Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Reflectance confocal microscopy (RCM) enables non-invasive in vivo imaging for basal cell carcinoma (BCC) diagnosis. Clinically, multiple RCM images are typically acquired per patient, while diagnostic labels are only available at the patient level, framing this task as a patient-level multiple instance learning (MIL) problem. Nevertheless, instance-level imaging artifacts, patient-specific macro-patterns, and modality-specific styles often act as confounders, introducing spurious correlations. We propose DualCaMIL, a dual-level causality-driven MIL framework that disentangles non-causal factors at both micro and macro scales. At the instance level, we introduce Cross-modal Microscopic Causal Disentanglement (CMCD), which leverages histopathology as a structural reference to isolate pathology-relevant content from imaging interference. At the patient level, we propose the Patient-aware Macro-level Causal Module (PMCM), which weakens shared non-causal structure among patient instances and prevents dependence on patient-specific contextual biases. Experiments on a public multicenter patient-level RCM dataset show that DualCaMIL outperforms representative protocol-compatible MIL baselines, with an AUC of 0.9447. Counterfactual stress testing further suggests robustness to non-causal factors, including imaging style variations and patient-specific constant patterns, while maintaining sensitivity to lesion-relevant signals. Code is available at https://github.com/Taffee9/DualCaMIL.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3710_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Taffee9/DualCaMIL

Link to the Dataset(s)

N/A

BibTex

@InProceedings{WanCha_DualCaMIL_MICCAI2026,
        author = { Wang, Changxin AND Chen, Lifang AND Zou, Yunmin},
        title = { { DualCaMIL: Dual-Level Causal Multi-instance Learning for Patient-Level Diagnosis in Reflectance Confocal Microscopy } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16895},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes DualCaMIL, a dual-level causality-driven MIL framework for basal cell carcinoma diagnosis from RCM images, which jointly addresses instance-level (e. g. , imaging style) and patient-level (e. g. , structural bias) confounders. It introduces a cross-modal disentanglement module guided by histopathology to isolate pathology-relevant features, and a patient-aware modeling module to suppress non-causal structural patterns during aggregation. Together with a counterfactual evaluation strategy, the method improves robustness and generalization without requiring additional annotations.

  • 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 to mitigate confounding factors is important for improving robustness and generalization in medical AI.
    • Leveraging histopathology images as a cross-modal reference to guide RCM representation learning is an interesting idea.
  • 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 is limited to basal cell carcinoma diagnosis using RCM. Although the proposed method appears broadly applicable, its generalizability to other tasks and datasets is not evaluated.
    • Relatedly, the effectiveness of the patient-level causal modeling module is only validated on a single task, and should be further assessed across diverse settings.
    • Most competing methods are relatively simple MIL baselines. Stronger comparisons with domain-invariant or causality-based learning approaches are needed.
    • A clearer conceptual comparison with existing domain-invariant learning methods would help justify the advantages of the proposed approach.
    • The paper uses several unclear or unconventional terms (e. g. , “patient instances”), which makes some technical details difficult to follow.
  • 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.

    (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?

    While the studied topic is important and the idea is somewhat interesting, the paper has several significant limitations, as discussed in Weaknesses.

  • Reviewer confidence

    Very confident (4)

  • [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 rebuttal provides useful clarifications and appropriately calibrates several claims, particularly regarding the causal interpretation and the limited scope of the current evaluation. However, my main concerns are not fully addressed.

    • The issue of generalizability remains largely unresolved. The authors acknowledge that the method is only validated on a single RCM BCC classification setting and that broader applicability is left for future work. While the architectural motivation may indeed be generic, the current evidence is still limited to one task/domain combination. In particular, the effectiveness of the patient-level causal modeling component is difficult to assess without experiments on additional datasets or patient-level MIL problems where confounding structures differ substantially.
    • The rebuttal does not sufficiently address the concern regarding comparisons with stronger baselines. The explanation that many invariant-learning methods require explicit environments or additional annotations is reasonable, but it still leaves the empirical comparison space relatively weak. The paper currently demonstrates improvements mainly over standard MIL variants, while the claimed contribution is closely related to domain-invariant and causality-inspired representation learning. As such, stronger or more carefully adapted baselines would still be important to establish the actual novelty and practical advantage of the proposed formulation.
    • The reproducibility concern is only partially addressed. The rebuttal acknowledges missing implementation details and indicates they will be added later, but at present the submission still lacks sufficient procedural detail for reliable reproduction, especially regarding pathology matching, patient bag construction, and training configurations.



Review #2

  • Please describe the contribution of the paper

    The paper proposes DualCaMIL, a dual-level causality-driven multiple instance learning framework for patient level basal cell carcinoma diagnosis from reflectance confocal microscopy. The method combines a micro level cross-modal disentanglement module guided by histopathology with a macro level patient-aware module that suppresses patient-shared structural confounders during aggregation. On the public RCM cohort, the method achieves the best reported performance among the listed baselines (AUC 0.9447), and the paper further presents ablations and counterfactual stress tests intended to assess robustness to style, patient-context, and lesion-level perturbations. This matters because patient-level RCM diagnosis is clinically relevant, weakly supervised, and especially vulnerable to spurious correlations.

