acceptodds
Under review as a conference paper at ICLR 2027

Center-Primed Multiple Instance Learning for Reducing Shortcuts in Cross-Center Whole-Slide Image Classification

Abstract

Weakly supervised multiple instance learning (MIL) requires only slide-level labels with low annotation labor costs and is widely used for whole-slide image (WSI) classification. However, inter-hospital differences in staining, scanning, and patient composition may cause models to rely on center-related cues that are spuriously correlated with disease labels, leading to reduced cross-center classification performance. To meet this challenge, this study proposes CenPrime-MIL, which uses fixed center Primes and zero–real Prime dual-path learning with a shared task head to establish an auxiliary pathway for center-related information that can be removed at inference. A Prime-difference center identification constraint directs center-related changes into this auxiliary pathway, while class-conditional center adversarial regularization suppresses center-predictive information in bag-level representations while also preserving pathological discriminative information. Evaluation on three multicenter datasets of breast cancer and skin diseases shows that CenPrime-MIL consistently improves the classification performance of the base MIL models, with average gains of 3.86% in AUC and 3.79% in ACC. Conditional representation analysis further shows that CenPrime-MIL more effectively disentangles center-related encodings () and diagnostic encodings (), supporting its ability to enhance pathological discriminative while reducing the influence of center information. Overall, CenPrime-MIL provides an effective approach to improving the reliability of cross-center WSI classification.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Whole-slide image classification, Multiple instance learning, Cross-center generalization, Computational pathology

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Ue24OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces CenPrime-MIL, a framework designed to mitigate center-specific shortcut learning in cross-center whole-slide image (WSI) classification. The approach combines fixed center embeddings, dual-path learning, Prime-difference center identification, and class-conditional adversarial regularization to reduce center-related information while preserving pathological discriminative features. The method is evaluated on three datasets (CAMELYON17, CAMELYON16, and Skin-PEAN) using ABMIL, CLAM, and TransMIL. The authors report consistent improvements in classification AUC, alongside improved tumor localization, reduced linear center decodability, and higher performance under staining perturbations. However, no comparison with current SOTA methods is provided.

Critical strengths

One aspect I appreciated is that the authors did not limit their evaluation to a single MIL architecture. The method improves mean AUC across all nine combinations of datasets and backbones, which is encouraging. I also found the patch-level tumor localization experiments valuable, since they provide evidence beyond slide-level classification accuracy. The improvements in localization suggest that the method may help the models focus on more relevant pathological regions.

Critical weaknesses

The role of the Prime-difference mechanism is not entirely convincing. My main concern is related to the PDCIH component. Since the center code is explicitly injected into one of the two paths, I wonder whether predicting the center from the difference between their representations actually encourages the model to separate center-related information, or whether the auxiliary classifier simply learns to recognize the injected code. The ablation results show that PDCIH contributes to classification performance, which is encouraging, but they do not fully explain why it helps. I also think a comparison with a simpler model using only class-conditional adversarial regularization, without the Prime mechanism, would be important to better understand the contribution of the proposed approach.

The comparison with other SOTA methods is absent. This is probably the most important weakness. The proposed method is not compared against other SOTA methods, as the ones discussed in the Introduction.

Questions for authors
  1. Regarding the PDCIH objective, since the center code is explicitly injected into the Real-Prime path, how can we be sure that the difference representation captures meaningful center-related information rather than simply encoding the injected center identity? Also, would it be possible to compare the full method with a simpler MIL + CCAR model without the Prime mechanism? I think this would help clarify the actual contribution of the proposed architecture.

  2. Some of the reported AUC improvements are relatively small compared with the variability across the five seeds. Could the authors provide additional statistical significance tests to better assess whether these improvements are consistent and reliable?

  3. In Figure 3, I noticed that both the baseline and CenPrime-MIL seem to follow a similar decreasing trend as the perturbation strength increases. How can we conclude that CenPrime-MIL is more robust to these changes, rather than simply maintaining the performance gain it already showed in the previous experiments?

  4. Why not comparing the proposed method against the other SOTA approaches in cross-center robustness and debiasing in computational pathology, presented in the Introduction ? Is there a major difference or reason ? Please discuss about this point.

2 comments

We thank the reviewer for recognizing the evaluation across MIL architectures and the value of patch-level localization.

  1. Interpretation of PDCIH and the CCAR-only comparison PDCIH supervises the hidden change induced by injecting a center code. Its classifier can recognize that code, so successful center prediction from the difference cannot establish that naturally occurring center information has been removed from the slide representation.

