Abstract

Cross-view interference caused by impure shared information for multiview mammogram representation. Existing methods are accustomed to assuming that purely complementary shared information are provided between multiple views, ignoring the negative side of the shared information. To address this issue, we propose the first Dual-view Mammography Causal Graph (DMCG) to model multi-view representation by capturing direct and mediation effects. Based on DMCG, we propose MammoCRKAN, the first counterfactual reasoning paradigm integrating the Kolmogorov-Arnold theorem for decoupling interfering information. MammoCRKAN comprises two key modules: the Spherical Sample Module (SSM), which enhances the direct effect of tumor features by aligning consistent geometric representations, and the Kolmogorov–Arnold Aggregate Module (KAAM), which decomposes complex joint causality into univariate effects to mitigate negative side of mediation effects. Moreover, We find that heterogeneous channel allocations across views outperform fixed matching channels. Extensive experiments on four publicly available mammogram datasets demonstrate the effectiveness of MammoCRKAN. Code is available at \url{https://github.com/**}.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0579_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://guoli-w.github.io/MammoCRKAN

Link to the Dataset(s)

N/A

BibTex

@InProceedings{WanGuo_Rethinking_MICCAI2025,
        author = { Wang, Guoli and Wei, Benzheng and Li, Shuo},
        title = { { Rethinking Multi-view Mammogram Representation Learning via Counterfactual Reasoning with Kolmogorov-Arnold Theorem } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15967},
        month = {September},
        page = {427 -- 437}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper introduces MammoCRKAN, a new framework that tackles cross-view interference in dual-view mammography by reframing representation learning as a causal problem. First, it proposes the Dual-view Mammogram Causal Graph (DMCG), which decomposes the tumour’s influence on the final diagnosis into a helpful Direct Effect (DE) and a potentially harmful Mediation Effect (ME) carried by “impure” shared information (redundant tissue, density, artefacts) between the craniocaudal and mediolateral-oblique views. To maximise the Total Direct Effect (TDE) while suppressing the negative ME path, the authors design the MammoCRKAN counterfactual-reasoning pipeline. MammoCRKAN comprises two modules: (i) a Spherical Sample Module (SSM) that learns view-specific offset grids to project both views into a common spherical coordinate system, aligning the same tumour region geometrically and amplifying the direct tumour signal; and (ii) a Kolmogorov–Arnold Aggregate Module (KAAM) that uses a maximum-coverage patch selector plus K-A-based univariate mapping to decompose the high-dimensional mediation effect into patch-level contributions and down-weight those dominated by background, neutralising cross-view interference. Experiments on four public datasets (INBreast, VinDr-Mammo, CBIS-DDSM, CMMD) show that MammoCRKAN outperforms eight recent state-of-the-art methods ; ablations confirm that SSM, MCP, and KAAM each contribute, and hyper-parameter sweeps identify the optimal patch coverage and size.

  • 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’s key strengths are: 1) Novel causal formulation: the Dual-view Mammogram Causal Graph is the first to model CC and MLO views via distinct Direct and Mediation paths, overturning the common “all-features-are-complementary” assumption and exposing cross-view interference as a causal problem. 2) First counterfactual framework for mammography: MammoCRKAN applies causal counterfactual reasoning and embeds the Kolmogorov-Arnold theorem to separate helpful from harmful shared information. 3) Two original modules: the Spherical Sample Module introduces learnable spherical resampling to align tumour geometry across views, while the Kolmogorov-Arnold Aggregate Module uses a max-coverage patch selector plus univariate K-A networks to down-weight background patches, yielding an interpretable decomposition of mediation effects. 5) Strong evaluation on four public datasets: MammoCRKAN beats eight recent baselines on every metric, and ablations rigorously demonstrate each component’s value.

  • 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) Unproven benefit of the Kolmogorov–Arnold (K-A) layer: no experiment compares the spline-based aggregator to a vanilla MLP or attention pool, so the claimed theoretical advantage of invoking the K-A theorem remains unclear.

    2) High computational costs: SSM doubles feature maps and KAAM runs multiple extra backbone passes (one per patch), yet the paper omits GPU memory, preprocessing time, and throughput.

    3) Hyper-parameter sensitivity: performance depends on finely tuned patch size, cover-rate, and asymmetric channel splits, but no guidance is offered for unseen data.

    4) Clarity of the paper could be improved, some sentences do not hold up from a grammar point of view.

