Abstract

Child-Pugh (CP) grading, a widely used clinical standard for therapeutic decision-making in hepatocellular carcinoma, integrates biochemical and clinical markers to assess hepatic reserve. Estimating CP grade from routinely acquired contrast-enhanced CT (CECT) may provide opportunistic, complementary imaging evidence and facilitate investigation of associations between structural imaging phenotypes and clinically assessed hepatic reserve. However, inferring CP-related functional status from structural imaging requires overcoming a fundamental structure-function semantic gap. Because CP grade is determined by multiple biochemical and clinical factors, its imaging manifestations are heterogeneous and deeply entangled. Conventional Multiple Instance Learning (MIL) architectures do not explicitly decouple these complex representations, making them susceptible to associative shortcut learning rather than capturing stage-relevant pathological patterns. To explicitly model this structure-function mapping, we propose SLICE-MIL, a spatial-logical framework for ordinal evidence disentanglement. Structurally, an anatomical mask-guided mechanism explicitly decouples the entangled intra- and extra-hepatic features. Functionally, a hierarchical evidence disentanglement module employs representation-level counterfactual ablations to anchor stage-specific semantics, encouraging the decoupled representations to align with the progression logic of CP deterioration. Multicenter validation demonstrates that explicitly bridging this semantic gap mitigates reliance on spurious correlations, substantially improving cross-domain robustness and interpretability.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2557_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/cutecattutu/SLICE-MIL

Link to the Dataset(s)

N/A

BibTex

@InProceedings{DenYih_SLICEMIL_MICCAI2026,
        author = { Deng, Yihan AND Huang, Hongwei AND Peng, Yunsong AND He, Lei AND Cui, Wenju AND Leng, Yilin AND Wang, Shuai AND Cheng, Qi AND Zheng, Jian},
        title = { { SLICE-MIL: Counterfactual Semantic Anchoring of Disentangled Evidence for Child-Pugh Grading } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16887},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    SLICE‑MIL addresses the structure‑function semantic gap and feature entanglement issues that exist in assessing Child‑Pugh grade from CT images. Through anatomy‑guided spatial decoupling, orthogonal decomposition, residual routing, and counterfactual semantic anchoring, it achieves explicit decoupling and causal regularization of ordered pathological evidence, thereby improving cross‑domain generalization and interpretability.

  • 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.Separate intra/extra-hepatic features via anatomy-gated queries to reduce cross-regional confounding. 2.Orthogonal decomposition, residual routing, and counterfactual semantic anchoring achieve causal regularization of ordered pathological evidence

  • 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.anatomical disentanglement only separate the lesion and liver parenchyma, maybe more detailed anatomical disentanglement (liver segements, portal vein and so on). 2.It’s hard to understand the Logical Dependency Graph. what is difference compare with SCM? 3.Through counterfactual intervention, the model transform the CP classifier to a binary classify task. What’s more, the three manually defined classification rules directly hard-coded clinical priors into the loss, rather than the true causal relationships learned by the model.

  • Please rate the clarity and organization of this paper

    Satisfactory

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

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

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

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

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

    Despite some methodological issues, the overall approach is sound. The implemention of DO operation method is novel, and the experiments are thorough.

  • 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.

    N/A



Review #2

  • Please describe the contribution of the paper

    The authors propose SLICE-MIL, a spatial-logical framework based on Multiple Instance Learning for the assessment of Child-Pugh grading using contrast-enhanced CT scans. The method introduces an Anatomy-Driven Dual-Query Attention for spatial disentanglement. Overall the method appears novel, but I have some major comments.

  • 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 authors framed the clinical challenge as a “tructure-function semantic gap and try to solve the issue of shortcut learning prevalent in standard MIL architectures. (2) The proposed ADDA used clinical priors to separate feature extraction, preventing the early cross-contamination of visual signals. They also applied counterfactual masking to guide the model, which is novel.