Its intended role is to discourage collapse of the auxiliary pathway, maintaining distinguishable offsets while the shared task classifier learns the same pathological label with and without them. CCAR provides the explicit adversarial constraint on center predictability in the bag representation. We will clarify this division of roles and revise wording that implies guaranteed disentanglement.

We will add base MIL plus CCAR with the entire Prime pathway removed. We will also compare difference-based supervision with a code-only auxiliary head receiving the projected code. Together with matched representation probes, these controls will examine whether the interaction between slide features and the injected offset contributes beyond code recognition. 2. Statistical reliability of the gains The current experiments use the same five seeds across methods. Reported means and variability describe these runs but do not directly test the paired improvements.

We will provide individual-run results, paired differences, confidence intervals, and appropriate paired statistical analyses. Patient-level resampling will preserve the grouping of slides from the same patient. We will distinguish test-cohort uncertainty from variability across training runs, specify the meaning of the reported variability, and account for multiple comparisons. Additional training repetitions will be included where needed to improve the assessment of run-to-run variability. These analyses will establish which gains are consistently supported and which remain uncertain. 3. Interpretation of the perturbation experiment The current experiment shows that CenPrime-MIL generally maintains higher absolute AUC across staining perturbations with frozen model weights. This supports persistence of the classification advantage when staining conditions change.

The reviewer correctly distinguishes this from reduced sensitivity to perturbation. Similar declining curves can preserve an initial performance gap without showing a smaller decline. We will therefore additionally report each model's AUC decrease from its own unperturbed level at every perturbation intensity, using matched perturbation realizations and uncertainty estimates. The wording will distinguish higher performance under perturbation from a smaller performance degradation, with the latter claim determined by the additional analysis. 4. Comparisons with related methods The manuscript already includes IBMIL with ABMIL, CLAM, and TransMIL on all three datasets. We will make this comparison clearer and expand the evaluation to include CaMIL and additional relevant methods with reproducible implementations.

Where applicable, comparisons will use matched patch features, data splits, and model-selection procedures. We will state the center labels and target data available to each method. CAMELYON17 includes evaluation on unseen centers, while CAMELYON16 and Skin-PEAN include some target-center training samples. This distinction will be made explicit when interpreting results.

The additional evaluations will assess performance against closely related approaches under comparable conditions. We will also strengthen the methodological discussion so that differences in assumptions and mechanisms are explained independently of the performance ranking.

Reviewer_Ue24OpenReview ↗

Promise is not a result...

I thank the authors for their answers but promises cannot replace results. Major experiments, such as MIL + CCAR ablation, or CaMIL/SOTA baselines, are deferred to a future revision, which cannot be seen nor evaluated by the Reviewer. Furthermore, statistical significance tests on the existing 5 seeds required no new model training, yet were still withheld from the rebuttal. I thus maintain my score. I believe that the paper will likely be stronger once these promised experiments and statistical checks are actually executed and verified, but I think that it is not ready for acceptance in its current state.

Reviewer_zdrAOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CenPrime-MIL, a framework for reducing center-specific shortcuts in weakly supervised whole-slide image (WSI) classification. It combines fixed center codes, a dual-path learning strategy, a center-identification objective based on the difference between the two paths, and class-conditional adversarial regularization. The framework is designed to use center information during training to improve the robustness of the primary prediction pathway, while requiring no center identity at inference. Experiments on CAMELYON17, CAMELYON16, and Skin-PEAN, using ABMIL, CLAM, and TransMIL, show improvements in several classification and tumor-localization metrics.

Critical strengths
  • The experiments cover three datasets and three MIL aggregators, with additional ablations, representation analysis, staining perturbations, and tumor-localization evaluation. This provides a broader assessment than a study limited to one dataset or one classification architecture, although the strength and consistency of the reported gains vary considerably.
Critical weaknesses
  • The novelty and methodological contributions are not sufficiently established: Although the paper introduces several named components, their individual methodological contributions and the differences from existing work are not clearly articulated. In particular, the relationship to prior work on shortcut mitigation in WSI classification needs a more critical discussion. For example, CaMIL already addresses spurious associations in WSI classification through causal intervention and front-door adjustment, and the paper cites this work without sufficiently clarifying the methodological distinction ("CaMIL: Causal Multiple Instance Learning for Whole Slide Image Classification"). The proposed combination of fixed center codes, dual-path learning, center prediction from representation differences, and adversarial regularization may be useful, but the paper does not explain convincingly why this combination constitutes a substantial advance over established techniques. The contributions should be stated more precisely, and the related work should explicitly compare the assumptions, mechanisms, and limitations of the closest approaches. As presented, the method reads more like a carefully engineered combination of known ideas than a clearly established methodological advance.

