acceptodds
Under review as a conference paper at ICLR 2027

OGPCMIL: Omics-Guided Prototype Calibration Multimodal Multiple Instance Learning for Few-Shot Whole Slide Image Subtyping of Rare Cancers

Abstract

Accurate subtype classification of rare cancers is essential for individualized clinical treatment. However, rare cancers suffer from limited cases and scarce annotations, so unimodal pathological data fail to capture subtype-specific features and stably distinguish fine-grained subtypes. Multimodal learning integrating whole slide images (WSIs) and genomic data provides a feasible solution, yet existing multimodal and multi-instance learning methods still have key limitations for few-shot rare cancer subtyping: (1) Most methods directly establish dense cross-modal interactions between high-dimensional omics features and numerous image patches through global fusion. Such operations easily cause overfitting under limited labeled data and fail to capture stable subtype distinctions. (2) Genomic profiles reflect the overall molecular state of patients, while subtype-specific pathological features only exist in local WSI regions. Current approaches cannot effectively connect global omics priors with local patch features, resulting in underutilization of cross-modal complementary information. To address these issues, we propose OGPCMIL, an Omics-Guided Prototype Calibration Multimodal Multiple Instance Learning framework for rare cancer subtyping. We first adopt a low-rank prototype calibration mechanism to transform high-dimensional omics information into subtle and reliable subtype prototype adjustments, effectively avoiding overfitting in few-shot cross-modal fusion. Furthermore, we design a class-conditional attention aggregation mechanism, which delivers global omics priors to local patches guided by calibrated prototypes, establishing precise associations between omics signatures and subtype-related pathological regions. Extensive 1-shot, 5-shot and 10-shot experiments on three rare cancer datasets show that OGPCMIL outperforms all compared baselines in 18 of 27 settings. Ablation studies validate the effectiveness of our core modules.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Rare cancer, whole-slide pathological image, few-shot subtype classification, multiple instance learning, multimodal learning, prototype learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_84JsOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes OGPCMIL for few-shot whole-slide image (WSI) subtyping of rare cancers using histopathology and transcriptomic data. The method introduces omics-guided low-rank prototype calibration and class-conditional attention aggregation to address overfitting in multimodal fusion and global–local misalignment. Experiments are conducted on three TCGA cohorts under different few-shot settings and pathology foundation models, with additional ablation and sensitivity analyses.

Critical strengths
  1. The paper identifies relevant challenges in few-shot multimodal pathology and presents a reasonably coherent framework. The low-rank calibration strategy provides a constrained mechanism for incorporating genomic information.

  2. The evaluation covers multiple datasets, few-shot settings, and pretrained backbones, with several ablation and sensitivity analyses.

Critical weaknesses

1 The main claim of outperforming baselines in 18 of 27 settings relies on win counts, while many performance differences are small relative to the reported variability. No paired statistical significance tests are provided. Moreover, the proposed method underperforms several baselines in certain settings, raising concerns about its robustness and generalizability.

  1. Despite focusing on multimodal learning, the paper lacks comparisons with established omics-aware or multimodal fusion methods. Consequently, the claimed advantages over conventional cross-modal fusion remain unverified. Additionally, key ablations are restricted to a single favorable dataset–backbone configuration, making it difficult to assess whether the proposed components consistently contribute across different settings.

  2. The pathway analysis relies on gradient-based attribution without sufficient validation, while the localization results lack quantitative comparisons or expert assessment. The calibration analysis is also limited to a single configuration. These results do not adequately establish that the omics-guided mechanism provides clinically meaningful or biologically reliable explanations.

  3. The trainable parameter counts of the proposed method and baselines are not reported, making it difficult to assess whether improvements arise from the proposed design or differences in model capacity. Some notation and terminology also require clarification.

Questions for authors
  1. Why are established omics-aware or multimodal MIL baselines excluded, given that multimodal fusion is central to the paper's contribution?
  2. Can the key ablation experiments be extended to additional datasets and foundation models to demonstrate consistent effectiveness?
Reviewer_xXJPOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

OGPCMIL uses patient gene expression to calibrate class prototypes through a bounded low-rank correction. The resulting prototypes guide class-conditional patch attention for rare-cancer subtyping. Evaluation covers three TCGA cohorts, three shot settings, and three frozen pathology encoders, with the best reported BACC in 18 of 27 settings.

Critical strengths

S1. The prototype calibration and class-conditional pooling address a specific modeling problem: using global molecular information to select localized subtype evidence.

S2. The evaluation spans three rare-cancer cohorts, three label regimes, and three pathology encoders. Tables 4-6 report the full comparisons, including failures, while Figure 6 examines the contribution of omics across all 27 settings.

Critical weaknesses

W1. Section 4.1 compares the multimodal model primarily with image-only MIL baselines. Without an omics-only classifier and competitive multimodal few-shot controls across cohorts, the results cannot distinguish improved fusion from simply having access to an additional predictive modality.