  • 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 addresses a clinically meaningful and technically challenging setting of patient-level diagnosis from multiple RCM images with only patient-level labels.

    The proposed dual-level formulation is well motivated. The instance-level CMCD module and patient-level PMCM module target different confounding mechanisms rather than using a single generic invariance objective.

    The quantitative results are strong relative to the listed baselines. In Table 1, the method improves over TransMIL, MicroMIL, and IRM, and the ablation in Table 2 supports that both proposed components contribute.

    The paper includes patient-level counterfactual stress tests in addition to standard classification metrics which is a useful robustness analysis and strengthens the empirical story beyond reporting a single AUC.

    The clinical motivation and overall pipeline are clearly communicated and the paper makes a reasonable effort to connect the method design to domain-specific structural correspondence between RCM and histopathology.

  • 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.

    My main concern is that several methodological details necessary for confident interpretation and reproducibility are missing or insufficiently precise. The paper states that “label-matched pathology tiles provide teacher supervision for alignment,” but does not make clear whether the RCM pathology supervision is patient-matched, region-matched, or only class-matched. Given how central the cross-modal alignment is to the method, this ambiguity matters. Likewise the paper mentions orthogonality and reconstruction constraints in CMCD but the main text only writes out L_align and L_inv, leaving some key losses and their weights unspecified.

    A second important issue is that the causal framing is somewhat stronger than the evidence currently shown. The interventions in CMCD and the patient-subspace suppression in PMCM are plausible and useful design choices but the manuscript does not provide a formal causal model or an identifiability argument. This does not invalidate the method but it does suggest the claims should be framed more carefully as causally inspired robustness modeling rather than established causal identification.

    The private OOD evaluation is limited. The paper explicitly describes the private dataset as 86 BCC positive patients and uses it for OOD stress testing without fine-tuning. That makes the OOD robustness analysis directionally useful, but it does not support a standard out-of-distribution classification claim because no negative class performance is reported there.

    There appears to be a mismatch between the PMCM figure and the formal description. The text defines B_p through eigendecomposition of the patient covariance, while the figure visually suggests a context-encoder pathway. This should be reconciled.

    The manuscript does not report variability across runs, confidence intervals, or statistical testing, which makes it harder to judge the robustness of the reported improvements.

  • 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.

    (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 slightly positive because the paper tackles an important clinical ML problem with a method that is both plausible and reasonably novel for this setting. The dual-level treatment of confounding is well motivated while the public-cohort gains over the included baselines are meaningful and the ablation plus stress-test analyses provide supportive evidence that the method is doing something nontrivial. My reservations are mainly about methodological specification and strength of claims rather than a fatal flaw.

  • 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.

    The authors have clarified the concerns. The paper can be accepted based on the wording changes to the paper.



Review #3

  • Please describe the contribution of the paper

    DualCaMIL introduces a dual-level causality-driven multi-instance learning framework for reflectance confocal microscopy diagnosis that effectively eliminates spurious correlations by addressing confounders at both micro and macro scales. The approach utilizes a cross-modal microscopic causal disentanglement module to isolate pathology-relevant content using histopathology as a structural reference, while a patient-aware macro-level causal module identifies and suppresses non-causal constant structures shared across a patient’s instances. The experiments demonstrate that this framework significantly enhances diagnostic accuracy and robustness against imaging artifacts and patient-specific biases compared to existing MIL baselines.

  • 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.

    DualCaMIL addresses spurious correlations in RCM diagnosis via a dual-level causal MIL framework that utilizes cross-modal microscopic disentanglement with histopathology references to eliminate imaging artifacts and a patient-aware macro-causal module to suppress individual background biases. The method demonstrates superior robustness and clinical relevance.

  • 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.

    Minor: The text in the figures is too small and cannot be read clearly when printed. Major: Figure 1(a) can be understood, but the clustering trend in Figure 1(b) is not obvious. Please provide a clearer description. The paper mentions using label-matched histopathology patches for alignment. If the selected patches are of poor quality, or not from the same patient (although from the same disease type), how much impact would this have on the model’s disentanglement performance? Table 1 needs to evaluate results on three datasets; performance on a single dataset is insufficient to demonstrate effectiveness. Additionally, some other latest non-MIL testing methods need to be added, especially methods that have been validated on publicly available datasets.

  • 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?

    please refer to the weaknesses.

