acceptodds
Under review as a conference paper at ICLR 2027

Uncertainty Estimation in Pathology Foundation Models via Deep Mutual Learning

Abstract

Pathology foundation models (PFMs) offer generalizable representations for whole-slide image (WSI) analysis, yet their clinical adoption remains limited. Specifically, their predictions lack reliable confidence estimates, and no single PFM is universally best across tasks, which severely undermines trust in medical settings. To overcome this, we propose DICE, a plug-and-play framework that ensembles frozen PFMs and estimates uncertainty based on their consensus. We align the ensemble members via deep mutual learning and theoretically show that this objective controls an upper bound on epistemic uncertainty. Additionally, we demonstrate that the ensemble localizes abnormalities at the patch level without any explicit supervision. We evaluate DICE on three challenging WSI benchmarks. Notably, our framework provides reliable uncertainty estimates that accurately flag failure-prone cases under in- and out-of-distribution settings, while matching or outperforming SOTA baselines in classification, calibration, and localization. Overall, DICE takes a crucial step toward translating PFMs into uncertainty-aware decision-support systems.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
uncertainty estimation, digital 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_9e8JOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the authors introduce the DICE approach that fuses five frozen pathology foundation models with distinct trainably learnable MIL heads. In the training process, the method utilizes supervised classification, pairwise deep mutual learning, and Gramian volume regularization of projected slide embedding. Predictive entropy based ensemble method is employed for unreliable sample detection, and averaged attention score for localization.

Critical strengths
  1. The uncertainty evaluation contains valuable practical checks. The paper considers retained errors as uncertain slides are delayed, applies validation-selected rejection thresholds to test data, and examines both CAMELYON cohort transfer directions. The conclusion reached is that entropy of the ensemble can detect potentially error-prone cases.

  2. There is a valuable empirical improvement on the largest multiclass problem. On PANDA, the submission presents F1 score of 76.6% for the full DICE, while the best individual PFM has 72.8%, and early fusion 73.6%. NLL also gets better. The paper uses various frozen encoders and considers feature extraction cost, thus making extra computations evident.

Critical weaknesses
  1. The theoretical bound does not lead to uncertainty estimation improvements. Proposition 2 derives from the convexity property of KL divergence and upper-bounds the ensemble Jensen-Shannon disagreement with the KL divergence between two experts. However, making the models agree may also make them confidently wrong about something. Minimizing this measure does not prove better detection of epistemic uncertainty or the error under shift. In addition, the uncertainty estimations do not consider simple late fusion of the same five PFMs, even though this baseline is feasible for classification tasks. Without this comparison, it remains unclear whether the proposed mutual learning improves over heterogeneous ensembling.

  2. There is no justification provided for the Gramian penalty's importance and its interpretation. The Gramian penalty is zero if there is any linear dependency among the projected expert vectors. It does not mean that all experts agree on anything or that the diagnostic property of this representation is preserved.

  3. No matched heterogeneous ensemble baseline for uncertainty analysis is provided. Late fusion of the same five PFMs without uncertainty analysis is evaluated for classification but is not included in the comparison of uncertainties. In the absence of the risk coverage curve, error detection score, and transferred threshold with matched tuning, the comparison experiments cannot disentangle the effect of mutual learning from the effect of heterogeneous ensembling.

  4. The benefits of the entire problem setting are inconsistent across cohorts. With Gramian regularization removed, the improvement of F1 becomes negligible compared to late fusion on PANDA, suggesting that most of the gain comes from the penalty. However, on CAMELYON17, the full DICE obtains an even lower F1 than both without penalty and the best single PFM.

Questions for authors
  1. How do typical late fusions of the same five PFMs behave relative to each other on the curves for risk/coverage, error detection measures, validation-selected deferral, and cohort shift?

  2. What are the empirical results showing that the smaller volume of the Gramian leads to meaningful alignment and not to redundancy or collapsed alignment?

  3. What is the empirical result that directly links the bound on the disagreement term to error detection and not to the overall entropy of the ensemble?