  • The performance gains do not establish the effectiveness of the proposed mechanism: Table 1 reports improvements across multiple datasets and MIL aggregators, but the gains are uneven. The ablation study in Table 3 shows that removing individual components often reduces performance, but it does not establish whether the proposed design is better than simpler alternatives.

Questions for authors
  • What is the fundamental methodological difference between CenPrime-MIL and existing approaches such as CaMIL? Which specific limitation of prior methods does the proposed design overcome, and what new insight does it provide beyond combining center encoding with adversarial learning?

  • Can the authors provide controlled comparisons against ordinary center-adversarial training, class-conditional adversarial training, and a conventional auxiliary center classifier? How much additional benefit does PDCIH provide beyond these simpler alternatives?

  • Why are fixed simplex ETF codes preferable to learned center embeddings? Does the fixed geometry provide a measurable advantage, or could the same performance be achieved with learned codes?


After the rebuttal, most of my concerns are answered with a promise of future work rather than actual results. I still don’t see a comparison with simpler options, like CCAR alone without the Prime pathway or a standard center-adversarial baseline, so it’s hard to tell whether the Prime mechanism really helps. The rebuttal also agrees that the method mostly combines known techniques. I maintain my rating.

2 comments

We thank the reviewer for requesting a clearer account of the methodological contribution and comparisons with simpler alternatives.

  1. Methodological distinction and novelty CaMIL estimates interventional predictions through front-door adjustment with a mediator derived from patch features. CenPrime-MIL uses observed training-center labels to construct a supervised training procedure. It introduces a removable center-conditioned offset after aggregation, trains a shared classifier on zero-Prime and real-Prime representations, and applies class-conditional adversarial learning to the bag representation.

The hypothesis is that an identifiable response to center-conditioned variation can be maintained during training while the diagnostic pathway learns to operate without center input. PDCIH encourages distinguishable auxiliary responses, shared pathological supervision encourages correct decisions with and without those responses, and CCAR opposes conditional center prediction from the bag representation.

The proposed contribution is this coordinated training construction and its evaluation in cross-center pathology. The individual ideas of center encoding and adversarial learning are established techniques. Their combination requires controlled evidence to establish its additional value. We will state the contribution more precisely and compare the assumptions, mechanisms, metadata requirements, and inference procedures of the closest methods, including CaMIL. We will also add matched empirical comparisons. 2. Simpler alternatives and the contribution of PDCIH Existing ablations show that removing PDCIH can reduce performance within the full framework. For TransMIL, AUC changes from 0.709 to 0.650 on CAMELYON16 and from 0.914 to 0.885 on Skin-PEAN. These comparisons do not establish superiority over all simpler alternatives or prove that PDCIH removes center information.

We will add unconditional center-adversarial training, CCAR alone with the Prime pathway removed, and an ordinary auxiliary center classifier applied to the bag representation. The ordinary auxiliary classifier will use normal gradient propagation, allowing its effect to be distinguished from gradient reversal. We will retain the comparison between the complete model and its PDCIH-removal variant and add representation probes to examine center decodability and diagnostic discriminability.

These evaluations will use matched features, data protocols, architectures, random seeds, and model-selection budgets. They will assess how much benefit arises from adversarial regularization and how much is added by the Prime mechanism. 3. Fixed ETF codes versus learned embeddings A fixed simplex ETF provides symmetric, centered training-center codes. The codes have equal norms and pairwise similarities, and their mean is zero. Because the Prime projection is linear and bias-free, the zero code also corresponds to the mean projected center offset before activation. The zero-Prime pathway is directly supervised during training. Fixing the codebook additionally removes embedding updates as a source of optimization variability.

These properties motivate the design but do not establish an empirical advantage over learned embeddings. The trainable projection can change distances and angles, and the mean offset before activation does not imply an average of final predictions.

We will compare fixed ETF codes with learned embeddings and a fixed-code control under matched dimensions, objectives, data splits, and optimization budgets. This will distinguish the rationale for a symmetric parameterization from any measurable performance advantage.

Thank you for your detailed response. The clarified role of PDCIH and how the method differs from CaMI are helpful. But most of my concerns are answered with a promise of future work rather than actual results. I still don’t see a comparison with simpler options, like CCAR alone without the Prime pathway or a standard center-adversarial baseline, so it’s hard to tell whether the Prime mechanism really helps. The rebuttal also agrees that the method mostly combines known techniques. I maintain my rating.

