Abstract

Few-shot Whole Slide Image (WSI) classification is challenged by extreme label scarcity and gigapixel resolution. While Vision–Language Models (VLMs) offer a promising route via text-guided retrieval, they often suffer from discriminative sparsity: rigid selection can either overlook key diagnostic regions or include excessive background. We propose Spectral-Semantic Consensus (SSC), a patch selection framework that couples distributed budget prompt retrieval with spectral knowledge projection selection. SSC consists of two complementary modules: (1) a Distributed Budget Prompt Consensus (DBPC) module that employs shared learnable prompts with distributed selection budget to improve the coverage of heterogeneous, diagnostically relevant patches; (2) a Spectral Knowledge Projection (SKP) module that selects patches by their alignment with subspace spanned by principal components (spectral components) derived from textual priors, thus reducing task-irrelevant regions. To adapt aggregation of the two modules, we further introduce an Uncertainty-guided Fusion mechanism that calibrates module contributions using retrieval uncertainty, emphasizing spectral alignment when semantic matching becomes unreliable. By aligning visual evidence selection with the spectral geometry of domain knowledge, SSC provides a robust and principled paradigm for data-efficient computational pathology. Experiments on TCGA-NSCLC, CAMELYON16, and UBC-OCEAN demonstrate consistent and substantial gains over strong baselines in few-shot regimes.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4444_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Hanyu1124/ssc

Link to the Dataset(s)

N/A

BibTex

@InProceedings{DuHan_SpectralSemantic_MICCAI2026,
        author = { Du, Hanyu AND Zhou, Peilin AND Shi, Weizhuo AND Wei, Zhikai AND Chen, Yajie AND Du, Bo AND Xu, Yongchao},
        title = { { Spectral-Semantic Consensus 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

    The SSC model introduces a novel framework for few-shot WSI classification by leveraging the multimodal capabilities of Vision-Language Models (VLMs). Moving beyond previous research that primarily focused on a learnable prompt, the authors propose the DBPC module. This module utilizes a multi-view approach to effectively handle text-visual token alignment, aiming to bridge the gap between sparse pathological labels and rich semantic information.

  • 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 introduction of a multi-view perspective for text-visual token alignment is a significant technical contribution. This approach helps the model induce more stable and robust representations, which is particularly beneficial in the data-constrained environment of few-shot learning.

  • 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 comparison seems potentially biased. Most baselines are standard MIL models that inherently struggle in few-shot scenarios. To claim superiority in few-shot learning, the authors should include more robust few-shot baselines or ensure the comparison is fair.

    • The ablation study regarding uncertainty-guided fusion is insufficient. How does the model behave when the fusion weight is forcefully increased or decreased? A more thorough interpretation of this mechanism’s impact on performance is needed.

    • The mention of SVTC lacks adequate explanation. It is unclear how it contributes to the final performance. An ablation study specifically testing the presence/absence of SVTC would be beneficial.

    • The core strengths of the proposed model over baselines are not clearly articulated. It would be helpful to include a more detailed explanation of the baseline architectures to highlight the specific architectural improvements made in this work.

    • There is a concern that the model relies too heavily on text features. It appears that the text information acts more as a feature engineering tool rather than boosting the weak association between images and labels. The authors should demonstrate that the image features are not merely auxiliary but are substantively integrated with the text.

  • 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?

    The primary factor for my score is the lack of a fair and rigorous performance evaluation. While the paper targets few-shot WSI classification, it predominantly compares its method against standard MIL baselines that are not designed for data-scarce scenarios.

  • 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.

    After reading the authors’ rebuttal, I am revising my score from Reject to Accept. My original concerns centered on the adequacy of baseline comparisons, but having carefully considered the authors’ responses alongside the comments from other reviewers, I am now convinced that the manuscript presents scientifically meaningful findings of sufficient merit. I therefore update my final decision to Accept.