  • 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 authors justify the use of an MIL architecture by stating that 2D CNNs suffer from label noise and 3D CNNs face a strict resolution-context trade-off. However, these claims are not empirically validated in the manuscript. In Table 1, the proposed SLICE-MIL is exclusively compared against other generic MIL frameworks. To convincingly demonstrate the necessity of this complex MIL approach, the authors must include standard non-MIL baselines, such as a 3D ResNet or a 2.5D CNN taking the abdominal/liver region (or dual-stream inputs) as data. Without these, the superiority of the MIL paradigm for this specific task remains unproven. (2) In the abstract, the authors claim that deriving the CP grade directly from CECT enables “opportunistic, non-invasive profiling of hepatic reserve.” I think this claim is overstated and misrepresents the clinical reality of the Child-Pugh score. The CP grade is a multi-source metric heavily reliant on biochemical blood biomarker, and neurological assessment. CECT can only capture morphological changes. Predicting systemic biochemical functional deficits solely from morphological features relies on correlation rather than direct measurement. The clinical utility claims should be toned down, acknowledging the fundamental information gap between imaging morphology and blood-based biomarkers. (3) The propsoed spatial disentanglement module heavily relies on the upstream extraction of binary liver and extrahepatic masks, presumably generated by TotalSegmentator. In my experiments of such models, I found in real-world clinical scenarios involving advanced cirrhosis or massive hepatocellular carcinoma (conditions frequently seen in patients assessed with the CP score), the liver and spleen are often severely deformed, leading to high failure or inaccuracy rates in automated segmentation tools. The authors should address how sensitive the downstream orthogonal decomposition and routing mechanisms are to upstream mask inaccuracies. An ablation study or a robustness analysis is necessary. (4) The authors use terms like do-calculus, causal grounding, and counterfactual interventions to describe the proposed semantic anchoring. In my opionin, they were rfeature masking combined with label modification. While this is a highly effective and good regularization technique, it may not constitute strong causal inference. Could the authors clarify the terminology to better reflect the mathematical nature of the operations, in order to avoid over-claiming causal discoveries? (5) Minor points: In Section 2.1, the definition of Q assumes the masks are quasi-disjoint. A brief statistical summary would strengthen this assumption; Please clarify the computational overhead introduced by the QR decomposition and dynamic routing; In Fig.1 their is Anatomy-Driven Differential Attention, but it should be Anatomy-Driven Dual-Query Attention; ‘Consequently, Current’, C should be c.

  • Please rate the clarity and organization of this paper

    Good

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

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

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

    (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?

    I keep open in raising my score if the concerns are explained.

  • 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.

    Thank you for the authors’ response. I appreciate that the authors have partially addressed my concerns; however, several issues remained.

    I agree that patient-level Child-Pugh labels fit a weakly supervised MIL formulation. I also acknowledge the practical challenges of training 3D models under heterogeneous CT acquisition protocols. However, I do not think this fully resolves the concern regarding missing non-MIL baselines. Standard 2.5D or 3D models can still be trained with patient-level supervision, for example through scan-level aggregation, liver/abdomen cropping, or volume-level classification. I understand that the MICCAI rebuttal policy does not allow additional experiments at this stage, but similar liver fibrosis staging studies have included broader baseline comparisons, including single-view, whole-image, and multi-view methods [1,2]. Therefore, the current rebuttal justifies the motivation for using MIL, but does not empirically establish that MIL is superior to reasonable non-MIL alternatives.

    Regarding segmentation robustness, the response is partially convincing but still limited. The use of downsampled masks may reduce sensitivity to minor boundary errors, but the rebuttal does not provide direct evidence of robustness against systematic segmentation failures. This limitation should be explicitly acknowledged in the final manuscript, preferably in the discussion or conclusion.

