List of Papers Browse by Subject Areas Author List
Abstract
Medical Visual Question Answering (MedVQA) models often exhibit limited generalization due to reliance on dataset-specific correlations, such as recurring anatomical patterns or question-type regularities, rather than genuine diagnostic evidence. Existing causal approaches are typically implemented as static adjustments or post-hoc corrections and do not explicitly refine representations during training. To address this issue, we propose a Learnable Causal Trimming (LCT) framework that integrates causal pruning into end-to-end optimization. We introduce a Dynamic Anatomical Feature Bank (DAFB), updated via a momentum mechanism, to capture global prototypes of frequent anatomical and linguistic patterns, serving as an approximation of dataset-level regularities. We further design a differentiable trimming module that estimates the dependency between instance-level representations and the global feature bank. Features highly correlated with global prototypes are softly suppressed, while instance-specific evidence is emphasized. This learnable mechanism encourages the model to prioritize causal signals over spurious correlations adaptively. Experiments on VQA-RAD, SLAKE, SLAKE-CP and PathVQA demonstrate that LCT consistently improves robustness and generalization over existing debiasing strategies.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0307_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{XuZib_Learning_MICCAI2026,
author = { Xu, Zibo AND Li, Qiang AND Nie, Weizhi AND Su, Yuting},
title = { { Learning to Trim: End-to-End Causal Graph Pruning with Dynamic Anatomical Feature Banks for Medical VQA } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16878},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper addresses the problem of spurious correlations and poor out-of-distribution generalization in medical visual question answering (MedVQA). The authors propose a new framework called Learnable Causal Trimming (LCT), which integrates debiasing into end-to-end model training rather than relying on static causal correction or test-time adaptation.
The method has two main technical components. First, the paper introduces a Dynamic Anatomical Feature Bank (DAFB), updated through a momentum mechanism, to capture dataset-level recurring anatomical and multimodal patterns that may act as shortcut cues. Second, it proposes a Causal Trimming (CT) module that measures the similarity between instance-level visual features and the feature-bank prototypes, then applies a differentiable soft mask to suppress features that are highly correlated with these global priors while preserving more instance-specific evidence. An additional orthogonality loss is used to further reduce residual correlation between the trimmed features and the feature bank.
The main contribution of the paper is therefore an end-to-end trainable debiasing framework for MedVQA that is inspired by causal reasoning and designed to improve robustness under both standard and out-of-distribution settings. The method is evaluated on VQA-RAD, SLAKE, PathVQA, and SLAKE-CP, where it shows consistent improvements over prior work.
- 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 paper addresses an important and clinically relevant problem. Shortcut learning and poor robustness under distribution shift are major limitations of current MedVQA systems. The paper targets this issue directly, which makes the work timely and relevant to the MICCAI community.
2.The proposed framework is coherent and reasonably novel in how it integrates debiasing into training. The combination of a momentum-updated dynamic feature bank and a differentiable trimming mechanism is intuitive and well motivated. A key strength is that the method incorporates debiasing into end-to-end optimization instead of using static adjustment or only test-time adaptation.
3.The experimental evaluation is solid and supports the paper’s main claim. The method is evaluated on multiple MedVQA benchmarks, including the OOD benchmark SLAKE-CP, and shows consistent gains over prior work. The ablation study and qualitative visualizations further strengthen the empirical support.
- 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 causal interpretation is stronger than what is formally justified by the method. While the framework is motivated by causal reasoning, the actual implementation is closer to a confounder-aware feature suppression approach. The paper does not fully justify why the feature bank is a faithful proxy for a latent confounder or under what assumptions the trimming operation corresponds to causal adjustment.
2.The component-level novelty is moderate. Several ingredients of the method, such as momentum-updated prototype/memory-style representations, similarity-based masking, and decorrelation losses, are related to existing ideas. The main novelty lies in their integration for MedVQA debiasing, but the distinction from prior debiasing/causal MedVQA works could be made clearer.
3.Some additional analysis is still missing. The paper would be stronger with more discussion of hyperparameter sensitivity, computational overhead, and the possible risk of suppressing clinically meaningful but common anatomical structures together with shortcut features.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I lean toward weak acceptance because the paper addresses an important problem in MedVQA and presents a coherent method with encouraging empirical results. In particular, the focus on reducing shortcut learning and improving OOD robustness is highly relevant for medical AI, and the proposed framework is a meaningful step beyond static debiasing or test-time-only adjustment. The evaluation is also reasonably comprehensive for this domain, including multiple datasets, an OOD benchmark, ablations, and qualitative analysis.