Review #2

  • Please describe the contribution of the paper

    The paper proposes a Spectral-Semantic Consensus (SSC) framework. It contains of two modules: (1) a Distributed Budget Prompt Consensus (DBPC) module (2) a Spectral Knowledge Projection (SKP) module. These two modules are complementary with each other. The paper further use an uncertainty-guided fusion mechanism to aggregate these two modules. Experiments on TCGA-NSCLC, CAMELYON16, and UBC-OCEAN show the effectiveness of SSC under the few-shot setting.

  • 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.Novel Problem Formulation: The authors mentioned “discriminative sparsity,” “semantic collapsing,” and “rigid semantic matching” found in existing text-guided WSI classification methods. The authors proposed a text-guided patch selection framework for few-shot WSI classification, which has clinical meaning. 2.Adaptive Dual-Route Patch Selection: The paper seamlessly integrates two complementary strategies. The DBPC module broadens evidence coverage and prevents overfitting by distributing the selection budget across multiple learnable prompts. Simultaneously, the SKP module filters background artifacts by scoring patches against a PCA-derived semantic subspace, acting as a robust structural prior. Crucially, these modules are dynamically balanced via an Uncertainty-guided Fusion mechanism. 3.Rigorous Experiments: The authors conducted experiments on three different datasets on different organs, and utilized a ten-fold cross-validation protocol to report mean and standard deviation, ensuring robust and trustworthy results.

  • 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.Hyperparameter Sensitivity: The framework introduces specific hyperparameters, such as the number of learnable prompts ($K=5$) , the variance ratio ($\rho$) for SKP , and the divergence loss weight ($\lambda=0.2$). The paper does not provide sensitivity analysis on how varying these values affects cross-dataset generalization. 2.Limited Interpretability of the Spectral Subspace: While the SKP module improves robustness by projecting patches onto a principal semantic subspace, the specific “meaning” of these spectral components can be difficult for clinicians to interpret compared to direct keyword matching. 3.Computational Bottleneck of PCA and Eigendecomposition: The SKP module requires performing eigendecomposition and SVD on text tokens to build the knowledge subspace. This will add additional mathematical preprocessing. It is unknown how much additional computational efforts this method will introduce.

  • Please rate the clarity and organization of this paper

    Good

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

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

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

    N/A

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

    N/A

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

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

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

    Despite some concerns regarding hyperparameter sensitivity and the computational efficiency, the paper’s novel formulation and strong empirical performance on multiple benchmarks make it a contribution to data-efficient computational pathology. However, limited Interpretability seems to be a weakness cannot be easily solved.

  • 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.

    The paper presents a novel method for few-shot WSI classification. The rebuttal helps clarify several concerns such as the comparison to few-shot baselines raised by R1.However, some concerns remain. The analysis of hyperparameter sensitivity and the validation of the fusion mechanism are largely qualitative and would benefit from more thorough ablation. In addition, the efficiency of the SKP module is not quantitatively evaluated. Moreover, while the title is “spectral-semantic,” the semantic interpretability of the learned subspace is not clearly demonstrated, especially from a clinical perspective. Overall, I would give a weak accept according to the motivation and current result, while noting that additional experimental validation and interpretability analysis would still be limitations and concerns which can further strengthen the work.



Review #3

  • Please describe the contribution of the paper

    The paper presents few-shot learning computational pathology method for classification of whole slide images (WSI). Proposed approach is motivated by limitations of two commonly used methodologies: multiple instance learning (MIL) and vision-language models (VLM). With MIL supervision is made only on the slide level while VLM often suffer from discriminative sparsity: rigid selection can either overlook key diagnostic regions or include excessive background. The paper proposed Spectral-Semantic Consensus (SSC), a patch selection framework that couples distributed budget prompt retrieval with spectral knowledge projection selection. The SSC works under MIL setting. A frozen pretrained backbone is used to generated a bag of embeddings. SSC goal is to identify a compact set of informative embeddings and aggregate them to obtain prediction. SSC consists of two complementary modules: (1) a Distributed Budget Prompt Consensus (DBPC) module that employs shared learnable prompts with distributed selection budget to improve the coverage of heterogeneous, diagnostically relevant patches; (2) a Spectral Knowledge Projection (SKP) module that selects patches by their alignment with subspace spanned by principal components (spectral components) derived from textual priors, thus reducing task-irrelevant regions. Both (1) and (2) are important because diagnostically informative regions cover very small percentage of the slide. The DBPC and SKP modules are balanced through uncertainty-guided fusion strategy based on DBPC retrieval uncertainty of the selected patches. It has been demonstrated that SSC improves over strong MIL and VLM baselines in label-scarce regimes.