    [1] Gao, Z., Liu, Y., Wu, F., Shi, N., Shi, Y., & Zhuang, X. (2023). A reliable and interpretable framework of multi-view learning for liver fibrosis staging. In International Conference on Medical Image Computing and Computer-Assisted Intervention (pp. 178–188). [2] Liu, Y., Gao, Z., Shi, N., Wu, F., Shi, Y., Chen, Q., & Zhuang, X. (2025). MERIT: Multi-view evidential learning for reliable and interpretable liver fibrosis staging. Medical Image Analysis, 102, 103507.



Review #3

  • Please describe the contribution of the paper

    This paper proposes a Slice-MIL framework for predicting Child-Pugh grade from contrast-enhanced CT scans, aiming to enable non-invasive estimation of hepatic functional reserve in patients with hepatocellular carcinoma. The authors highlight that Child-Pugh scoring is clinically important for treatment planning but is traditionally based on both biochemical and clinical parameters, making direct imaging-based prediction challenging. To address this, the method treats each CT scan as a bag of slices and applies a multiple-instance learning framework with anatomy-guided feature disentanglement and counterfactual reasoning. The proposed approach separates intrahepatic and extrahepatic features using anatomical masks and further applies mechanisms intended to reduce shortcut learning and enforce more clinically meaningful representations. The model is evaluated on a multicenter dataset with internal and external validation, showing improved performance and robustness compared to baseline approaches.

  • 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 study addresses a clinically important problem, as Child-Pugh scoring is central to management decisions in hepatocellular carcinoma.

    The attempt to derive systemic hepatic function from CT imaging is innovative and clinically relevant, particularly for opportunistic assessment.

    Use of a multicenter dataset with external validation improves the strength of evidence for generalizability.

    The anatomical masking and slice-level learning approach is consistent with clinical interpretation, where radiologists assess both parenchymal changes and signs of portal hypertension.

    The focus on reducing shortcut learning and improving robustness is appropriate for real-world deployment scenarios.

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

    The study lacks clear clinical validation of predictions against standard Child-Pugh scoring components, particularly biochemical and clinical parameters.

    It is not fully clear how well the model performs in clinically ambiguous or borderline cases, where grading is subjective.

    The reliance on CT-only inference may limit applicability, since Child-Pugh grading is inherently a multimodal clinical construct.

  • 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

    How does the model perform in borderline or mixed presentations, where clinical scoring is less definitive?

    Can the model distinguish between hepatic parenchymal dysfunction and extrahepatic causes of ascites or portal hypertension?

    How robust is the method to variations in CT acquisition protocols across centers?

  • 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 paper addresses an important clinical problem of predicting Child-Pugh grade from CT scans, which has strong relevance for treatment planning in liver cancer patients. The proposed Slice-MIL framework is well motivated and aligns with clinical reasoning by separating liver-related and extrahepatic features and using anatomical guidance for prediction. The use of a multicenter dataset with external validation adds strength to the results and suggests good generalizability. Overall, the work is clinically meaningful, methodologically sound, and shows clear improvement over baseline approaches, making it suitable for acceptance.