Reviewer_dawVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CenPrime-MIL, a framework designed to reduce center-specific shortcut in multiple-instance learning (MIL). The method designs a dual-path architecture with a shared classification head and the Prime-difference center identification objective encourages center-identifiable changes between the two pathways. The class-conditional adversarial regularization is introduced in addition, aiming to reduce center-related information in the bag-level representation. The authors evaluate the method on three pathology datasets with different base architectures. Experimental results show improvements in classification performance, patch-level tumor localization, and robustness.

Critical strengths
  1. The paper addresses an important challenge in computational pathology, where the generalization of models across medical centers is difficult.

  2. The authors evaluate the proposed method on three datasets utilizing different MIL architectures.

  3. Ablation studies were conducted to understand the contributions of different components in the model.

Critical weaknesses
  1. The explanation of the formulations seems to be insufficient. The motivations and hypotheses behind several mathematical formulations are not sufficiently explained, making the Methods Section difficult to follow. For example, The rationale behind the dual-path architecture and its associated objectives is not clearly explained. It is unclear why previously known center code is not utilized in the inference, which might help in the prediction. The motivations of introducing both prime-difference center identification and class-conditional adversarial regularization are also unclear.

  2. It is unclear how the prime-difference objective disentangles the center-related information. Since the difference representation directly depends on this code, the center identification objective may be satisfied simply by learning distinguishable transformations of the provided codes. As a result, the it is not guaranteed that the center-related information contained in the representation has been removed.

  3. The authors conduct ablation studies. However, the version with class-conditional adversarial regularization alone but without other components is missing. Class-conditional adversarial learning provides a mechanism for suppressing center-related information. Therefore, the current result is insufficient to examine whether the improvements is primarily introduced by the class-conditional adversarial regularization alone .

  4. Table 4 suggests that center information is not fully removed but is reduced. it is unclear whether such reduction also generalizes to previously unseen centers. It is also unclear whether the reductions in center predictability are associated with classification performance.

  5. One of the motivations of the paper is the correlations between center identity and disease labels can lead to shortcut learning. However, The systematic experiments with different degrees of class imbalance in different centers were not conducted.

Questions for authors
  1. What are the motivations and hypotheses behind the formulations?

  2. Since the Prime-difference representation is explicitly a function of center code, why is it expected that predicting center identity from this difference help in removing center-related information?

  3. Is the proposed model better than a version with conventional adversarial regularization alone?

  4. Since the center identity remains predictable from the learned representations, are there evidence that the disentanglement is successful? Are the reduction in center predictability generalizes to unseen centers?

  5. How the method performs when class imbalance and sample sizes differ substantially between centers?

2 comments

We thank the reviewer for recognizing the importance of cross-center generalization and the evaluation across multiple datasets and MIL architectures. The comments identify several places where the relationship between the architectural design, training objectives, and empirical evidence needs to be explained more precisely.

  1. What are the motivations and hypotheses behind the formulations? Why is center identity omitted at inference?

Our starting point is that center identity can be associated with both acquisition conditions and disease prevalence. Consequently, a classifier trained on multicenter data may use center-associated features to predict disease labels. The design of CenPrime-MIL addresses this problem through complementary constraints on the training pathways and the learned representation.

For a bag representation z , the two pathways compute

u^{0}=\tanh(q),\qquad u^{+}=\tanh(q+W_pm_d),

where q=W_zz+b_z and m_d identifies the training center. Both pathways use the same classifier and receive supervision from the same pathological label. Our hypothesis is that learning the correct diagnosis with and without this center-conditioned offset encourages the shared classifier to preserve diagnostic decisions under the introduced variation. The zero-Prime pathway receives direct pathological supervision throughout training.

PDCIH encourages the auxiliary pathway to retain distinguishable center-conditioned changes. CCAR acts directly on the bag representation through gradient reversal, reducing center predictability within each pathological class. Class conditioning is motivated by differences in disease prevalence across centers, because unconditional center alignment can also penalize features useful for distinguishing pathological classes.

We use the zero-Prime pathway at inference because a target hospital may be absent from training and therefore has no corresponding learned center offset. The deployed pathway is already trained for this setting. When a test slide comes from a known training center, using its real Prime is a feasible alternative. We will add a controlled comparison of zero-Prime and real-Prime inference for known centers and clarify how this differs from evaluation on unseen centers.

We will reorganize the Methods section around these hypotheses and explain the purpose of each objective before presenting its formulation.

  1. Since the difference representation explicitly depends on the center code, why should predicting center identity from it remove center-related information?

The reviewer correctly identifies that center prediction from the difference representation can be achieved by recognizing the injected code. Successful PDCIH classification therefore does not, by itself, establish that center information has been removed from the bag representation.