W2. The supplied UCS ten-shot test set contains only five patients, although evaluation.py scores individual slides. Repeating training on the same test patients does not estimate population uncertainty; patient-aggregated predictions and patient-level intervals are needed for reliable conclusions.

W3. Tables 1 and 3 show small mean gains for low-rank calibration, gating, and individual regularizers relative to their run variability. Ablations on one UCS configuration do not support the claim that every component is necessary.

Questions for authors

Q1. Do the gains remain against omics-only and matched multimodal baselines with the same annotation budget?

Q2. Section 4.1 excludes MPNST from ten-shot SARC, and the split files confirm four active classes. Which model and cohort produce the five-subtype ten-shot interpretation in Appendix B?

Reviewer_p4jMOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes OGPCMIL, a multimodal multiple instance learning framework for few-shot subtyping of rare cancers from whole-slide images (WSIs) that uses patient-level RNA-seq expression as a low-capacity prior for calibrating class prototypes. The authors first extract patch features with a frozen pathology foundation model, refine them with a lightweight self-attention encoder, and compress the gene-expression profile into a compact molecular state vector. They then introduce an Omics-Guided Low-Rank Prototype Calibration module, in which the molecular state generates per-class coefficients over a small set of shared calibration axes. The resulting offset is norm-bounded and blended with the uncalibrated representation through a gating network, producing a patient-specific dynamic prototype for each subtype. In this way, genomic information adjusts the class representations rather than interacting densely with patch features. Next, a Class-Conditional Attention Aggregation module assigns each patch a class-specific attention weight based on its distance to the calibrated prototype of each subtype. This yields a separate bag representation per class, which is matched against the corresponding prototype to produce slide-level scores. Training combines cross-entropy on the calibrated and an auxiliary uncalibrated branch with topology-preservation, sparsity, and consistency regularizers. The method is evaluated against five MIL baselines on three TCGA rare cancer cohorts (UCS, THYM, SARC) under 1-, 5-, and 10-shot settings with three foundation-model feature extractors. The authors report the best balanced accuracy in 18 of 27 settings. The proposed components are examined through ablation studies, hyper-parameter sensitivity analysis, a unimodal comparison, and interpretability analyses at the region and pathway levels.

Critical strengths

The paper tackles the underexplored and clinically meaningful problem of few-shot multimodal subtyping for rare cancers, with a fairly systematic patient-level evaluation across three datasets, three foundation models, and three shot settings, and the code is publicly released.

Critical weaknesses
  1. The proposed global--local alignment mechanism is not sufficiently demonstrated, and the method formulation is difficult to follow because of inconsistent or insufficiently defined notation. The current attention visualizations only show the model's internal patch weighting and do not establish that omics guidance meaningfully changes attention toward subtype-relevant regions. More direct analyses, such as comparing attention with and without omics, perturbing relevant molecular features, or validating highlighted regions against pathological evidence, are needed.

  2. The baseline comparisons do not isolate the architectural contribution of OGPCMIL. The proposed model uses both WSI and gene-expression data, whereas the main competing methods use only pathological images. Consequently, the reported improvements may partly result from access to an additional modality. Omics-only controls, the image-only OGPCMIL variant, standard fusion baselines, and established WSI--omics methods evaluated under the same few-shot protocol are needed for a fair comparison.

  3. The evaluation protocol does not provide sufficient evidence that the reported improvements are statistically reliable. Test-set sizes and class distributions are not reported, and repeated training on a fixed test set does not capture uncertainty arising from test-set composition. Moreover, several gains over the strongest baseline are small relative to the reported variation, so counting wins across the 27 settings is insufficient to establish a consistent advantage. Paired statistical comparisons and a clear validation-only hyperparameter selection protocol are needed to rule out chance variation and potential test-set tuning.

Questions for authors
  1. The claimed global–local alignment lacks direct evidence. The paper does not show how omics guidance changes patch attention or whether these changes highlight regions relevant to the subtype. Attention maps alone do not establish alignment.
  2. The method section uses inconsistent notation. Several symbols change meaning or lack clear definitions, making the formulation difficult to follow.
  3. The comparisons do not isolate the contribution of the proposed method. OGPCMIL uses both images and gene expression, while the main baselines use images alone. Omics-only baselines and established WSI–omics fusion methods are needed to distinguish architectural gains from the benefit of additional input.
  4. Test-set sizes and class distributions are not reported. Given the small cohorts, the test sets may be too small for reliable evaluation. Repeating training on a fixed test set does not address this uncertainty.
  5. Several gains over the strongest baseline are marginal relative to the reported variability. The count of 18 wins out of 27 settings is therefore not sufficient evidence of a consistent advantage. Paired statistical comparisons would help establish whether these gains are meaningful.
  6. The hyperparameter selection protocol is unclear. The paper reports that the defaults achieve the best performance in the sensitivity analysis but does not specify which split was used to select them. This raises a concern about potential test-set tuning. The authors should clarify that all choices were made using validation data only.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.