List of Papers Browse by Subject Areas Author List
Abstract
While fine-grained prompt-to-patch matching has significantly advanced few-shot whole-slide image (WSI) classification, it introduces a critical vulnerability: the massive aggregation of local interactions inherently amplifies alignment noise.
Such noise stems from both intrinsic pathological ambiguity, where shared local morphologies trigger spurious non-target responses, and extrinsic textual unreliability, arising from incomplete or noisy LLM-generated prompts. Together, these factors severely blur inter-class boundaries. To tackle this, we propose Confusion-Aware Evidence Calibration (CAEC), a robust framework designed to explicitly rectify alignment noise. Specifically, CAEC maintains a dynamic confusion memory module to capture recurring misleading activations, which guides a sample-adaptive calibrator to suppress spurious text-guided responses. Furthermore, we introduce a morphology evidence compensator that adaptively integrates pure visual features to anchor predictions when text guidance is weak. Evaluated on the TCGA-RCC and TCGA-NSCLC benchmarks, CAEC achieves state-of-the-art performance, demonstrating remarkable robustness in data-scarce settings.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3046_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{LiPen_ConfusionAware_MICCAI2026,
author = { Li, Peng AND Yan, Xuefeng AND Ma, Gaoxiang AND Tang, Hao AND Wei, Mingqiang AND Qin, Jing},
title = { { Confusion-Aware Evidence Calibration for Few-Shot Whole Slide Image Classification } },
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 addresses a critical weakness of prompt-to-patch matching for few-shot whole-slide image (WSI) classification: alignment noise accumulates when aggregating thousands of local patch–prompt similarities, blurring inter-class boundaries. The authors propose CAEC, a confusion-aware evidence calibration framework with three components: (i) a dynamic confusion memory that tracks over-activated non-target prompts, (ii) a sample-adaptive calibrator that uses cross-attention with confusion prompts to predict residual logit corrections, and (iii) a morphology evidence compensator that adds gated, pure visual logits when language guidance is weak. Across TCGA-RCC and TCGA-NSCLC under 4/8/16-shot settings, CAEC delivers consistent improvements over strong baselines, with clear gains in the lowest-shot regime and ablations confirming complementary benefits of both modules.
- 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 proposed dynamic confusion memory and adaptive calibrator effectively track, predict, and mitigate alignment noise within the VLM environment. 2.The Morphology Evidence Compensator (MEC) acts as a reliable “failsafe” when text guidance is suboptimal, integrated alongside parameter-efficient designs. 3.Comprehensive evaluation across multiple few-shot settings (4/8/16-shot) with detailed ablation studies and strong interpretability visualizations. 4.Successfully overcomes the noise accumulation issue inherent in prompt-to-patch methods, achieving SOTA or near-SOTA performance (particularly excelling in the very low-shot regime).
- 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 currently only reports limited 5-fold cross-validation experiments evaluated on just two datasets: kidney cancer (TCGA-RCC) and lung cancer (TCGA-NSCLC).
2.Confusion memory update relies on several heuristics (top-r anchor averaging, margin m, EMA with η) without sensitivity analysis; stability or convergence properties are not characterized.
3.The method directly depends on LLM-generated prompts but lacks an analysis of prompt quality/noise sensitivity or a strict protocol for prompt curation
- 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?
The strengths slightly outweigh the weaknesses. The authors are recommended to address the weaknesses during rebuttal.
- 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 proposes Confusion-Aware Evidence Calibration (CAEC), a framework for few-shot whole slide image (WSI) classification that aims to mitigate alignment noise in prompt-to-patch vision-language matching. The method introduces a dynamic confusion memory module to capture recurrent misleading prompt activations, a confusion-aware calibrator to suppress spurious responses via sample-adaptive correction, and a morphology evidence compensator to integrate pure visual features when text guidance is unreliable. The framework is evaluated on TCGA-RCC and TCGA-NSCLC datasets and demonstrates improved performance in low-shot settings.
- 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) Clear problem motivation: The paper identifies an important limitation of prompt-to-patch matching, namely the amplification of alignment noise when aggregating large numbers of patch–prompt interactions (p.2) .
2) Meaningful methodological contribution: The introduction of a confusion memory to explicitly model inter-class confusion is a novel and intuitive idea that goes beyond standard MIL or prompt-learning approaches.
3) Complementary design: The combination of confusion-aware calibration and morphology evidence compensation is well-motivated and addresses both intrinsic ambiguity and extrinsic prompt unreliability.
4) Strong empirical results: The method achieves consistent improvements across multiple few-shot settings (4, 8, 16-shot) on two datasets (Table 1, p.6) .
5) Interpretability effort: Visualization of confusing prompts and logit corrections (Fig. 3, p.8) provides useful insight into model behavior.
6) Well-presented framework: The pipeline is clearly illustrated (Fig. 2, p.3), and the methodology is logically structured.