5 comments
AuthorsOpenReview ↗

Our Rebuttal to Reviewer 9e8J (W1a & W1b & Q1 & W3)

Thank you for reviewing the paper. We appreciate that you find the proposed framework practical and the empirical performance strong. We address your remaining comments below, which we hope will help clarify and improve your evaluation of the paper.

W1a: Relationship between the theoretical analysis and uncertainty estimation

Proposition 2 establishes how the DML objective controls an upper bound on ensemble Jensen–Shannon disagreement through the aggregate pairwise KL divergences between experts. We do not claim that this bound guarantees improved error detection, calibrated disagreement, or protection against confidently incorrect consensus.

The uncertainty measure used in our evaluations is total predictive entropy, which incorporates both the average entropy of individual predictions and their disagreement. Proposition 2 therefore provides a theoretical connection between the training objective and one component of the evaluated uncertainty measure.

Finally, we acknowledge that confidently incorrect consensus is a failure mode that DICE’s uncertainty scores may not detect. However, this failure mode is not specific to our method and exists for all the other reported baselines. In fact, as shown in Figure 2 (right), baseline methods assign misclassified slides uncertainty values comparable to those of their most certain correctly classified slides. This pattern appears less pronounced for both DICE variants, suggesting our method’s superiority.

W1b/Q1/W3: Heterogeneous late-fusion baseline

We agree that including this baseline in the uncertainty analysis is essential to assess the contribution of DML beyond heterogeneous ensembling. We have therefore added Late Fusion to Figures 2 and 3 in the main paper and Figures 6, 7, and 8 in the appendix. We highlighted these changes in blue in the revised manuscript and include the results of Late Fusion, DICE (w/o reg), and DICE in the tables below. These results clearly demonstrate the benefit of our DML loss in uncertainty analysis.

Figure 2, left — PANDA retained error rate (%) ↓

Method0% deferred20%40%60%80%
Late Fusion20.6514.047.642.020.50
DICE (w/o reg)20.6514.297.392.140.76
DICE18.4411.716.131.390.50

Figure 2, right, and Figure 3(b) — uncertainty separation ↑

MethodFigure 2: PANDA, Cliff’s ΔFigure 3(b): C16 → C17, Cliff’s Δ
Late Fusion0.630.65
DICE (w/o reg)0.630.66
DICE0.650.64

All listed separation comparisons have p < 10⁻⁴.

Figure 3(a) — PANDA classification with rejection

SplitMethodF1 before rejection (%) ↑F1 after rejection (%) ↑F1 gain (pp) ↑
ValidationLate Fusion73.89 ± 1.4281.17 ± 1.99+7.28
ValidationDICE (w/o reg)73.54 ± 1.6680.19 ± 3.23+6.65
ValidationDICE77.58 ± 1.8486.23 ± 4.10+8.65
TestLate Fusion73.41 ± 0.8179.43 ± 1.13+6.03
TestDICE (w/o reg)73.62 ± 1.1179.41 ± 1.70+5.78
TestDICE76.64 ± 1.0286.49 ± 2.23+9.85
AuthorsOpenReview ↗

Our Rebuttal to Reviewer 9e8J (W2 & Q2 & W4 & Q3)

W2/Q2: Role of the Gramian term

The motivation for the Gramian term is stated in Section 3.2: it introduces representation-level alignment between slide embeddings, complementing the prediction-level coupling provided by DML. Its practical contribution is assessed through our experiments with and without this regularization. We agree that the Gramian penalty vanishes whenever the projected expert vectors are linearly dependent, however, we did not observe the Gramian loss reaching zero during training.

The Gramian term is optimized jointly with supervised classification and DML. The supervised objective encourages each expert to retain information relevant to the diagnostic task, while DML enables knowledge exchange through predictive distributions. Furthermore, the representations being aligned originate from the same slide, providing a shared semantic basis despite differences in the PFMs’ architectures and pretraining. The intended role of Gramian regularization is to encourage alignment across these different representations during task learning.

