List of Papers Browse by Subject Areas Author List
Abstract
Pathological staging is critical for cancer diagnosis and treatment. However, extracting interpretable features from high-resolution whole slide images (WSIs) for accurate staging remains challenging. Existing methods divide WSIs into numerous patches but face two fundamental issues: lack of fine-grained stage labels and mixed-stage morphology within a single patch. Simply assigning a single pseudo-label discards heterogeneity and leads to decisions lacking medical interpretability. To address this, we propose DGFMIL, a dual-granularity fuzzy multiple instance learning framework centered on uncertainty modeling and semantic prior alignment. First, we introduce a differentiable interval type-2 fuzzification function that explicitly assigns a possibility interval rather than a single stage label to capture intrinsic heterogeneous staging characteristics. Second, cross-granularity calibration and fuzzy-guided cross-attention serve as auxiliary mechanisms to refine the representation and fusion of fuzzy features. Third, we construct a pathological concept text bank and impose semantic prior constraints on each patch via a routing soft alignment, enforcing visual representations to align with interpretable medical concepts. Extensive experiments on TCGA-CRC and Camelyon17 datasets demonstrate that DGFMIL achieves strong competitive performance.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1127_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{CaiMei_DGFMIL_MICCAI2026,
author = { Cai, Meiling AND Zhao, Juanjuan AND Yue, Wensong AND Zhao, Lin AND Zeng, Yue AND Zhang, Lijing AND Qiang, Yan},
title = { { DGFMIL: Fuzzy Multi-instance Learning with Cross-Granularity Calibration and Soft Semantic Routing for Histopathological Image Staging } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16891},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper primarily tackles the issue of mixed-stage morphology patches in whole-slide image (WSI) analysis. Common approaches, using multi-instance learning (MIL), treat patches as single-labeled, which might not hold. The authors propose to use type-2 fuzzification to assign a probability interval to the patches to capture heterogeneous staging characteristics.
The authors also implement a pathological concept text bank to impose semantic constraints via routing soft alignment. This improves interpretability by aligning each patch representation to a medical concept.
With the two stages, the authors aim to unify uncertainty quantification and semantic interpretation. However, the core contribution of this paper is the type-2 fuzzy membership assignment for patch representation.
- 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 itself is well-structured, and the reader can easily follow.
- The core idea sounds. It tackles an important problem in patch-level representation, related to the heterogeneity of the tissue morphology.
- The implementation of type-2 fuzzification for patch-level representation is novel.
- The evaluation performance seems convincing.
- 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.
Major weaknesses of the paper 1.My major concern is the complexity of the framework. It contains many components (i. e. , Membership Prior Attention Module, type-2 fuzzification, Fuzzy-Guide Cross Attention, Granularity Alignment Module, and Vision-language Alignment Matrix). Although the authors provided an ablation study (leave-one-out style), we cannot tell which combination of components is really important. 2.The authors claimed that their method achieves improvement, particularly in cases with high morphological ambiguity, but this statement is not proven nor implied. 3.One of the key hypotheses is that type-2 fuzzy sets would help to improve uncertainty quantification; however, this is never evaluated. For example, with the Expected Calibration Error or “Uncertainty vs. Error” correlation. 4.In Subsection 2.2., how are $m_k$, $\sigma_k$, and $\delta_k$ calculated?
Minor weaknesses
- The authors did not reference any paper on Type-2 Fuzzy Logic or the Karnik-Mendel algorithm (e. g. , Mendel, Jerry M. , Robert I. John, and Feilong Liu. “Interval type-2 fuzzy logic systems made simple. “ IEEE transactions on fuzzy systems 14.6 (2006): 808-821.).
- References to the datasets are missing.
- No code is provided.
- 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
- In the first sentence of the introduction, please leave a space between “histopathology” and the reference (histopathology\cite{. . . } to histopathology \cite{. . . }).
- Measuring uncertainty will greatly improve this work.
- 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?
There are some concerns about the methodology of this work, but I think the key ideas are novel and beneficial for digital pathology. But I wouldn’t mind if the paper is rejected.
- 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.
After reading the rebuttal, I maintain my “weak accept” recommendation. The rebuttal addressed several of my concerns:
- The hyperparameters are confirmed as learnable.
- The overall loss function was explained.
- The interaction between the modules was also clarified.
- Moreover, the authors promised to release the code and add the missing references.
However, two concerns remain partially unresolved:
- The claim that the model provides meaningful uncertainty estimates is still not experimentally validated.
- The authors’ rebuttal reframes IT2FS as not a general-purpose uncertainty estimation method, which is somewhat confusing, because the paper’s own stated contribution is “unified uncertainty quantification.”