  • 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

Thanks for the valuable feedback. We are encouraged by the recognition of the clinical relevance(R3), methodological novelty(R1,R2)and thorough evaluation(R1,R3).Below we address the main concerns. Q1(MR,R2):Rationale for MIL We agree that including systematic 2D/2.5D/3D baselines could make the argument more complete, but we didn’t include because they are not suitable to this task: 1.2D/2.5D models lack slice-level gold-standard labels; individual slices often contain only partial evidence. 2.3D models face computational trade-offs and are sensitive to variations in slice thickness, spacing, and sequence length, making them less suitable for CT scans with heterogeneous protocols. In early exploration, we ran systematic 2D/2.5D/3D experiments. The results confirmed these principles. MIL performs weakly supervised learning at the patient level, matching CP grading’s multi-source nature and enabling interpretable slice aggregation. Additionally, our proposed SLICE-MIL is lightweight, with only 1.26M additional parameters (+4.5%) and 0.087G extra FLOPs (+1.9%).); QR decomposition adds no inference overhead once the projection is fixed; dynamic routing uses linear-complexity gating. Q2(MR,R1,R2):Causal claims and Logical Dependency Graph(LDG) R1 referred to Structural Causal Model (SCM), which is a mathematical framework that represents causal relationships between variables using structural equations and a directed acyclic graph (DAG). SCM assumes that the true variables are observable with clearly defined causal paths, but for Child-Pugh grading these variables cannot be directly observed from imaging, making it impossible to define explicit causal paths or structural equations. The LDG in SLICE-MIL is a visualization of the latent embedding’s logical disentanglement, not a DAG. Instead, we construct a latent-space representation to stage-specific CP increments, using counterfactual ablations to functional deterioration logic, maintaining three-class classification. Due to the multi-source nature of CP assessment, the model would struggle to learn the progression logic directly from CT-only data without these constraints. R2 mentioned “do-calculus” and “counterfactual interventions,” which in our method correspond to latent-space ablations that guide stage-specific progression. This implements a form of weak causal inference, regularizing stage-specific progression and reducing shortcut learning, rather than performing strong causal inference on the specific clinical variables. We treat these rules as weak causal priors with clinical guidance, rather than strict causal discoveries. This design aims to help bridge the structure–function semantic gap in CP grading. We apologize for the Fig.1 labeling error-“Causal grounding”,it will be corrected to “semantic anchoring”, consistent with the Fig.1 caption and the manuscript. Q3(MR,R1,R2):Segmentation robustness Our original intention is to design the model without relying on fine anatomical structures. Simply distinguishing body regions and liver parenchyma is sufficient. TotalSegmentator achieves a Dice score of ~0.94 for liver segmentation, and minor boundary errors in malignant cirrhosis or small tumors do not affect downstream processing, as the model incorporates all regions. We designed ADDA uses the masks only as regional guidance at a downsampled 7×7 resolution without pixel-level constraints. Experiments show downstream modules performed independently. Q4(R2,R3):Clinical scope As R3 noted, CP grading indeed relies on multiple modalities, but our work is limited to exploring CECT features related to CP, not as a substitute for standard assessment. We thank R3 for highlighting borderline cases, CP components, and etiology, and will include these clinical considerations in the discussion of the final manuscript. Further experimental exploration will be addressed in future work. The final manuscript will incorporate these revisions.The code will be released upon acceptance.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    The three reviewers gave two weak reject recommendations and one accept. Given the split opinions and some rather valid concerns raised, the paper would benefit from a rebuttal (clarify the missing baseline comparisons, rephrase the causal claims, and address segmentation robustness).

  • 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.

    One reviewer has increased their scores and one reviewer has lowered their score. My impression from the paper also remains somewhat mixed. Given the importance of the clinical application, I will still recommend an accept given the authors pay close attention to R2 post-rebuttal comment and implement the requested clarifications in the camera-ready (if ultimately accepted).



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.

    The decision is to accept the paper. It addresses a clinically relevant problem by proposing a novel framework to map structural imaging to functional clinical scores. In the rebuttal, the authors effectively justified their architectural choices, clarified their terminology, and resolved concerns regarding segmentation robustness.



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 rebuttal, the reviewer opinions lean positive but remain mixed. Reviewer #1 changed to Accept, and Reviewer #3 already recommended acceptance, emphasizing the clinical relevance, multicenter evaluation, external validation, and reasonable anatomy-guided MIL design. However, Reviewer #2 maintains Reject because the rebuttal does not fully resolve the lack of non-MIL baselines, such as 2.5D/3D or scan-level aggregation models trained with patient-level supervision, and provides limited direct evidence for robustness against systematic segmentation failures.

    Overall, the paper is clinically meaningful and methodologically interesting, but still has unresolved limitations in baseline completeness, segmentation robustness, and the clinical scope of CT-only Child-Pugh prediction. I therefore recommend acceptance, provided that these limitations and overclaims are clearly addressed in the final version. The authors should provide the full open-source implementation to improve the reproducibility and broader impact of the paper.



back to top