  • 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.

    Significance of proposed SSC method is in addressing the fundamental weakness in few-shot MIL approach to WSI classification: a single retrieval query that can miss heterogenous diagnostic cues. The SSC method combines two approaches, DBPC and SKP, to select informative tokens (patches) which are fused and afterwards aggregated via prompt queries through cross-modal attention to produce the slide-level prediction. DBPC selects top-r patches per query. SKP selects top-L patches most aligned with principal subspace spanned by principal components of learnable tokens. Selected tokens (patches) are fused by newly proposed uncertainty-guided fusion method which is based on entropy of DBPC-selected patches. The idea is to fuse DBPC- and SKP-selected patches that are minimally redundant.

  • 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.

    To my understanding, the strength of proposed method described previously seems to be also its main weakness. To be specific, I mean on uncertainty-guided fusion method. It is, however, possible that presentation of this method was not clear enough. In equation (4) authors compute normalized entropy of DBPC scores over N patch embeddings. DBPC-module ratios are computed based on 1-entropy, and SKP-module ratios are computed based on entropy. However, entropy is only computed from DPLC-scores. It is based on idea that SKP module excludes prediction-task irrelevant regions (patches). Ablation study conducted in Figure 3 confirms performance improvement due to use of SKP module. I wonder whether authors considered to compute SKP-module ratios based on entropy computed from SKP scores and then using them for fusion model together with DBPC-module ratios? Corresponding ablation study should be carried out to confirm (reject) this assertion.

  • 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

    In overall, presentation is clear although it is occasionally difficult to follow in parts where familiarity with the extensively used meta-language is required. After equations (2) and (3) authors say that L patches are kept after DBPC scoring. However, in equation (1) the number of selected patches is defined as max(1, lowest integer(L/K)). Authors should clarify this.

  • 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?

    My recommendation is accept. My proposal for acceptance is justified by the fact the paper presents method that is addressing the fundamental weakness in few-shot MIL approach to WSI classification: a single retrieval query that can miss heterogenous diagnostic cues. Proposed methodology is presented reasonably clear and conducted experimental analysis on three histopathology datasets yielded convincing performance improvement over few-shot MIL WSI classifiers. Also important, conducted ablation study demonstrated contributions in performance improvement due to the use of DBPC- and SKP modules. My confidence is 75%.

  • 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.

    In rebuttal authors provided satisfactory answers to my comments. By my judgement they also answered in details comments raised by other two reviewers. Thus, I confirm my initial recommendation for acceptance of this paper.



Author Feedback

We thank all reviewers for their constructive feedback. Code will be released.

1.[R1-Q1] Comparison baselines: Great suggestion! We have compared to ViLa-MIL (CVPR 2024), a strong and widely used dedicated few-shot VLM for WSIs. Regarding the comparison to more recent few-shot baselines, we respectfully argue that our main consideration is to maintain the same annotation budget, feature extraction pipeline, data splits, and evaluation protocol. Some recent few-shot WSI methods introduce additional annotation assumptions, e.g., FAST (NeurIPS 2024), or unaligned version of released source codes, e.g., FOCUS (CVPR 2025, see the issue “confusion about text prompt #9” in its official implementation). Besides, since our SSC is a versatile framework rather than being exclusively tailored for few-shot scenarios, for a fair comparison, we evaluate our SSC against diverse WSI-MIL paradigms under the same strictly matched few-shot protocol. Therefore, in our defense, although the baselines are not exhaustive, the current comparison is fair and provides enough informative benchmark.

