, while MDM provides additional spatial regularization, interpretability. Because ADNI and NACC contain few severe artifacts or local noise, MDM’s marginal gain is expected. This complementary behavior aligns with our framework’s design rationale, where CAPM handles global covariate-related variation and MDM addresses local image instability.

G2: Implementation of CAPM We acknowledge confounders may have spatial dependencies. We clarify three points: (i) Each \gamma is computed from all confounders, enabling cross-patch learning, and acts on patch-level representations, not raw pixels. (ii) \gamma scale is unconstrained, but a residual connection prevents numerical instability. (iii) Pooling is applied per patch, preserving selective suppression; a global pooling claim is a misunderstanding. The current \gamma is a per-patch scalar (uniform across feature dimensions) for stable optimization;

G3: Prior Loss The mean image of the training set is not used for diagnostic supervision. Instead, the ranking-consistency loss acts as a weak, early-stage regularizer with small weight, linearly decaying to zero during training. It stabilizes initial mask learning, preventing degenerate or symmetric score distributions. In ablation experiments, removing this term led to larger variance in early training loss and slower convergence. The final masks are determined predominantly by the task-driven classification objective and perturbation-based MDM, ensuring that the ranking term does not bias final predictions.

R1Q1: Advantage CAPM is not a simple tabular branch, it conditions local patch features on covariates. TabFormer alone achieves AUC 0.795 but ACC/F1 only 0.656/0.558 on ADNI, showing tabular data alone is insufficient. Replacing CAPM with FiLM/HyperFusion degrades performance, validating covariate-guided modulation. CAPM is the main empirical contributor; MDM provides complementary regularization and interpretability.

R2Q1: Architecture MDM relies on perturbation-based sensitivity, not strict causal effect estimation. Despite CNN receptive field overlap, masking a patch selectively suppresses local activations; the resulting change in prediction still reflects the model’s reliance on that region, yielding a “blurred” but effective localization. This perturb-and-measure mechanism provides a useful regularization signal, as confirmed by our ablation results. We also tested a ViT backbone, which further improved performance, confirming the framework’s generality across architectures. We chose CNN for efficiency and fair baseline comparison; ViT ablation and discussion will be added to the manuscript.

R2Q4: Calculation of AUC Multi-class AUC uses one-vs-rest with macro averaging, giving equal weight to NC, MCI, AD to avoid majority-class bias.

R3Q1: Excessive Use of Causal Terminology We will replace “causal/non-causal branch” with “covariate-sensitive/non-sensitive branch” and rephrase causal claims as hypothesized dependencies.

R3Q2: Baseline Comparisons. Image-only baselines (CausalMixNet, etc.) assess visual representation robustness. Tabular-only (TabFormer) and multimodal models (HyperFusion) ensure fair comparison. Our three-category baseline design gives a comprehensive evaluation.




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 paper proposes a causality-inspired, confounder-aware multimodal framework (MDM + CAPM) for Alzheimer’s disease diagnosis. The reviews are mixed: R1 recommends weak accept, while R2 and R3 recommend weak reject. The ablation study shows that MDM’s independent contribution is limited and the main performance gain appears to come from CAPM, which creates a mismatch with the paper’s narrative emphasis and calls for a clearer, more balanced explanation; the validity of using a population-level voxel-wise mean image as a prior is also questioned as a meaningful supervisory signal. The authors are encouraged to focus their rebuttal on the attribution of each module’s contribution, as well as the specific concerns raised by the reviewers.

  • 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.

    During the rebuttal phase, the authors provided positive responses that resolved some detailed issues and addressed concerns regarding the prior score loss. R2 and R3 still have remaining concerns regarding the MDM CAPM. However, considering the comprehensive setup and interpretability, I lean towards acceptance.



Meta-review #2

  • 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 proposal is technically interesting but there is a large field of unbiased or confounder-aware learning that the authors did not consider in the comparison. The validation is mainly driven by the overall prediction accuracy, a metric that is irrelevant to the studied topic of confounder modeling. There is no evidence showing whether there are confounding effects in the prediction and whether those effects are captured and mitigated.



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.

    The rebuttal reasonably clarified the relative roles of CAPM and MDM, the weak regularizing nature of the prior loss, the AUC computation, and the intended softening of causal claims; although some architectural concerns remain, the consistent empirical gains and improved positioning support acceptance with these clarifications incorporated in the camera-ready version.



back to top