  • Please rate the clarity and organization of this paper

    Satisfactory

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

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

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

    (4) Weak Accept — could be accepted, dependent on rebuttal

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

    I would like to see the concerns expressed in the limitations addressed in the rebuttal.

  • Reviewer confidence

    Somewhat confident (2)

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

    N/A

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

    N/A



Review #2

  • Please describe the contribution of the paper

    This is a technical (method-heavy) paper that introduces MammoCRKAN, a framework for 2‑view mammogram learning that models and reduces cross‑view interference via a causal perspective. The claimed contributions include: 1)SSM that projects features into a spherical coordinate system and learns consistent geometric embeddings to strengthen DE.

    2) KAAM that leverages K-A theorem to decompose high‑dimensional mediation into univariate effects, selecting and aggregating key patches via a Maximum Coverage strategy to promote the positive side of ME while suppressing interference .

  • 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) Reviewer is able to validate that this is truly the first application of a causal graph in multi‑view mammography to explicitly separate beneficial tumor cues from harmful shared information 

    2) Non-cliche use of the K-A theorem, at first look at title people may think it is a follow-up to the trendy K-A networks but it is actually much better than that.

    3) Comprehensive empirical validation and ablation well included within page limit. I do not think any additional experiments are needed.

  • 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 counterfactual estimation of P₂ requires masking background b without knowing the exact tumor location. More details to prove that KAAM’s patch selection reliably targets true lesions are needed.

    2) While DMCG is a compelling abstraction, the authors should discuss identifiability conditions and potential confounders beyond B (e.g., breast density heterogeneity, imaging artifacts) that may violate the assumed causal structure.

    3) Shared vs Unique informaiton, for mammography, or more general medical imaging, has been a hot topic in recent medical AI field. No need to run more baselines, but a few lines in intro to compare with info-theory based methods could be helpful to clarify your strengths. Ive attached some links below.

    4)A brief intuitive explanation for readers unfamiliar with K-A theory would improve accessibility.

  • Please rate the clarity and organization of this paper

    Satisfactory

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

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    A brief intuitive explanation for readers unfamiliar with Kolmogorov–Arnold superposition theory would improve accessibility.

    Comparison with the related works (shared information) will be convincing:

    https://openreview.net/pdf?id=otHZ8JAIgh https://openreview.net/pdf?id=NJxCpMt0sf https://openaccess.thecvf.com/content/CVPR2023/papers/Wang_Multi-Modal_Learning_With_Missing_Modality_via_Shared-Specific_Feature_Modelling_CVPR_2023_paper.pdf

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

    (4) Weak Accept — could be accepted, dependent on rebuttal

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

    In general very novel & solid paper, I would raise score to strong accept after authors help me clarify major issues 1) 2) & 4).

  • 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 #3

  • Please describe the contribution of the paper

    This paper tries to address the multi-view differences when conducting mammogram segmentation. The authors have observed the view differences may not be directly learned by network and used a KAN to address this issue. The experiments demonstrated state-of-the-art performance by the novel method.

  • 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 very insightful and mathematically solid. The experiments are persuasive and the method has achieved SOTA performance.

  • 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 obvious weakness as I can see.

  • 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.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

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

    (5) Accept — should be accepted, independent of rebuttal

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

    The motivation, methodology and experiments are all arranged well and clear. It is quite insightful, and the similar idea could be transfered to other applications.

  • Reviewer confidence

    Somewhat confident (2)

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

    N/A

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

    N/A




Author Feedback

N/A




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    This paper presents a novel and interesting idea by modeling dual-view mammography as a causal inference problem. The proposed MammoCRKAN framework introduces a counterfactual reasoning pipeline, with a geometric alignment module and a Kolmogorov–Arnold-based aggregator playing key roles.

    However, as pointed out by Reviewer 2, there are important concerns that should be addressed. First, the actual benefit of the Kolmogorov–Arnold (K-A) module remains unproven, as no experiment compares it against standard alternatives such as MLPs or attention pooling. Second, the method introduces non-trivial computational overhead, but runtime and memory costs are not reported or discussed.

    The authors are encouraged to address these points clearly in the cam-ready version. Additionally, also it is curious how this theoretically driven approach compares empirically to works focusing on shared-specific feature disentanglement in multi-view or multi-modal learning, such as [a].

    [a] Wang H, Chen Y, Ma C, Avery J, Hull L, Carneiro G. Multi-modal learning with missing modality via shared-specific feature modelling. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.



back to top