acceptodds
Under review as a conference paper at ICLR 2027

EC-MIL: Disentangling Evidence and Context for Intrinsic Instance-Level Explanations

15 views

Abstract

Modern multiple instance learning (MIL) models enhance bag-level prediction via instance interactions, but the same interactions entangle individual contributions and hinder instance-level attribution. This tension is particularly important in biomedicine, where instance-level supervision is scarce yet identifying supporting or opposing instances is essential. To address this, we introduce Evidence-Context Disentangled MIL (EC-MIL), a framework separating instance-local, class-specific evidence from bag-dependent context weighting. EC-MIL expresses each class logit as a weighted sum of instance evidence, preserving backbone context modeling while making individual contributions directly accessible. The product of each instance's evidence and context weight naturally defines Weighted Evidence (WE), an intrinsic, signed, class-specific attribution whose sum exactly recovers the class logit in a single forward pass. Under explicit assumptions, we justify this decomposition and show that excess population bag-level cross-entropy risk controls instance-contribution decomposition error. Across synthetic, whole-slide image, and mutation-bag tasks, EC-MIL maintains competitive predictive performance, while WE performs well in evidence recovery and instance-removal faithfulness at lower attribution cost than post-hoc alternatives. On mutation bags, WE also identifies gene-level signals consistent with established tissue–gene associations. Together, these results demonstrate that expressive context modeling and efficient intrinsic attribution can be achieved within a unified MIL framework. The code can be obtained from: https://anonymous.4open.science/r/EC-MIL-6449.

Primary area:
interpretability and explainable AI
Keywords:
multiple instance learning, explainable AI

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_oMhqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1(−1)
Read review
Paper summary

EC-MIL introduces a MIL framework that separates the estimation of instance-class-specific contribution from the context of other instances. This framework provides the possibility of intrinsically explaining the signed contribution of each instance to the bag prediction in a single forward pass and is adaptable to different MIL backbone architectures.

Critical strengths
  • Targeting instance-level explanation in MIL architectures as a critical requirement for medical and biomedical fields.
  • The proposed method is flexible to be adjusted to different backbone architectures (ABMIL, TransMIL, and etc) and not limited to a specific backbone design.
  • Providing an efficient single forward pass solution for instance-level explanation during inference.
  • Wide evaluation covering different backbone architectures, datasets, and components of the proposed method.
  • The source code is provided.
Critical weaknesses
  1. Insufficient comparison with competing intrinsic attribution methods (Additive MIL is insufficiently compared, and AttriMIL [1] is not included at all). In particular, there is no systematic comparison of the proposed approach with competing intrinsic methods under the same backbones and datasets. Therefore, the evaluation does not sufficiently establish the advantages of EC-MIL over its main intrinsic competitors.

  2. In Section 4.2, Assumption C is strong and is not guaranteed by the proposed architecture. The authors aim to prove that a low risk of bag prediction implies low instance attribution error. Their proof assumes that the attribution error of each instance does not change too much when other instances change. But even if the evidence (E(xi)E(x_i)) remains fixed, its context weight (αi(B)\alpha_i(B)) can change significantly (under context effect). Therefore, the attribution error may also change. The authors should explain why Assumption C is expected to hold for EC-MIL, or evaluate it on synthetic data where the true instance contributions are known.

  3. The authors claim that WE is a signed attribution for each class, but their theoretical result in Section 4.2 concerns the difference between the contributions of two classes. If we add the same value to the evidence of an instance for both classes, the logit difference and therefore the cross-entropy loss will not change, while the WE of each class changes. Therefore, bag-level CE cannot guarantee the accuracy or sign of each class-specific WE.

  4. The experimental results show that EC-MIL improves performance in some settings but leads to underperformance in others. The authors should provide an explanation of what limits or boosts the performance of their method in those cases. Since the proposed method aims to improve interpretability, understanding its own behavior and limitations is also important. For example, EC-MIL behaves differently on BRACS vs. CAMELYON16. The question is: which characteristics of these datasets relate to the positive or negative performance changes of the proposed model compared to the baseline backbone?

  5. The writing is hard to follow. The introduction and related work need to be polished to cover only the relevant explanation. Also, in the introduction, the distinction between post-hoc and intrinsic approaches is not clarified. For example, in Section 2.2, the descriptions and limitations of GI and IG are combined into one long sentence with multiple citations, while they could be explained in two shorter sentences.

  6. The complete performance comparisons should be moved to the main text, as the differences are difficult to see in the charts. Figure 2 has poor presentation. The differences are not visible. Consider magnifying the relevant performance range (e.g., above 50%) or showing the performance differences separately.

  7. A statistical comparison would be beneficial to assess the significance of the improvements provided by the proposed method.

  8. While WE is supposed to be signed contributions, Figure 4a shows only positive WE values. A distribution plot of the computed WE values, showing the proportion and magnitude of positive and negative contributions, would help clarify how signed attribution behaves in practice.

Reference: [1] AttriMIL: Revisiting attention-based multiple instance learning for whole-slide pathological image classification from a perspective of instance attributes, L. Cai et al., Medical Image Analysis 2025

Questions for authors

Please refer to weaknesses.

10 comments

Thank you for recognizing the importance of instance-level explanations in MIL, the flexibility of EC-MIL across backbones, and its efficient single-forward-pass attribution. Below, we address your concerns and questions point by point.