Despite all these, I believe the core contribution is novel and addresses a real problem in digital pathology. The performance results are convincing, and the ablation study provides reasonable evidence for the framework design.
Review #2
- Please describe the contribution of the paper
This paper proposes a fuzzy MIL model with cross-granularity calibration and soft semantic routing for WSI disease staging. The framework consists of two novel synergistic pathways: one for fuzzy multi-granularity feature modeling and another for incorporating text-based semantic concepts. The design aims to capture boundary uncertainty, preserve global–local context, and improve interpretability. Experiments on two datasets show competitive performance against multiple 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.
a novel way to integrate both the uncertainty and interpretability into MIL model. compared with multiple baselines performed comprehensive ablation studies
- 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 Methods section lacks clarity and key details: for example, the interaction of the Semantic Prior Path with the main model is unclear, and the overall loss function is not specified
important experimental details are missing: e.g., Choice of K interval, Construction of the text concept library Selection of temperature in Eq. (3), Baseline implementation settings
Some claims are not supported by the experiment results. The paper claims that “While outputting high-precision staging results, DGFMIL provides uncertainty estimates and semantic explanations for each patch, significantly enhancing the model’s decision-making reliability on borderline cases.” However, in the experiment, only classification performance is evaluated, and there is no evidence provided for: uncertainty estimation, interpretability, robustness on borderline cases
There are some inconsistencies in Figure 1, e.g. the caption mentions three paths, while the figure and Section 2.1 describe only two; the figure is not fully consistent with the method description
- 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?
I think this paper introduces a novel way to integrate the uncertainty and the interpretability into the MIL model, which is good. However, essential details are missing from the methods and experiment sections. Additionally, the experimental results focus solely on classification performance, which does not fully support the authors’ claim about the model’s benefits.
- 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 studies pathological staging from whole slide images (WSIs) and proposes DGFMIL, a framework designed to address patch-level uncertainty, cross-scale information fusion, and the alignment of model decisions with pathological semantic concepts. The method is evaluated on the TCGA-CRC and Camelyon17 datasets.
- 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 idea of introducing fuzzy logic into MIL is innovative. Conventional MIL methods typically treat patches as instances with implicitly deterministic labels, whereas this work models the stage assignment of each patch as a membership interval. This formulation is more consistent with the reality of pathology, where stage boundaries are often inherently ambiguous. 2.The method outperforms baseline approaches on both datasets.
- 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.Although Table 2 provides some ablation evidence, the analysis remains relatively shallow given the overall complexity of the framework. The paper introduces five major components. However, each component is described rather briefly, without sufficiently in-depth analysis or validation, making it difficult to assess whether such a high level of complexity is truly necessary. 2.The evaluation of interpretability is relatively weak. The paper presents “establishing a unified paradigm for uncertainty quantification and semantic interpretation” as one of its main contributions, yet the experimental section only provides, in Figure 2(a), the attention map and key patches for a single random sample. A more systematic and quantitative evaluation of interpretability is lacking.
- 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
1.The paper uses “DGFMIL” in the title, whereas the caption of Figure 1 refers to “DFGMIL.” This inconsistency should be corrected. 2.Reference [18] appears to have formatting issues and also seems to duplicate Reference [15]. 3.I would encourage the authors to further clarify how the pathological concept bank was constructed, including the complete list of concepts, as well as the rationale and sources used for concept selection.
- 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 introduction of IT2FS into MIL for uncertainty modeling constitutes a certain degree of novelty. However, the paper has several important weaknesses at the execution level. In terms of model design, the framework is overly complex, yet the analysis of each component is not sufficiently deep. In the experimental validation, one of the paper’s central claims, interpretability, is not systematically evaluated. In addition, the absence of key implementation details affects reproducibility, and the manuscript also contains writing and formatting errors.