  • 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 reviewers for recognizing DualCaMIL’s importance and promising results, and address the main concerns below. (i) Generalizability/datasets. DualCaMIL uses no BCC-specific handcrafted cues. It targets a broader patient-level MIL pattern: multiple images per patient, patient-level labels, style variation, and patient-shared background. This suggests broader applicability, but evidence validates patient-level RCM BCC diagnosis; we will narrow claims. One main classification dataset reflects task validity/public-data availability: valid testing needs RCM bags with patient-level positive/negative labels. Public RCM datasets matching this protocol are limited, and Monnier provides the target protocol. SkinCancer histopathology is not a second RCM test set; it provides cross-modal supervision motivated by RCM-H&E correspondence. The private RCM cohort is BCC-positive only; as R2 noted, it is not standard OOD classification, but no-finetuning stress testing under RCM domain shift. We acknowledge this limitation; broader patient-level MIL tasks and stronger compatible baselines are future work. (ii) Baselines/comparison. IRM seeks environment-level invariance, whereas patient-level RCM MIL involves two-level confounders: RCM style/artifacts affect individual instances, while site, pigmentation, scanning conditions, and tissue context can be shared within a patient bag. DualCaMIL addresses these separately through CMCD and PMCM, which differs from generic invariant learning. Many domain-invariant/causal methods require explicit environments, multi-environment training, or image/lesion labels unavailable in our setting. Preliminary experiments with several such methods also did not provide stable gains under this protocol. Thus Table 1 reports IRM as a compatible invariant-learning baseline and focuses on protocol-consistent MIL methods. (iii) Method clarity. “Patient instances” means multiple RCM images/patches forming one patient bag. Patient-level structural bias means stable same-patient background, e.g., site, texture, pigmentation, or tissue context, which naive MIL may mistake for evidence. For R2/R3’s pathology question, pathology supervision is class/semantic matched, not patient- or region-registered. CMCD uses pathology as a structural teacher, not paired registration; non-same-patient tiles need no spatial correspondence. Low-quality or semantically inconsistent tiles may weaken the teacher signal; semantic matching plus reconstruction, orthogonality, and intervention consistency mitigate, but do not eliminate, this risk. For R2’s PMCM figure concern, PMCM is defined in equations/text by eigendecomposition of the patient covariance matrix. Fig.2(c) only schematizes patient-context estimation/suppression, not an extra context encoder. We will clarify terminology, pathology matching, PMCM caption, and loss weights without changing the method. (iv) Claim calibration/causal interpretation. Evidence should not be read as formal causal identification. We retain the causality-driven framing because the method is motivated by intervention and counterfactual robustness, but will revise wording implying causal proof or identifiability. The supported claim is that DualCaMIL achieves mechanism-consistent robustness in ablation and PCST while preserving lesion sensitivity; it is not a formal causal theorem. (v) Reproducibility/statistics/presentation. We agree with R1/R2 that procedural details should be clearer. We will specify pathology selection/semantic matching, patient bag construction, PMCM k, CMCD losses/weights, training/inference flow, and PCST. Code will be released upon acceptance. Results are consistent across metrics/ablations/PCST; we will further explore repeated-run variability and statistical testing. For R3, Fig.1(b) is qualitative: each point is one patient-level embedding; dispersion shows naive pooling can amplify patient-specific background over lesion evidence. We will enlarge text and revise the caption.



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.

    There is broad agreement among the reviewers that the work addresses an important and relevant problem, and that the proposed solution is well-motivated and potentially useful with promising results. However, reviewers have highlighted issues with the methodology clarity, limitation to evaluating on a single dataset and unjustified interpretations. In rebuttal, the authors should (i) comment on the works generalizability to other tasks, include experimental results if they exist already; (ii) extend the comparison to more relevant baselines or justify limiting to the existing comparisons presented; (iii) similarly, include comparisons with other datasets if available or justify limiting to the single dataset used; (iv) clarify the methodology, including terms used; and (v) review the claims made to ensure that they are supported by the included results or provide additional justification.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    N/A

  • Please justify your recommendation.

    N/A



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.

    After rebuttal, this paper receives two accept recommendations and one rejection. The AC acknowledges that DualCaMIL addresses an important patient-level RCM diagnosis problem and proposes a well-motivated dual-level framework for mitigating instance- and patient-level confounding. Reviewers recognized the clinical relevance, promising empirical performance, and useful ablation and counterfactual stress-test analyses. The rebuttal clarifies several issues, including the scope of generalizability, the use of class/semantic-matched pathology supervision, the PMCM formulation, baseline selection, and the need to calibrate causal claims. However, substantial concerns remain regarding validation on only a single standard classification dataset, limited comparison with stronger causality/domain-invariant baselines, incomplete reproducibility details, and insufficient evidence for broader generalizability. Given that these limitations affect the strength of the main claims and cannot be fully resolved by presentation changes, the AC recommends rejection.



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 paper proposes DualCaMIL, a dual-level causality-driven MIL framework for RCM-based BCC diagnosis, addressing both instance-level and patient-level confounders. Reviewers find the method well-motivated, clinically relevant, and empirically strong against baselines, with meaningful ablations and counterfactual stress tests. While Reviewer #1 raises concerns about generalizability and baseline comparisons, Reviewers #2 and #3 see the contributions as solid and the rebuttal sufficiently addresses key issues. Recommendation: Accept.



back to top