My main reservations concern the strength of the causal claims and the fact that some method components are built on relatively standard ideas. I also think the paper would benefit from more analysis regarding efficiency and sensitivity to design choices. However, despite these limitations, the paper offers a useful and practically relevant contribution, and the OOD improvements make the work valuable for the MICCAI audience.
- 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 addresses the problem of spurious correlations and poor out-of-distribution generalization in Medical Visual Question Answering (MedVQA). The main contribution is a Learnable Causal Trimming (LCT) framework that incorporates causal-style debiasing directly into end-to-end training, rather than relying on static causal adjustment or test-time adaptation. Specifically, the paper introduces a Dynamic Anatomical Feature Bank (DAFB) to model dataset-level priors and a differentiable Causal Trimming (CT) module to suppress visual features that are highly correlated with such global prototypes, with the goal of encouraging the model to focus on instance-specific diagnostic evidence. The method is evaluated on VQA-RAD, SLAKE, PathVQA, and SLAKE-CP, where it shows consistently improved performance, especially under distribution shift.
- 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 paper is clearly written and easy to follow. 2.The overall method is conceptually appealing. 3.The experimental results are convincing overall.
- 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 method relies on region-level trimming, but the paper does not clearly explain how regions are defined in practice. It seems likely that patch-level features are used, but the exact setup should be clarified. 2.Since the paper argues that DAFB captures dataset-level confounders, it would be helpful to visualize which regions are most similar to the learned bank after training, and whether they indeed correspond to the expected shortcut-related patterns. 3.The ablation baseline differs from the PMC-CLIP result reported in Table 1.The paper should clarify whether this is due to re-implementation or different training settings. 4.The related work could be expanded. The paper [1] is relevant to representation alignment and robustness in MedVQA.
[1] Bridging the Semantic Gap in Medical Visual Question Answering With Prompt Learning
- 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 is clearly written, the method is easy to follow, and the experimental results are overall convincing, especially under OOD evaluation. The idea of integrating causal-style trimming into end-to-end MedVQA training is meaningful and appears effective in practice. My concerns are mainly about clarity and additional validation rather than fundamental flaws. In particular, the paper should better explain how regions are defined, provide stronger evidence for what DAFB actually captures, and clarify the baseline inconsistency. Overall, I believe the paper is above the acceptance threshold for the venue.
- 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
This paper proposes a spurious shortcut-mitigating MedVQA method centered on a learnable causal trimming framework. It suppresses regions highly correlated with dataset-level priors, while emphasizing instance-specific diagnostic evidence. Experiments validate its effectiveness in boosting both accuracy and out-of-distribution generalization.
- 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.
- End-to-end learnable causal trimming is new and beneficial to the literature. The momentum-updated DAFB provides a dynamic, annotation-free proxy for latent confounders, addressing the limitation of static confounder dictionaries in prior causal MedVQA works.
- Improved OOD accuracy and Grad-CAM quality.
- Compared to competitive prior works.
- The proposed trimming module can be packed into the ViT encoder and easily injected to VLM models.
- 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 contribution of L_{orth} remains unknown in the ablation studies.
- 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.
(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?
End-to-end learnable causal trimming is new and beneficial to the literature.
- 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
We thank the reviewers and the meta-reviewer for their positive feedback and constructive suggestions. We will carefully refine the camera-ready version in light of the reviewers’ valuable comments to further improve the clarity and completeness of the paper.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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 learnable causal trimming framework for Medical VQA, integrating a dynamic anatomical feature bank with a differentiable feature suppression mechanism to improve robustness under distribution shift.
The reviewers agree that the paper tackles an important and clinically relevant problem, and presents a coherent and practically meaningful end-to-end debiasing framework. The method is conceptually appealing, and the experimental results are solid and consistently support the claimed improvements. While modeling the confounder C as dataset-specific priors via the feature bank raises a potential concern of learning dataset-dependent bias that may limit generalization, the improved performance on OOD benchmarks (e.g., SLAKE-CP) partially alleviates this issue. However, validation under more diverse distribution shifts would further strengthen this claim. Meanwhile, the work could also be further strengthened by better justifying the causal interpretation of the framework (particularly the role of the feature bank as a proxy for confounders) and improving the methodological clarity of key components (e.g., region/feature definition and implementation details).