Finally, we argue that our Gramian objective leads to meaningful alignment, rather than a collapsed one, since DICE outperforms both deep ensembles (one model repeated five times) and late fusion (five-model ensembles) in classification accuracy and uncertainty estimation. We respectfully draw the reviewer’s attention to the fact that this strong empirical performance (Table 1, Figure 2, and Figure 3) shows successful alignment. For example, on PANDA (the most complex and largest dataset in our study), adding Gramian regularization improves F1 from 73.6% to 76.6%, together with improvements in NLL and Brier score.

W4: Clarification on empirical benefits

We respectfully disagree with the reviewer that our empirical benefits are inconsistent. Table 2 reports statistically significant improvements for DICE on PANDA and C16. On C17, however, DICE performs within fold-level variability with the best method, and hence, the slight underperformance is not statistically significant. We would like to draw your attention to the fact that DICE matches or outperforms the aforementioned baselines while also providing better uncertainty estimates.

Q3: Clarification on the theoretical result

Thank you for this question. Our experiments assess error detection using total predictive entropy, while Proposition 2 connects the DML objective to its epistemic component. We do not focus on evaluating a direct relationship between the disagreement term and error detection.

Our choice is motivated by prior medical-classification studies reporting favorable performance of total predictive uncertainty compared with its components or alternative uncertainty measures [1,2]. Furthermore, although the entropy decomposition is mathematically exact, interpreting its components as a faithful separation of aleatoric and epistemic uncertainty has empirically documented limitations [3]. These considerations motivate our focus on total predictive uncertainty for identifying unreliable predictions.

References

[1] Tchokponhoue, G. A. D., & Idri, A. (2026). On the value of uncertainty quantification in deep learning based breast cancer molecular subtype classification. Applied Soft Computing, 186, 114249.

[2] Tchokponhoue, G. A. D., & Idri, A. (2026). A comprehensive empirical evaluation of Bayesian evidential deep learning for uncertainty quantification of image-based medical classification. Results in Engineering, 30, 110453.

[3] Wimmer, L., Sale, Y., Hofman, P., Bischl, B., & Hüllermeier, E. (2023). Quantifying aleatoric and epistemic uncertainty in machine learning: Are conditional entropy and mutual information appropriate measures? In Proceedings of the Thirty-Ninth Conference on Uncertainty in Artificial Intelligence, Proceedings of Machine Learning Research, 216, 2282–2292.

We thank Reviewer 9e8J once again for the time and effort they put into our work. As the discussion period approaches its end, we would kindly like to ask if there are any further comments or concerns that we could address.

Reviewer_3679OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes DICE, a framework that ensembles frozen pathology foundation models for whole-slide image analysis. Each model is paired with a trainable MIL head, and the experts are jointly trained using supervised classification, deep mutual learning to align their predictions, and a Gramian loss to align their representations. During inference, DICE averages the experts’ predictions and uses predictive entropy and residual disagreement as uncertainty signals to predict slide misclassification. It also combines the experts’ attention maps for weakly supervised lesion localisation. Experiments on PANDA and CAMELYON16/17 show competitive results, probabilistic prediction, selective deferral, and localisation performance. The theoretical result shows that the DML objective controls an upper bound on ensemble disagreement. However, it does not prove that reducing disagreement makes the remaining disagreement a faithful measure of epistemic uncertainty. The paper shows a promising ensemble for uncertainty-aware pathology, but its calibration claims, uncertainty interpretation, and the necessity of the Gramian term require stronger justification.

Critical strengths
  • Clinically relevant problem: DICE addresses uncertainty and failure detection in computational pathology, which are important problems in medical AI.

  • Reasonable evaluation: The method is assessed on classification, uncertainty-based deferral, cross-cohort generalisation, and lesion localisation.

Critical weaknesses
  • Overstated calibration claims: Calibration is assessed using NLL, which combines accuracy and calibration, without sufficient analysis using ECE or reliability diagrams.

