acceptodds
Withdrawn from ICLR 2027

Statistical Testing for Multiple Instance Learning via Selective Inference with Applications to Computational Pathology

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multiple instance learning (MIL) is widely used in computational pathology for weakly supervised analysis of whole-slide images (WSIs) without patch-level annotations. In attention-based MIL, high-attention instances are often interpreted as diagnostically important regions, but attention scores alone do not establish whether these instances significantly differ from normal tissue. We formulate this problem as a statistical hypothesis test that compares a selected high-attention instance with a representative normal reference selected based on feature similarity. Because both the target and reference instances are selected through data-dependent procedures, standard hypothesis testing is invalid. We therefore develop a selective inference (SI) framework that accounts for attention-based instance selection and adaptive reference selection, enabling valid selective -values. Experiments demonstrate Type-I error control on synthetic and MNIST-based data and practical applicability to pathological WSIs, with higher statistical power than the conventional over-conditioning approach.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Statistical hypothesis testing, Attention-based multiple instance learning, Selective inference, Computational pathology, Whole-slide image analysis

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_QvWDOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Traditionally, in the multiple instance learning (MIL) field, especially for pathology image classification task, because of hundreds of millions pixels of pathology images, it is impossible to label each pixel by pathologists rather only give a slide-level classification label. So, this transforms WSI classification task to a weakly supervised classification task.

At the same time, to obtain the truly important region of the WSI, the researches usually use attention-based method calculate the attention score to focus high-score patch. But the previous research found these high-score region may not representing the diagnostic related region.

So, aiming at above question, this paper propose a hypothesis testing method to discriminate whether selected patches are obviously different with normal tissue or not, based ABMIL model.

Critical strengths

None

Critical weaknesses
  1. From the model designs, the proposed method lacks of evident motivation about reference patch selection and design. All section method mainly focus on how to perform statistical inference. But in my view, the selection of reference patches are truly important for the testing and the author only state they use adaptively choice method from a collected patch set.

  2. The paper lacks of necessary experiments to support proposed model, such as, the main text only contain one experimental results. I thought the author should carefully revise the paper, perform more experimental designs and consider submit to other publications.

  3. There are some statistical terminology, such as type I error in statistical hypothesis testing, these need to give some simplified background introduction or description. This will help other disciplinary reader understand this more clearly.

Questions for authors

Because of my previous review on weakness and negative score now, I thought at the stage, it is much more important for author to improve the manuscript to support proposed method rather than answer my question. So i don't have any question for authors at current stage.

Reviewer_hJz2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework for testing whether a high-attention instance selected by ABMIL significantly differs from normal tissue. The reference is chosen adaptively as the medoid of the k nearest neighbors among normal instances. Because both the target and the reference are selected in a data-dependent manner, the authors use selective inference to condition on both selection events and characterize the entire truncation region via parametric programming. Experiments on synthetic data, MNIST, and CAMELYON16 demonstrate Type-I error control and higher power than over-conditioning.

Critical strengths
  1. Validating attention regions with a statistical criterion is an interesting direction for computational pathology.
  2. The derivation of the selection events is careful and sound. Conditioning only on the selected medoid, rather than on the k-NN set, is a reasonable design that avoids excessive conditioning.
  3. The stress test (Fig. 3) clearly demonstrates the need to account for both selection events.
Critical weaknesses
  1. Insufficient comparison and limited novelty. The baselines are limited to naive testing, OC, and an ad hoc Bonferroni correction, with no comparison to alternatives such as sample splitting. The methodological difference from Niihori et al. (2025) and Shiraishi et al. (2024; 2026) is also unclear. Given that the selection mechanism considered here is relatively simple, the contribution appears incremental.
  2. Limited experimental scale. The real-data evaluation is restricted to a single center and macro-metastases in CAMELYON16, with only 9 tumor and 30 normal test WSIs. Although tumor masks are available, they are used only for visualization, and the correspondence between rejected patches and tumor regions is not quantitatively evaluated.
  3. Validity of assumptions. The theoretical guarantees rely on a known variance and an isotropic Gaussian model, whereas the real-data experiments use an estimated variance. No sensitivity analysis is provided with respect to misestimation of σ² or the dimensionality reduction (1536 → 192). WSI-level multiplicity control and computational cost are also not discussed.
Questions for authors
  1. How do the rejection proportions change when σ² is misspecified?
  2. Can you provide a quantitative evaluation using the tumor masks?
  3. How does the power of the proposed method compare with that of sample splitting?
Reviewer_zgzWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework for assessing the statistical significance of high-attention instances in attention-based multiple instance learning (ABMIL). For each selected high-attention instance, the method finds its nearest neighbors in the normal reference instances and selects their medoid as the reference. The authors formulate a hypothesis test comparing the underlying signal vectors of the target and reference instances. The authors derive p-values using a truncated chi distribution and parametric programming.

Critical strengths
  1. The paper addresses an interesting problem in computational pathology, where limited statistical evaluation is developed.

  2. The proposed framework accounts for both attention-based target selection and adaptive reference selection.

  3. The method is based on theoretical formulation.

Critical weaknesses
  1. The authors formulate the problem as a pairwise hypothesis test between each high-attention instance and one single selected normal reference instance. I wonder whether this formulation adequately captures the intended scientific question, because normal tissue can exhibit substantial heterogeneity. As a result, the underlying feature representations may differ considerably. Therefore, rejecting the null hypothesis may not indicate pathological abnormality.

More importantly, the authors should justify why comparing the target instance with one representative normal instance is preferable to comparing it with the distribution of normal tissue. Established distribution-based testing approaches, such as Maximum Mean Discrepancy (MMD), could be used to compare the feature distributions of high-attention instances from diseased and normal WSIs. Although these methods address a distribution-level hypothesis, which differs from the proposed instance-level hypothesis, they might better capture the underlying biological question.

  1. The real-world evaluation primarily relies on the rejection rates on normal and tumor WSIs. However, the interpretation of these results is unclear. A higher rejection rate on tumor WSIs does not necessarily indicate improved tumor detection, because tumor WSIs generally contain both normal and malignant tissue. Note that CAMELYON16 provides ground-truth tumor annotations, which can be used in quantitative evaluation. However, the author did not provide such a quantitative evaluation.

  2. The theoretical guarantees rely on independent isotropic Gaussian noise with known variance. It is unclear whether the assumption holds for the representations extracted from real histopathology images. The features from patches extracted from the same WSI are also likely to be dependent.

  3. The authors did not perform sensitivity analyses for the number of nearest neighbors and the attention threshold.

  4. It is not obvious what biological or clinical insight was generated from the analysis of CAMELYON16. It is not demonstrated that the proposed method can help in obtaining biological insights or in making clinical decision.

Questions for authors
  1. What is the reasons of using a single medoid reference rather than a distribution-based approach, such as MMD?

  2. Could the authors provide quantitative tumor-localization results utilizing the CAMELYON16 annotations?

  3. How the choice of number of reference instances and the estimated noise variance impact the performance?

  4. How is the proposed method useful in obtaining biological insights or in making clinical decision?

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

Related papers

Discussion (0)

Sign in to comment.