W1: Systematic Comparison with Competing Intrinsic Attribution Methods (AdditiveMIL and AttriMIL)

(1) Our original manuscript already compares EC-MIL with AdditiveMIL using the same ABMIL backbone across prediction, evidence recovery, and deletion evaluation. Table 1 in the original manuscript reports classification AUC and attribution AUPRC-2 on three synthetic tasks, with EC-ABMIL exceeding AdditiveMIL on both metrics in all three tasks. For example, Adjacent-Pairs AUC is 0.8149 vs. 0.7803, and AUPRC-2 is 0.8000 vs. 0.6027. Original Appendix Tables A19–A25 also report AdditiveMIL prediction or attribution results on synthetic tasks, MUT, and CAMELYON16 with ResNet18 features. The original manuscript thus includes intrinsic-method comparisons using the same backbone, as well as prediction and explanation evaluations across tasks.

(2) Thank you for suggesting the inclusion of AttriMIL and a more systematic comparison. In this rebuttal, we include AttriMIL and compare it jointly with AdditiveMIL and EC-MIL using the same tasks, input features, and data splits. We use the ABMIL implementation of AdditiveMIL and EC-ABMIL for comparison within the same backbone, and also report AttriMIL, EC-TransMIL, and EC-ILRAMIL to show the prediction and attribution performance of the intrinsic methods. AdditiveMIL uses its native Patch-Logits, AttriMIL uses its native instance attribution, and EC-MIL uses WE. The table below reports classification AUC, recovery of known instance contributions (AUPRC-2), and deletion AUPC. Revised Table 1 and Figure 3 include AttriMIL and all four synthetic tasks; complete classification, evidence-recovery, and deletion results with standard deviations are consolidated in revised Appendix Tables A8 and A11–A12.

The synthetic evaluation covers the three original tasks and adds Pos-Neg-L. Digits 1, 2, and 3 each contribute +1; digits 4, 5, and 6 each contribute −1; and digits 0, 7, 8, and 9 contribute 0. When digits 1, 2, and 3 form an L shape with 2 at the corner, each participating instance contributes +2. The bag label is determined by the sum of instance contributions. This evaluates local instance evidence and three-instance spatial relations within the same task. On MUT, we use the scGPT and text2vec features from the original manuscript to further compare prediction and deletion faithfulness on a real-world task.

Rebuttal Table 1: Prediction and attribution comparison of intrinsic methods across four synthetic tasks. Values for AdditiveMIL and EC-MIL on the first three tasks are from the original manuscript; AttriMIL and Pos-Neg-L results are newly added. Bold indicates the highest mean AUC and AUPRC-2 and the lowest mean AUPC for each task.

TaskModelAUC ↑\uparrowAUPRC-2 ↑\uparrowAUPC ↓\downarrow
4-BagsAdditiveMIL (ABMIL)0.9916 ±\pm 0.00150.9053 ±\pm 0.03700.3464 ±\pm 0.0169
4-BagsAttriMIL0.9966 ±\pm 0.00060.7615 ±\pm 0.02840.6244 ±\pm 0.0842
4-BagsEC-ABMIL0.9932 ±\pm 0.00180.9307 ±\pm 0.00740.3618 ±\pm 0.0164
4-BagsEC-TransMIL0.9942 ±\pm 0.00110.9386 ±\pm 0.00660.3635 ±\pm 0.0058
4-BagsEC-ILRAMIL0.9950 ±\pm 0.00140.9295 ±\pm 0.01720.3654 ±\pm 0.0040
Pos-NegAdditiveMIL (ABMIL)0.9492 ±\pm 0.00630.8891 ±\pm 0.01450.2599 ±\pm 0.0190
Pos-NegAttriMIL0.9679 ±\pm 0.00400.7869 ±\pm 0.04220.4148 ±\pm 0.0566
Pos-NegEC-ABMIL0.9702 ±\pm 0.00590.9351 ±\pm 0.00520.2396 ±\pm 0.0067
Pos-NegEC-TransMIL0.9652 ±\pm 0.00390.8998 ±\pm 0.00960.2641 ±\pm 0.0093
Pos-NegEC-ILRAMIL0.9729 ±\pm 0.00540.9183 ±\pm 0.00790.2533 ±\pm 0.0107
Adjacent-PairsAdditiveMIL (ABMIL)0.7803 ±\pm 0.05920.6027 ±\pm 0.05930.5767 ±\pm 0.0956
Adjacent-PairsAttriMIL0.9456 ±\pm 0.02910.5751 ±\pm 0.01390.8404 ±\pm 0.0551
Adjacent-PairsEC-ABMIL0.8149 ±\pm 0.01110.8000 ±\pm 0.02050.4689 ±\pm 0.0588
Adjacent-PairsEC-TransMIL0.9878 ±\pm 0.00390.7970 ±\pm 0.01370.4470 ±\pm 0.0186
Adjacent-PairsEC-ILRAMIL0.9910 ±\pm 0.00280.7967 ±\pm 0.01670.4360 ±\pm 0.0108
Pos-Neg-LAdditiveMIL (ABMIL)0.8660 ±\pm 0.00660.9403 ±\pm 0.00320.3251 ±\pm 0.0087
Pos-Neg-LAttriMIL0.8798 ±\pm 0.00580.8331 ±\pm 0.04640.4485 ±\pm 0.0337
Pos-Neg-LEC-ABMIL0.8777 ±\pm 0.00650.9824 ±\pm 0.00320.3471 ±\pm 0.0092
Pos-Neg-LEC-TransMIL0.9849 ±\pm 0.00230.9743 ±\pm 0.00430.3085 ±\pm 0.0076
Pos-Neg-LEC-ILRAMIL0.8939 ±\pm 0.00670.9773 ±\pm 0.00490.3514 ±\pm 0.0104

The additional results show that EC-ABMIL achieves higher AUPRC-2 than AdditiveMIL and AttriMIL on all four synthetic tasks. For example, on 4-Bags, AttriMIL achieves a classification AUC of 0.9966 and a native-attribution AUPRC-2 of 0.7615, compared with 0.9932 and 0.9307 for EC-ABMIL. On Pos-Neg-L, EC-ABMIL achieves an AUPRC-2 of 0.9824, compared with 0.9403 for AdditiveMIL and 0.8331 for AttriMIL. Jointly reporting these metrics evaluates predictive ability and instance evidence recovery separately.