- 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) Heavy reliance on LLM-generated prompts without sufficient analysis: The method critically depends on LLM-generated morphological descriptions (p.3–4) , yet the paper does not evaluate the quality, variability, or robustness of these prompts. It remains unclear how sensitive the method is to noisy, incomplete, or biased prompt generation, which is a central assumption of the framework.
2) Limited validation of confusion memory effectiveness: While the confusion memory is a key contribution, there is no quantitative analysis demonstrating that the learned confusion patterns correspond to meaningful pathological similarities rather than dataset-specific artifacts or overfitting.
2) Limited generalization analysis: Experiments are restricted to two TCGA datasets. The lack of cross-dataset or external validation raises concerns about generalization, particularly in real-world clinical settings where domain shifts are common.
3) Moderate methodological novelty: The framework builds on existing paradigms (MIL, VLMs, prompt learning, cross-attention), and the main novelty lies in the calibration mechanism. While useful, this may be considered incremental relative to recent advances.
4) Lack of clinical or expert validation: No pathologist evaluation is provided to assess whether the corrected predictions or confusion patterns are clinically meaningful.
5) Computational and scalability aspects are not discussed: The addition of memory and calibration modules may introduce overhead, but efficiency and scalability are not analyzed.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
The proposed framework is promising and addresses an important challenge in vision-language WSI classification.
However, the paper would benefit from:
1) Analysis of robustness to prompt quality (e.g., noisy or perturbed prompts)
2) Quantitative validation of the confusion memory (e.g., alignment with known class similarities)
3) Cross-dataset evaluation to assess generalization
4) Discussion of computational overhead and scalability
5) Potential expert validation to confirm clinical relevance
Addressing these points would significantly strengthen confidence in the proposed method.
- 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
na
- 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?
This paper addresses an important limitation in prompt-based WSI classification, namely the accumulation of alignment noise in patch-level interactions. The proposed confusion-aware calibration framework is well-motivated and demonstrates consistent improvements across multiple few-shot settings and datasets.
However, the method relies heavily on LLM-generated prompts, and the robustness of the approach to prompt quality is not sufficiently analyzed. Additionally, the effectiveness of the confusion memory is not quantitatively validated, and evaluation is limited to two datasets without broader generalization or clinical validation.
Overall, the paper presents a solid and technically sound contribution with meaningful improvements, but would benefit from deeper analysis of its core assumptions and broader validation. I am therefore slightly above the acceptance threshold, but my decision is contingent on addressing these concerns.
- 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
The paper introduces a confusion-aware evidence calibration framework for few-shot whole slide image classification, which goes beyond standard prompt-to-patch matching by explicitly modeling and correcting alignment noise. In particular, it proposes a dynamic confusion memory to capture recurrent misleading prompt activations and a sample-adaptive calibrator that refines similarity-based predictions at the logit level. In addition, a morphology evidence compensator is designed to incorporate pure visual cues when textual guidance is unreliable, leading to more robust multimodal predictions.
- 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 main strengths of the paper lie in its clear motivation and targeted design for addressing alignment noise in prompt-to-patch matching. The proposed framework is well-structured, with complementary components that handle both spurious text-guided responses and unreliable language priors. The method is technically sound and integrates naturally with existing vision–language pipelines. In addition, the experimental results are comprehensive, covering multiple datasets and shot settings, and show consistent improvements over strong baselines, supporting the effectiveness of the approach.
- 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 main weaknesses of the paper lie in its limited methodological novelty and somewhat incremental nature. The overall framework builds upon standard prompt-to-patch matching and introduces additional modules for calibration and compensation, but these components are largely based on existing ideas such as attention, and logit adjustment, without fundamentally new formulations. In addition, the actual performance gains, while consistent, are relatively modest over strong baselines, and it remains unclear whether the added complexity is fully justified. The presentation could also be improved, as parts of the pipeline (e.g., Fig. 2) are not clearly aligned with the described computation, which may hinder understanding of the core contributions.
- 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.
(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?
It introduces a logit-level calibration mechanism on top of standard prompt-to-patch matching, which is a reasonable but incremental extension. However, the presentation (e.g., Fig. 2) does not clearly reflect this key idea.
- 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
Author Feedback
We sincerely thank all reviewers for their constructive comments. We are grateful that the reviewers recognized the motivation of addressing alignment noise in prompt-to-patch matching, the rationale and design of the proposed framework, and the consistent improvements across few-shot settings.
We also appreciate the reviewers’ valuable suggestions. These suggestions are helpful for better understanding the limitations of the current work and will provide useful guidance for future research.
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 introduces a new few-shot learnig framework for whole slide image analysis task. All three reviewers give positive scores and recognize the rationality of the motivation in dealing with the noise in matching the patch and text embeddings. The proposed method is compared to strong recent baselines and show performance improvement on two datasets. Small number of evaluation datasets prevents me from recommending this paper for oral presentation.