- Reviewer confidence
Somewhat confident (2)
- [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
Author Feedback
We appreciate positive feedback from reviewers (R), we address reviewers’ concerns about the weaknesses(W). Q1: Module interactions and overall loss (R1W1, R2W1) A1: Since only WSI-level stage labels are typically available, individual unannotated patches may exhibit features spanning multiple stages. Therefore, in addition to the multi-granularity image features processed by Granularity Alignment Module (GAM), we perform additional fuzzification on patches and enhance them through Membership Prior Attention Mechanism (MPAM). Fuzzy-Guided Cross-Attention Module (FGCA) then fuses the multi-granularity image features with the fuzzy features to produce the staging result. Additionally, Path2 (sematic prior) imposes supervisory constraints on Path1 through semantic priors from the medical concept bank. Overall loss is: L_total = L_cls + λ_f L_fuzzy + λ_a L_align + λ_s L_sparse, L_cls is the cross-entropy loss, L_fuzzy is membership regularization. Q2: Module Combination Analysis (R1W1, R3W1) A2: We conducted module combination analysis. Our method consists of two paths, with three modules in Path1.Therefore, we performed ablation studies on both the two paths and the three modules. In Table 2, “w/o Text” indicates the ablation of Path2, leaving only the complete pure image path; “w/o fuzzy” indicates the ablation of fuzzy logic in Path1, leaving a deterministic image path with text constraints. The results show that as long as Path1 is complete, even without Path2, the performance reaches 0.608 (Acc); however, any internal breakage in Path1 leads to >10% loss. This proves that the complete combination of Path1 is the necessary foundation, Path2 is an enhancement, and the combination of both achieves the optimum. Q3: Medical concept bank (R2W1, R3) A3: The concept bank was constructed in accordance with the WHO histological staging criteria and in conjunction with pathology reports. It provides three microscopic features for each stage, presenting representative characteristics in terms of tissue structure, cellular morphology, and degree of invasion. Q4: Fuzzy parameter, Selection of K and τ, Baseline Settings (R1W4, R2W1) A4: In Equation (1), m_k is the center of the k-th fuzzy set, δ_k is the standard deviation, and σ_k is the uncertainty footprint parameter, all are learnable parameters. Hyperparameters are determined via grid search on the validation set with K∈{3,5,7,10,15}. K=5 achieves the optimal balance between performance and computational cost. Ablation experiments validate τ∈{0.01,0.03,0.07,0.1,0.2}, and τ=0.07 yields the sharpest concept discrimination while maintaining multi-concept routing capability. All baseline methods adopt their official implementations. Q5: Uncertainty Estimation, Interpretability, and Robustness in borderline cases (R1W2, R2W2, R3W2) A5: Interpretability is intrinsic rather than post-hoc. Under Path2’s semantic constraints, image patches are associated with medical concepts, and the model outputs attention-weighted lesion regions as a byproduct (Fig. 2(a)), demonstrating localized focus on pathological areas for staging decisions. We have experiments on borderline cases. In Camelyon17, the categories are pN0, pN1-mi, pN1, pN2, where pN1-mi is the borderline case between pN0 and pN1.From Table 1 and Fig. 2(b), our model remains competitive on such borderline cases, thus validating the effectiveness of our uncertainty estimation. Q6: Type-2 fuzzy logic (R1W3) A6: The core motivation for introducing IT2FS is to model the inherent fuzziness in pathological staging, rather than serving as a general-purpose uncertainty estimation method. The advantages of IT2FS over T1FS in complex fuzziness modeling have been widely recognized in literatures. Our ablation experiments show replacing IT2FS degrades performance noticeably, indirectly verifying its effectiveness. Q7/A7: We will correct all errors, add relevant references and dataset citations, and release the code in the final version. (R1, R2, R3)
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 reviews are mixed and recognize the novelty of introducing fuzzy logic into MIL for WSI staging, but several concerns need to be addressed in rebuttal. Please clarify the full methodology, including the interaction between the semantic prior path and the main model, the overall loss function, calculation of fuzzy parameters, choice of interval number/temperature, and construction of the concept bank. Please also justify the framework complexity with stronger component-level analysis, and provide evidence for the claimed uncertainty estimation, interpretability, and robustness on borderline or morphologically ambiguous cases. Clarifications on reproducibility, dataset references, baseline settings, and figure/name inconsistencies would also be important for the final decision.
- 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.
Reviewers appreciated the novel introduction of interval type-2 fuzzy logic into MIL for modeling pathological ambiguity, along with strong empirical performance and reasonable ablation support. Although concerns remain regarding framework complexity, insufficient validation of uncertainty/interpretability claims, and missing implementation details, the rebuttal clarified the formulation, module interactions, and reproducibility issues sufficiently. I therefore recommend 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.
Accept
- Please justify your recommendation.
Two reviewers raised their scores to accept after rebuttal. The reviewer with initial weak reject score did not provide the post-rebuttal score, while recognize the novelty of the method in the initial review. Thouth some concerns are not fully addressed, I recommend acceptance considering the motivation and the methodology contribution.
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 reviewers and the AC consistently recognoize the novelty of fuzzy logic-based MIL for WSI staging, which could improve the model’s interpretability and robustnesss under uncertainty. The current manuscript still has limitations from the experimental justification perspective.