Rebuttal Table 2: Prediction and deletion faithfulness of intrinsic methods on MUT. AdditiveMIL and EC-MIL values are from the original manuscript; AttriMIL is newly added. Bold indicates the highest mean AUC and the lowest mean AUPC for each feature setting.

FeatureModelAUC ↑\uparrowAUPC ↓\downarrow
scGPTAdditiveMIL (ABMIL)0.9442 ±\pm 0.00330.0190 ±\pm 0.0010
scGPTAttriMIL0.9123 ±\pm 0.00470.6027 ±\pm 0.0310
scGPTEC-ABMIL0.9482 ±\pm 0.00260.0232 ±\pm 0.0004
scGPTEC-TransMIL0.9620 ±\pm 0.00400.0257 ±\pm 0.0006
scGPTEC-ILRAMIL0.9752 ±\pm 0.00250.0322 ±\pm 0.0010
text2vecAdditiveMIL (ABMIL)0.9500 ±\pm 0.00220.0193 ±\pm 0.0007
text2vecAttriMIL0.9185 ±\pm 0.00300.6156 ±\pm 0.0379
text2vecEC-ABMIL0.9501 ±\pm 0.00480.0230 ±\pm 0.0007
text2vecEC-TransMIL0.9626 ±\pm 0.00260.0257 ±\pm 0.0005
text2vecEC-ILRAMIL0.9748 ±\pm 0.00220.0308 ±\pm 0.0019

The MUT results further show the prediction and attribution performance of the intrinsic methods. With scGPT and text2vec, EC-ABMIL achieves classification AUCs of 0.9482 and 0.9501, compared with 0.9123 and 0.9185 for AttriMIL. Their corresponding native-attribution AUPCs are 0.0232 and 0.0230 for EC-ABMIL, and 0.6027 and 0.6156 for AttriMIL. AdditiveMIL achieves lower AUPCs of 0.0190 and 0.0193 in these two settings. These deletion results are included in revised Figure 5, with complete numerical results in revised Appendix Table A15.

We also include AdditiveMIL and AttriMIL in the new TCGA-LUAD TP53 mutation prediction task, extending intrinsic-method comparisons to another real-world WSI setting. This task predicts molecular status from WSIs using the ResNet18 and pathology-pretrained DINO (SSL) features used in the original manuscript, with frozen feature extractors and the same data splits. Prediction and native-attribution deletion results are reported in Rebuttal Table 3, extending the comparisons to molecular prediction from pathology images. In the revised manuscript, Table 2 reports backbone-averaged classification performance for the original and EC variants, Figure 7 shows deletion faithfulness, and Appendix J Tables A20–A21 provide the complete per-model results.

Rebuttal Table 3: Comparison of intrinsic methods on TCGA-LUAD TP53 mutation prediction.

FeatureModelAUC ↑\uparrowIntrinsic Attribution AUPC ↓\downarrow
ResNet18AdditiveMIL (ABMIL)0.6403 ±\pm 0.03530.3781 ±\pm 0.0728
ResNet18AttriMIL0.6404 ±\pm 0.02980.7922 ±\pm 0.0667
ResNet18EC-ABMIL0.5867 ±\pm 0.04180.2629 ±\pm 0.2559
ResNet18EC-TransMIL0.5572 ±\pm 0.02390.3286 ±\pm 0.1643
ResNet18EC-ILRAMIL0.6325 ±\pm 0.06660.3602 ±\pm 0.1851
DINO (SSL)AdditiveMIL (ABMIL)0.6883 ±\pm 0.05590.3055 ±\pm 0.1131
DINO (SSL)AttriMIL0.7307 ±\pm 0.03120.8016 ±\pm 0.0932
DINO (SSL)EC-ABMIL0.7635 ±\pm 0.02720.3213 ±\pm 0.1122
DINO (SSL)EC-TransMIL0.7267 ±\pm 0.03930.3341 ±\pm 0.1191
DINO (SSL)EC-ILRAMIL0.7459 ±\pm 0.02530.4511 ±\pm 0.0595

With ResNet18 features, AdditiveMIL and AttriMIL achieve higher classification AUCs than the three EC variants, while WE achieves lower AUPCs (0.2629, 0.3286, and 0.3602) than their native attributions (0.3781 and 0.7922). With DINO features, all three EC variants achieve higher AUCs than AdditiveMIL, and EC-ABMIL and EC-ILRAMIL also exceed AttriMIL. AdditiveMIL has the lowest deletion AUPC of 0.3055; WE ranges from 0.3213 to 0.4511, all below AttriMIL at 0.8016. These results extend the intrinsic-method comparison to molecular prediction with two pathology feature representations, jointly reporting prediction and attribution performance.


W2: What Assumption C Constrains, How It Relates to EC-MIL, and Empirical Validation

(1) Our theory aims to explain why instance contributions can be learned using only bag labels. A single bag prediction constrains the sum of contributions, allowing instance-level errors to cancel each other out. EC-MIL shares local evidence across bags, so the same instance information is repeatedly supervised by bag labels under different companions. Its contribution decomposition is therefore jointly constrained across combinations. Assumption C characterizes a regime where contextual error cancellation is controlled, linking excess bag-level cross-entropy risk to instance decomposition error. This insight provides a theoretical explanation of how bag supervision can train meaningful instance contributions.