The intended role of PDCIH is to discourage collapse of the auxiliary center pathway. If the projected center offsets become zero, then u^{+}=u^{0} , the difference representation becomes identical across centers, and the two pathological classification pathways become duplicates. PDCIH encourages these offsets to remain distinguishable, allowing the shared classifier to be trained in the presence of nontrivial center-conditioned changes.

This distinction is evident from

\Delta u=\tanh(q+W_pm_d)-\tanh(q).

In a locally linear regime, \Delta u can approximately equal W_pm_d . A center classifier could then succeed primarily by decoding the injected offset. With the nonlinear activation, PDCIH can also influence the bag representation through its dependence on q , but this interaction does not guarantee a reduction in center information.

The explicit objective opposing center prediction from the bag representation is CCAR. We will revise statements that could suggest PDCIH independently removes center information or guarantees disentanglement. We will also add matched representation-probing analyses for the relevant ablations to examine how PDCIH and CCAR affect the deployed representation.

  1. Is the full model better than a version with class-conditional adversarial regularization alone?

The current ablations evaluate the contribution of each objective within the complete framework. They show that removing CCAR or PDCIH can reduce classification performance, but they do not establish how much of the improvement can be obtained by adding CCAR alone to a base MIL model.

We will add a MIL + CCAR configuration that removes the entire Prime pathway and its associated objectives. Simply disabling the real-Prime classification loss would not provide this control, because the real-Prime pathway would still participate in PDCIH. We will also include an unconditional center-adversarial baseline to evaluate the contribution of class conditioning.

These comparisons will use matched feature extractors, aggregators, data splits, random seeds, and model-selection budgets. Together with the existing ablations, they will assess the additional benefit of the Prime mechanism over adversarial regularization alone. We will report the outcomes without assuming that every component provides the same benefit in every setting.

  1. Does residual center predictability contradict successful disentanglement? Does the reduction generalize to unseen centers, and is it associated with classification performance?

The empirical objective is to reduce center dependence while retaining diagnostic information. The conditional probes support reduced linearly decodable center information in the deployed representation under most evaluated conditions. They do not establish complete statistical independence or removal of all center-related information.

There is also a distinction between marginal and class-conditional center predictability. When disease prevalence differs across centers, a representation that preserves pathological class information can remain informative about center identity through that association. Our class-conditional analysis controls for this effect by predicting the center within each pathological class. Residual predictability in these conditional probes indicates that some center information remains accessible.

The current manuscript provides unseen-center performance evidence on CAMELYON17 through held-out-center slide classification and five-fold, center-held-out patient staging. These results evaluate prediction generalization. They do not independently demonstrate that the reduction in center decodability transfers to unseen centers.

We will extend the representation analysis to held-out centers using frozen representations and patient-disjoint probe fitting and evaluation. A source-versus-held-out-center probe within each pathological class can assess residual domain distinguishability without requiring a source-trained center classifier to recognize an unseen center label. We will also document the probe cohorts, preprocessing, regularization selection, and sample counts.

Finally, we will examine the relationship between changes in center-probe AUC and classification AUC across matched runs and ablations within each dataset and architecture. The current results show that reduced center decodability and improved classification coexist, while their association requires this additional analysis.

  1. How does the method perform when class imbalance and sample sizes differ substantially between centers?

The formulation includes measures intended to reduce the dominance of large label–center groups. The two center-related objectives use weights

w_{y,d}= \operatorname{clip}!\left( c\sqrt{\frac{n_y}{M n_{y,d}}},,0.5,,2.0 \right),

where n_{y,d} is the number of training bags from center d with label y , and c normalizes the average training-sample weight to one. These weights increase the relative contribution of underrepresented observed label–center combinations. The square root and clipping limit the magnitude of the adjustment. The pathological classification losses retain the original training distribution.

CCAR is activated for a class only when that class occurs in at least two training centers. A class present in one center provides no within-class comparison from which to learn a center-invariance constraint.

These choices motivate the treatment of unequal sampling, but the existing datasets do not provide a controlled assessment across different imbalance levels. We will add experiments varying center-specific class proportions and center sample sizes, including settings with stronger center–label correlations. We will report per-class and per-center performance, including minority-class sensitivity, and compare the full model with the base MIL and CCAR-only configurations. This will evaluate both the benefit of the proposed weighting and the conditions under which limited cross-center class overlap constrains the method.

Comments on the paper

1 comment
Area_Chair_cbLcOpenReview ↗

Gentle Reminder: Discussion Closing Soon

Dear Reviewer,

As the discussion period closes soon, we kindly encourage you to engage in the discussion and share any remaining thoughts or feedback.

Thank you for your time and valuable contributions!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.