2.[R1-Q2] DBPC ablation: Excellent question! In our SSC, the fusion weight is a slide-adaptive budget allocation factor derived from DBPC retrieval uncertainty. If the fusion weight is forcefully increased, SSC becomes more prompt-similarity-dominant. This can help when text-image matching is confident. Under high uncertainty, it may over-select visually/textually similar yet incomplete or spurious regions, missing atypical diagnostic evidence. If the fusion weight is forcefully decreased, SSC relies more on the spectral semantic subspace, which can recover patches not captured by direct prompt matching. However, excessive SKP may introduce broadly relevant but less class-discriminative tissue regions. We will clarify this in the revision.

3.[R1-Q3] Contribution of SVTC: We adopt SVTC [21] for better redundant patch elimination. Yet, as claimed in the submission, SVTC [21] is not our proposed contribution. Therefore, we did not conduct ablation study on SVTC. As shown in Fig. 3, we have conducted ablation study on our proposed DBPC and SKP, where both the baseline and our method adopt the same SVTC [21].

4.[R1-Q4] Core strengths compared to the baseline: Our DBPC module distributes budgets across multiple prompts to capture heterogeneous morphology, which can capture more heterogeneous diagnostic relevant patches, compared to the mean pooling strategy in baseline. Our SKP module injects a textual spectral subspace to structurally filter out noise. Visual information is continuous, which may not be captured by text prompt. Thus, SKP can further capture diversified patches by the alignment on subspace spanned by main semantic directions.

5.[R1-Q5] Text feature reliance: Our SSC does not introduce a text-only shortcut to classification. Text priors are only used to guide visual evidence selection, while the final prediction is made from selected WSI visual patch embeddings.

6.[R2-Q1] Hyperparameter sensitivity: Excellent suggestion! Our SSC achieves performance gains across highly heterogeneous cancer types using the same hyperparameter settings. Optimal hyperparameter tuning may yield better results.

7.[R2-Q2] Interpretability of spectral subspace: SKP is not intended to replace clinician-readable concepts. We will state this robustness–interpretability trade-off as a limitation.

8.[R2-Q3] Computational bottleneck: The cost of PCA in SKP—performed on the text-token matrix—depends on text token count and embedding dimension. Compared to thousands of WSI patches, the text token length is small.

9.[R3] DBPC VS SKP scores: SKP is used as a complementary fallback when DBPC becomes uncertain.

10.[R3] Patch selection: The selection budget of each prompt in DBPC is r, computed by max(1, floor(L/K)). Then, we union the selected K*r patches, making the number of selected patches in DBPC <=L. We will clarify this in the revision.




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 paper proposes the Spectral‑Semantic Consensus (SSC) framework for few‑shot WSI classification to address the limitation that a single retrieval query can miss heterogeneous diagnostic cues. The reviewers recognize that this paper propose aovel problem formulation addressing a fundamental weakness of few‑shot MIL. The reviewers also pointed out serious problems, such as insufficient ablation and unclear presentation of the uncertainty‑guided fusion mechanism, lack of clarity in certain methodological details and terminology and the over‑reliance on text features. As pointed out by Reviewer #1, I think the biggest problem of the current paper is that most baselines are standard MIL models that inherently struggle in few-shot scenarios. Instead, recent few-shot papers should be included in baselines. Please respond to the reviewers concerns, including: why not include recent few-shot methods in comparison?How does the model behave when the fusion weight is forcefully increased or decreased? Clarify the inconsistency in patch selection numbers.

  • 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 received three scores of accept after rebuttal. The motivation and novelty are recognized by the reviewers. Though there are still some remainning concens, they can not be resolved without additional experiments. Overall the paper is acceptable.



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.

    This paper proposes a novel framework named Spectral-Semantic Consensus (SSC) for few-show histopathology whole slide image classification. The concerns regarding insufficient ablation and unclear presentation of the uncertainty‑guided fusion mechanism, and lack of clarity in certain methodological details have been properly addressed by the rebuttals. I rate accepting this paper.



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, all reviewers vote clear acceptance. R1 upgraded to Accept after the baseline concerns were addressed, R2 remains a weak Accept while noting residual limitations, and R3 confirms acceptance after satisfactory clarifications.



back to top