Regarding your point about changes in context weights, the true contributions themselves can also change with companions. Learning this variation is precisely the model's objective. When predicted contributions track changes in the true contributions, weights can change substantially while contribution errors remain small. Assumption C constrains the relative variation of this tracking error across contexts and is therefore compatible with the Context Stream substantially adjusting weights according to instance relations.

Reviewer_Gd12OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose a new more interpretable architecture for multiple instance learning called Evidence-Context Disentangled MIL (EC-MIL). The main idea is to separate the computation inside the MIL model into two branches: one evidence branch which is an independent MLP transformation of instance-level representations and a context branch which assigns each instance a positive weight alpha. The class is then predicted by multiplying the individual evidence scores with the context weight. The method is evaluated on a synthetic MNIST-based dataset, the Camelyon16 dataset (lymph node metastasis prediction), and the MUT dataset which is introduced in this paper as a new mutation-bag dataset for tissue-of-origin prediction. The authors claim that EC-MIL maintains competitive predictive performance, while WE performs well in evidence recovery and instance-removal faithfulness at lower attribution cost than posthoc alternatives.

Critical strengths
  • The motivation of the paper is clear: Better explainability of MIL models is an important topic and many current works try to approach this with post-hoc methods. Integrating this directly into the design of the model is an interesting area of research.

  • The paper is well written. The method is simple and easy to understand.

  • The method is evaluated on multiple datasets and compared to several baselines across multiple MIL base architectures (e.g. TransMIL and ABMIL).

Critical weaknesses
  • The "Weighted Evidence" calculation assumes that the context only affects the reliability of an instance (the weight α), not the interpretation of the instance itself. If a feature is ambiguous until seen next to another feature, EC-MIL cannot "fix" that ambiguity via the context stream. It will likely misclassify the ambiguous patch as negative evidence. This is acknowledged with the bounded contextual effects assumption but it is a rather strong assumption which I’m not sure is true in a lot of real-world histopathology slide-level tasks.

  • For the application of Whole Slide Image analysis, a rather simple task is considered (Camelyon 16) which does not really need any interaction between the different instances. Each patch can be independently classified on each own. That’s why the Patch Camelyon dataset exists. This simplistic dataset might not show whether the method really keeps the same performance on datasets with more complex prediction targets. Adding more difficult tasks with more complex relationships (e.g. survival prediction, mutation prediction) with a histopathology patch foundation model as an encoder would significantly strengthen this paper. This xMIL-LRP follow-up contains some more complex tasks [1] and [2] contains a wide range of possible slide-level tasks that could be tested. Achieving similar performance as a TransMIL model on these kind of tasks would be quite convincing.

[1] Idaji, Mina Jamshidi, et al. "Beyond attention heatmaps: How to get better explanations for multiple instance learning models in histopathology." Medical Image Analysis (2026): 104148.

[2] Zhang, Andrew, et al. "Accelerating data processing and benchmarking of ai models for pathology." arXiv preprint arXiv:2502.06750 (2025).

Questions for authors
  • Why did you not include the XMIL-LRP results in the plots shown in Figure 3 and 5 and only in the tables in the appendix? Including only the other methods make the method look significantly stronger compared to other works although XMIL-LRP is often on par or better in the tables.

  • Did you consider to allow the weight to be negative? That would allow more complex interactions based on the context.

8 comments

Thank you for recognizing the motivation for intrinsic MIL explainability, the clarity of our method, and the breadth of our evaluation. Below, we respond to your concerns and questions point by point.

W1 & Q2: Local Evidence, Contextual Modulation, and Nonnegative Weights

(1) Thank you for raising this insightful question regarding ambiguous instances and the role of context. We would like to clarify that the Evidence Stream outputs signed class evidence rather than an independent patch classification decision, and both streams are trained jointly under bag-level supervision. The Evidence Stream extracts the local task-relevant information carried by an instance, while the Context Stream modulates how strongly that information is used based on the entire bag. Because they jointly determine the bag prediction, the sign of the local evidence does not equate to a standalone patch-level classification.

Figure 4(a) in the original manuscript illustrates this mechanism: when keeping a digit-1 instance fixed and replacing a companion digit 7 with digit 2 to form a class-defining pair (1,2)(1, 2), the instance's local evidence remains unchanged at 3.3943, whereas its context weight increases from 0.0753 to 0.9079, flipping the bag prediction from negative to positive. Figure 4(b), Figure 3(d–f), and Appendix Table A13 in the original manuscript also provide lesion-localization results on real histopathology data, showing how local evidence and contextual modulation jointly contribute to prediction and attribution.

(2) Regarding your question about allowing negative weights, we did consider this option and deliberately chose nonnegative weights to preserve clear attribution semantics. Nonnegative weights ensure that the direction (sign) of WE is determined locally by the instance’s own evidence, while the context modulates its magnitude. Allowing negative weights would allow the contribution of the same local evidence to reverse sign across contexts. Our design is motivated by histopathology: the intrinsic morphological features of a local tissue patch remain unchanged regardless of its companions, whereas its relevance or importance to the slide-level diagnosis varies with context. Thus, nonnegative weights reflect a deliberate design choice to preserve a consistent local evidence direction while modeling contextual importance.


W2: More Complex Histopathology Prediction Tasks and Pathology Foundation Model Features

(1) We would like to clarify that our original manuscript already utilized self-supervised (SSL) features extracted from a pathology-pretrained DINO model, with the source cited. As stated in the WSI paragraph of Section 5.1 (Experimental Setup) and the WSI Appendix, our SSL features are extracted from a DINO model pretrained on 36,666 histopathology WSIs (Caron et al., 2021; Kang et al., 2023) and were evaluated on both CAMELYON16 and BRACS. We used the pathology-pretrained checkpoint provided by Kang et al. (2023); we have updated Section 5.1 and the Appendix in the revised manuscript (highlighted) to make this pathology-specific pretraining even more explicit.

Caron et al. Emerging Properties in Self-Supervised Vision Transformers. ICCV, 2021.

Kang et al. Benchmarking Self-Supervised Learning on Diverse Pathology Datasets. CVPR, 2023.

(2) Thank you for suggesting the inclusion of more complex slide-level prediction targets. In this rebuttal, we added a TP53 mutation prediction experiment on TCGA-LUAD, extending our WSI evaluation from histopathological diagnosis to molecular status prediction. Predicting TP53 mutation status directly from WSIs addresses your suggestion of testing mutation prediction. Following our existing protocol, we evaluate both ResNet18 and the pathology-pretrained DINO (SSL) features with frozen extractors and identical data splits, comparing paired original and EC variants across ABMIL, TransMIL, and ILRAMIL, alongside AdditiveMIL and AttriMIL. Classification and deletion faithfulness results are reported in Rebuttal Tables 1 and 2 below, directly testing whether EC-MIL (including EC-TransMIL) maintains competitive performance on complex molecular prediction targets. In the revised manuscript, Table 2 reports backbone-averaged classification performance for the original and EC variants, Figure 7 shows deletion faithfulness, and Appendix J Tables A20–A21 provide the complete per-model results.

Rebuttal Table 1: Classification performance on TCGA-LUAD TP53 mutation prediction. DINO (SSL) denotes the pathology-pretrained DINO features described above.

FeatureModelACC ↑\uparrowF1 ↑\uparrowAUC ↑\uparrow
ResNet18ABMIL0.6179 ±\pm 0.01850.6109 ±\pm 0.02360.6660 ±\pm 0.0459
ResNet18EC-ABMIL0.5582 ±\pm 0.02600.5223 ±\pm 0.09250.5867 ±\pm 0.0418
ResNet18TransMIL0.5209 ±\pm 0.04150.4425 ±\pm 0.09400.5953 ±\pm 0.0440
ResNet18EC-TransMIL0.5463 ±\pm 0.05070.5141 ±\pm 0.08910.5572 ±\pm 0.0239
ResNet18ILRAMIL0.5358 ±\pm 0.06850.4514 ±\pm 0.13520.5911 ±\pm 0.0578
ResNet18EC-ILRAMIL0.5896 ±\pm 0.04930.5537 ±\pm 0.11230.6325 ±\pm 0.0666
ResNet18AdditiveMIL0.5940 ±\pm 0.03080.5727 ±\pm 0.05880.6403 ±\pm 0.0353
ResNet18AttriMIL0.5970 ±\pm 0.03020.5930 ±\pm 0.03520.6404 ±\pm 0.0298
DINO (SSL)ABMIL0.6642 ±\pm 0.01950.6637 ±\pm 0.01940.7637 ±\pm 0.0247
DINO (SSL)EC-ABMIL0.6955 ±\pm 0.03640.6873 ±\pm 0.04490.7635 ±\pm 0.0272
DINO (SSL)TransMIL0.6507 ±\pm 0.02470.6454 ±\pm 0.03170.7389 ±\pm 0.0162
DINO (SSL)EC-TransMIL0.6269 ±\pm 0.03090.6239 ±\pm 0.03130.7267 ±\pm 0.0393
DINO (SSL)ILRAMIL0.6761 ±\pm 0.02930.6736 ±\pm 0.02980.7506 ±\pm 0.0185
DINO (SSL)EC-ILRAMIL0.6716 ±\pm 0.04330.6678 ±\pm 0.04720.7459 ±\pm 0.0253
DINO (SSL)AdditiveMIL0.6299 ±\pm 0.05140.6195 ±\pm 0.05500.6883 ±\pm 0.0559
DINO (SSL)AttriMIL0.6418 ±\pm 0.03630.6359 ±\pm 0.04300.7307 ±\pm 0.0312

Rebuttal Table 2: Deletion faithfulness on TCGA-LUAD TP53 mutation prediction. Lower AUPC (↓\downarrow) indicates better faithfulness.

ModelAttribution MethodResNet18: AUPC ↓\downarrowDINO (SSL): AUPC ↓\downarrow
ABMILRandom0.7941 ±\pm 0.05130.8177 ±\pm 0.0551
ABMILLOO0.6422 ±\pm 0.02700.6860 ±\pm 0.0297
ABMILGI0.6263 ±\pm 0.03680.6199 ±\pm 0.0748
ABMILIG0.5849 ±\pm 0.04420.5952 ±\pm 0.0832
ABMILxMIL-LRP0.3015 ±\pm 0.07510.4367 ±\pm 0.0588
EC-ABMILWE0.2629 ±\pm 0.25590.3213 ±\pm 0.1122
TransMILRandom0.6919 ±\pm 0.07500.7973 ±\pm 0.0709
TransMILLOO0.6551 ±\pm 0.07360.7252 ±\pm 0.0486
TransMILGI0.6559 ±\pm 0.07370.6598 ±\pm 0.0525
TransMILIG0.6000 ±\pm 0.06470.5677 ±\pm 0.0443
TransMILxMIL-LRP0.5756 ±\pm 0.06920.4730 ±\pm 0.0785
EC-TransMILWE0.3286 ±\pm 0.16430.3341 ±\pm 0.1191
ILRAMILRandom0.6444 ±\pm 0.08940.8407 ±\pm 0.0406
ILRAMILLOO0.5783 ±\pm 0.04740.7329 ±\pm 0.0475
ILRAMILGI0.6608 ±\pm 0.11850.6187 ±\pm 0.0279
ILRAMILIG0.5489 ±\pm 0.05370.5488 ±\pm 0.0210
EC-ILRAMILWE0.3602 ±\pm 0.18510.4511 ±\pm 0.0595
AdditiveMILRandom0.7449 ±\pm 0.02620.7511 ±\pm 0.0872
AdditiveMILLOO0.6958 ±\pm 0.01530.7142 ±\pm 0.0871
AdditiveMILGI0.6125 ±\pm 0.05020.3700 ±\pm 0.1193
AdditiveMILIG0.5939 ±\pm 0.04070.5060 ±\pm 0.2493
AdditiveMILPatch-Logits0.3781 ±\pm 0.07280.3055 ±\pm 0.1131
AttriMILRandom0.8020 ±\pm 0.06030.8437 ±\pm 0.0778
AttriMILLOO0.6757 ±\pm 0.05370.5925 ±\pm 0.0336
AttriMILGI0.7419 ±\pm 0.07750.8755 ±\pm 0.0700
AttriMILIG0.7214 ±\pm 0.07060.8871 ±\pm 0.0670
AttriMILNative (AttriMIL)0.7922 ±\pm 0.06670.8016 ±\pm 0.0932

With pathology-pretrained DINO features, the molecular prediction AUCs of the EC variants remain within 0.0122 of their original backbones; EC-TransMIL and TransMIL achieve 0.7267 and 0.7389, respectively. With ResNet18 features, EC-ILRAMIL improves AUC from 0.5911 to 0.6325, whereas the other two EC variants have lower AUCs than their original models. In deletion evaluation, WE achieves the lowest mean AUPC within each EC predictor under both feature settings, and lower AUPC than xMIL-LRP on the corresponding original ABMIL and TransMIL predictors. This experiment jointly evaluates molecular prediction and attribution faithfulness, showing how EC-MIL performs on a more complex pathology target with different feature representations.


Q1: Including xMIL-LRP Results in Figures 3 and 5

(1) In the original manuscript, Figures 3 and 5 were designed to compare different attribution methods evaluated on the exact same EC predictor, whereas xMIL-LRP is applied to the original backbones; hence, xMIL-LRP results were reported in the comprehensive Appendix tables. Specifically, WE, GI, IG, LOO, and Random in those figures all explain the same trained EC-MIL model. Because the xMIL-LRP implementation supports the original ABMIL and TransMIL architectures, we reported its results in Appendix Tables A11 and A13–A15, and explicitly noted in the main text of the original manuscript (Section 5.4) that original ABMIL with xMIL-LRP attains higher localization AUPRC-2 on CAMELYON16 in both feature settings.

(2) We appreciate your feedback and agree that plotting xMIL-LRP directly in Figures 3 and 5 provides a more complete visual comparison. We have updated Figures 3 and 5 in the revised manuscript to include the xMIL-LRP results from the original manuscript's Appendix. Both figures retain all original attribution results and standard deviation error bars, with legends and captions (highlighted in the revision) clearly distinguishing xMIL-LRP on the original ABMIL/TransMIL backbones from WE, GI, IG, LOO, and Random on the EC predictors. Additionally, the revised Figure 3 now covers four synthetic tasks, and both Figure 3 and Figure 5 incorporate the native attributions of AdditiveMIL and AttriMIL. The full numerical tables remain in the Appendix for comparison across tasks, feature settings, and backbones, while distinguishing attribution comparisons within the same predictor from those across different predictors.

Thank you again for your constructive feedback. Please let us know if you have any further questions; we would be happy to continue the discussion.

Reviewer_CbfJOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EC-MIL, a multiple instance learning framework that separates instance-local signed class evidence from bag-dependent context weighting. The Evidence Stream computes class-specific evidence for each instance independently, while the Context Stream uses a MIL backbone to produce nonnegative normalized weights. Their product gives Weighted Evidence (WE), which is signed, class-specific, complete with respect to the class logit, and available in a single forward pass. The authors also show under explicit assumptions that excess bag-level cross-entropy risk bounds instance-contribution decomposition error, using the Efron-Stein inequality to limit error cancellation across instances. Experiments on synthetic MIL, whole-slide image datasets CAMELYON16 and BRACS, and a new MUT mutation-bag dataset show competitive predictive performance, better signed evidence recovery and deletion faithfulness, and lower attribution cost than standard attention and post-hoc baselines. On MUT, WE also identifies gene-level signals consistent with known tissue-gene associations.

Critical strengths

A key strength is the formal link between bag-level risk and instance-level attribution. Under explicit assumptions, the authors show that excess bag-level cross-entropy risk bounds attribution decomposition error, using the Efron-Stein inequality to limit cancellation across instances. This addresses a central MIL issue where accurate bag predictions can hide unreliable instance contributions, though the result depends on assumptions such as bounded contextual effects.

A second strength is efficient signed attribution in a single forward pass. By multiplying instance-level class logits with scalar context weights, EC-MIL distinguishes supportive from opposing evidence without post-hoc gradients or perturbations. This reduces attribution cost and preserves sign information, but the method requires a dedicated architecture and cannot directly explain fixed pretrained MIL classifiers.

Critical weaknesses

The main difference from instance-based attention MIL such as IAMIL (Gao et al., 2025) is that EC-MIL decouples instance evidence from the context weights, allowing the Context Stream to use a more complex attention or backbone. This is a useful modularity, but it is not a new aggregation mechanism: however complex the context encoder, its output is still reduced to one nonnegative scalar weight per instance. The paper does not isolate whether gains come from this decoupling or simply from a stronger context encoder. Moreover, Adjacent-Pairs only requires magnitude modulation with fixed evidence direction, and a single pairwise-adjacency task does not establish general context modeling, since ternary, angular, or triangle relations may remain unrepresentable under the scalar-weight bottleneck.

The theoretical analysis links bag-level supervision to instance-level attribution, but its assumptions limit how much the guarantee establishes. The bound relies on conditional independence, low sensitivity to single-instance replacement, additive decomposition of the true bag log-odds, and bounded contextual effects that cannot flip an instance's sign. These conditions sit uneasily with WSI, where spatial continuity breaks independence, and with MUT, where pathway mutations are correlated, and relation-defined labels may not decompose additively. The paper does not test whether they hold in practice, and the synthetic benchmark appears to be designed in the regime where the bounded-context assumption is satisfied.

Reference:

Gao, Z., Mao, A., Dong, Y., Wu, J., Liu, J., Wang, C., He, K., Gong, T., Li, C., & Crispin-Ortuzar, M. (2025). SMMILe enables accurate spatial quantification in digital pathology using multiple-instance learning. Nature Cancer. https://doi.org/10.1038/s43018-025-01060-8

Questions for authors

Q1. Since EC-MIL mainly allows a more complex attention or context encoder than IAMIL, can the authors compare EC-MIL with IAMIL under matched context-encoder capacity, and separate the effect of decoupling from the effect of a stronger backbone? Testing harder relations, such as three instances interacting or spatial angles would show what the decoupling actually adds.

Q2. Is the synthetic benchmark constructed precisely to satisfy the theoretical assumptions, and do the authors have evidence on attribution quality when these assumptions are violated in WSI or MUT? Without such a test, the bound remains an idealized sufficient condition rather than a practical guarantee.

10 comments

Thank you for recognizing the formal link between bag-level risk and instance-level attribution, as well as the efficiency of single-forward-pass signed attribution. Below, we address your concerns and questions point by point.

W1-1: Methodological Differences from IAMIL and Contributions of EC-MIL

We disagree that the contribution of EC-MIL is primarily to allow a more complex context encoder. IAMIL is a framework for instance-level prediction and attention aggregation. EC-MIL uses separate evidence and context streams to output instance contributions alongside predictions (Figure 1 in the original manuscript). Its contributions concern both the predicted quantities and the information flow, as detailed below.

(1) IAMIL aggregates instance classification probabilities, whereas EC-MIL decomposes class logits.

In Eqs. (4) and (8) of Gao et al. (2025) cited in your review, instance classification outputs are first converted into probabilities picp_{ic} via sigmoid or softmax, and then aggregated using attention weights aic(B)a_{ic}(B). In contrast, EC-MIL aggregates real-valued evidence Ec(xi)E_c(x_i) to form the class logit before applying softmax. Formally:

\text{IAMIL:}\quad &P_c(B)=\sum_i a_{ic}(B)p_{ic},\\\\ \text{EC-MIL:}\quad &z_c(B)=\sum_i\alpha_i(B)E_c(x_i), \qquad \mathrm{WE}_i^c=\alpha_i(B)E_c(x_i). \end{aligned}$$ The two architectures decompose different predicted quantities: $a_{ic}p_{ic}$ in IAMIL is a nonnegative probability score, whereas WE in EC-MIL is a signed (positive or negative) contribution that sums *exactly* to the class logit. Because nonlinear activations and summation generally do not commute, this distinction is structurally determined by the prediction formulation, leading to different attribution semantics. **(2) EC-MIL’s separation of local evidence and contextual interactions explicitly governs how context influences instance contributions.** We preserve the independence of local evidence even when using interaction-based backbones. In the full SMMILe framework (Gao et al., 2025), both the classifier and the attention detector read from the same compressed features; in configurations utilizing spatial convolutions, both are influenced by the neighborhood. In EC-MIL, the Evidence Stream processes each instance independently, while the contextual representations from TransMIL or ILRAMIL are used exclusively to compute the class-agnostic weights $\alpha_i(B)$. Consequently, an instance's evidence remains invariant across different companions, while the context modulates its contribution via the weights. This allows predictions and signed, class-specific, logit-complete explanations to be obtained in a single forward pass.

W1-2 / Q1 (Capacity Matching): Distinguishing the Effects of Evidence-Context Decoupling and Context-Encoder Capacity

(1) To address whether the gains stem from decoupling or a stronger encoder, we examine paired implementations using the same underlying backbones and compare attributions on fixed predictors.

We examine the source of gains through backbone-paired architectures and attribution comparisons on fixed models. Figure 2 and Appendix Tables A8–A10 in the original manuscript compare the original and EC variants of three backbones across synthetic MIL, WSI, and MUT. On Adjacent-Pairs, TransMIL / EC-TransMIL achieve AUCs of 0.9831 / 0.9878, and ILRAMIL / EC-ILRAMIL achieve 0.9930 / 0.9910 (original Appendix Table A8), showing that the advantages of interaction modeling are preserved. Table 1 in the original manuscript also shows that, with ABMIL, EC-ABMIL achieves higher mean AUC and AUPRC-2 than AdditiveMIL on all three synthetic tasks; for example, Adjacent-Pairs AUPRC-2 is 0.8000 vs. 0.6027. On the same EC predictor, WE, GI, IG, and LOO are evaluated with identical model parameters and prediction outputs. WE achieves the highest mean AUPRC-2 in eight of nine settings (Figure 3(a–c) and Appendix Table A11 in the original manuscript). These results show that attribution gains also occur with a simple backbone and that attribution readouts differ even when the predictor is fixed.

(2) To evaluate more complex interactions and their impact on prediction and explanation, we designed and conducted an additional experiment on a new task, Pos-Neg-L. Task details are provided in Appendix I of the revised manuscript, and results are integrated into the revised Figure 3 and unified synthetic comparison tables.

To test three-instance spatial relationships, we introduce Pos-Neg-L: digits 1, 2, and 3 each contribute +1; digits 4, 5, and 6 each contribute −1; and digits 0, 7, 8, and 9 contribute 0. When digits 1, 2, and 3 form an L shape with 2 at the corner, each participating instance contributes +2. The bag label is determined by the sum of instance contributions, allowing three-instance spatial relations to modulate contributions and affect classification. Under this setup, we compare paired original and EC variants across all three backbones, alongside AdditiveMIL and AttriMIL, jointly evaluating classification and attribution. This allows relational modeling capacity and explanation quality to be benchmarked under the same setting.

Rebuttal Table 1: Classification performance on Pos-Neg-L (Revised Table A8). Bold values indicate the best mean across the table for each metric; Δ\DeltaAUC denotes the improvement of the EC variant over its original backbone in percentage points.

ModelACC ↑\uparrowF1 ↑\uparrowAUC ↑\uparrowΔ\DeltaAUC (pp)
ABMIL0.7624 ±\pm 0.00710.7473 ±\pm 0.01110.8670 ±\pm 0.0078—
EC-ABMIL0.7656 ±\pm 0.00630.7543 ±\pm 0.00830.8777 ±\pm 0.0065+1.07
TransMIL0.9112 ±\pm 0.01020.9081 ±\pm 0.01040.9728 ±\pm 0.0025—
EC-TransMIL0.9336 ±\pm 0.00510.9317 ±\pm 0.00540.9849 ±\pm 0.0023+1.21
ILRAMIL0.7756 ±\pm 0.00790.7608 ±\pm 0.01010.8924 ±\pm 0.0065—
EC-ILRAMIL0.7780 ±\pm 0.00370.7678 ±\pm 0.00800.8939 ±\pm 0.0067+0.15
AdditiveMIL0.7546 ±\pm 0.01010.7446 ±\pm 0.00830.8660 ±\pm 0.0066—
AttriMIL0.7658 ±\pm 0.00720.7572 ±\pm 0.00710.8798 ±\pm 0.0058—

Rebuttal Table 1 shows that all three EC variants achieve higher mean ACC, F1, and AUC than their respective original backbones. TransMIL leads in both original and EC configurations, with AUC improving from 0.9728 to 0.9849. Thus, the EC architecture preserves the backbone's advantage on this spatial relational task.

To simultaneously evaluate explanation quality, we compare WE on EC models against LOO, GI, IG, and xMIL-LRP (where supported) on the original backbones, as well as the native attributions of AdditiveMIL and AttriMIL. Rebuttal Table 2 reports both AUPRC-2 and deletion AUPC to separately assess evidence recovery and predictive faithfulness.

Rebuttal Table 2: Evidence recovery and deletion faithfulness on Pos-Neg-L (Revised Tables A11–A12). Bold values indicate the highest AUPRC-2 and lowest AUPC; values in parentheses for EC-TransMIL denote absolute changes relative to xMIL-LRP on the original TransMIL.

ModelAttribution MethodAUPRC-2 ↑\uparrowAUPC ↓\downarrow
ABMILLOO0.9295 ±\pm 0.00620.3747 ±\pm 0.0101
ABMILGI0.9401 ±\pm 0.00660.3452 ±\pm 0.0065
ABMILIG0.9392 ±\pm 0.00670.3412 ±\pm 0.0079
ABMILxMIL-LRP0.9185 ±\pm 0.01180.3393 ±\pm 0.0161
TransMILLOO0.7184 ±\pm 0.02190.3969 ±\pm 0.0063
TransMILGI0.4961 ±\pm 0.07900.6319 ±\pm 0.0640
TransMILIG0.9050 ±\pm 0.02400.3699 ±\pm 0.0214
TransMILxMIL-LRP0.9110 ±\pm 0.01960.3511 ±\pm 0.0199
ILRAMILLOO0.8677 ±\pm 0.02700.4285 ±\pm 0.0184
ILRAMILGI0.4907 ±\pm 0.03620.7164 ±\pm 0.0278
ILRAMILIG0.9329 ±\pm 0.01680.3599 ±\pm 0.0095
AdditiveMILPatch-Logits0.9403 ±\pm 0.00320.3251 ±\pm 0.0087
AttriMILAttriMIL0.8331 ±\pm 0.04640.4485 ±\pm 0.0337
EC-ABMILWE0.9824 ±\pm 0.00320.3471 ±\pm 0.0092
EC-TransMILWE0.9743 ±\pm 0.0043 (+0.0633)0.3085 ±\pm 0.0076 (-0.0426)
EC-ILRAMILWE0.9773 ±\pm 0.00490.3514 ±\pm 0.0104

In Rebuttal Table 2, all three WE variants achieve higher mean AUPRC-2 than the listed attribution methods on the original models. WE achieves 0.9743–0.9824, compared with a maximum of 0.9403 among the listed original-model methods. EC-TransMIL simultaneously achieves the highest classification AUC (0.9849) and the lowest deletion AUPC (0.3085). Relative to xMIL-LRP on the original TransMIL, its AUPRC-2 increases by 0.0633 and its AUPC decreases by 0.0426. Thus, the paired comparison within the same backbone shows strong performance in both relational prediction and instance-level explanation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.