  • Contradictory treatment of disagreement: DICE uses disagreement as a proxy for epistemic uncertainty while DML explicitly trains the experts to produce similar predictions.

  • Unproven preservation of complementarity: Neither DML nor the Gramian loss distinguishes useful model diversity from spurious disagreement, so alignment may suppress the complementary information motivating the heterogeneous ensemble.

  • Unclear Gramian contribution: The representation-alignment term is poorly motivated and does not consistently improve classification, uncertainty estimation, or localisation across datasets.

  • Limited theoretical result: The theorem shows that DML contracts an upper bound on expert disagreement, but not that the residual disagreement is calibrated, or reflects true epistemic uncertainty.

Questions for authors
  1. Can the authors provide dedicated calibration results, e.g. ECE, Brier score, or reliability diagrams?

  2. How does varying the DML weight affect expert diversity, classification, calibration, and error-detection performance?

  3. Does DML improve uncertainty quality over independently trained heterogeneous late fusion?

  4. What does the Gramian loss contribute independently, and why does it help on some datasets but hurt on others?

  5. How does DICE behave on confidently incorrect consensus cases where all experts agree on the wrong class?

5 comments
AuthorsOpenReview ↗

Our Rebuttal to Reviewer 3679 (W1 & Q1 & W2 & W3)

We thank the reviewer for finding our setting relevant and our experiments rigorous. We address the remaining comments below, which we hope will improve their evaluation of our paper.

W1/Q1: Clarification on calibration claims

We thank the reviewer for this suggestion. We emphasize that, in addition to NLL, Table 5 in Appendix A.9 reports Brier scores for all methods. For example, on PANDA, DICE achieves a Brier score of 4.5 ± 0.15, compared with 4.9 ± 0.12 for heterogeneous late fusion, in the table’s ×100 units. We provide this table here for your reference and observe that DICE variants either outperform all the UQ baselines or stay within fold-level variability.

MethodPANDACAMELYON16CAMELYON17
Virchow25.12 ± 0.142.45 ± 0.3714.64 ± 1.28
UNI2-h5.51 ± 0.083.01 ± 1.2814.61 ± 2.04
H-optimus-15.41 ± 0.123.46 ± 0.4316.08 ± 1.54
CONCHv1.55.91 ± 0.063.67 ± 0.7015.38 ± 2.39
Hibou-L5.63 ± 0.045.28 ± 0.4515.53 ± 2.25
MC Dropout5.13 ± 0.152.97 ± 0.3915.52 ± 3.63
Deep Ensemble4.92 ± 0.072.42 ± 0.1114.78 ± 1.88
EDL6.80 ± 0.072.80 ± 0.4615.74 ± 1.52
SNGP5.04 ± 0.053.39 ± 0.2820.88 ± 1.23
DivDis5.34 ± 0.122.85 ± 0.6215.45 ± 3.43
Early Fusion4.96 ± 0.123.19 ± 1.0714.31 ± 1.99
Late Fusion4.90 ± 0.122.56 ± 0.2314.64 ± 1.71
DICE (w/o reg)4.99 ± 0.242.85 ± 0.2513.80 ± 0.60
DICE4.47 ± 0.152.66 ± 0.1515.45 ± 0.68

W2: Treatment of disagreement

We would like to clarify that the uncertainty measure used throughout our experiments is the ensemble’s total predictive entropy, which decomposes into the sum of aleatoric (data) and epistemic (model) uncertainty. Proposition 2 establishes a formal link between the training-time DML loss and this epistemic component. In the asymptotic limit, zero DML loss implies zero epistemic uncertainty (i.e., complete model agreement), meaning total predictive entropy reduces directly to data uncertainty. Consequently, minimizing disagreement during training does not create a contradiction with using total predictive entropy at test time. Rather, it regularizes spurious model disagreement while preserving the data uncertainty signal.

W3: Preservation of complementarity

Our motivation is to exploit the complementary knowledge provided by heterogeneous PFMs through collaborative learning. Differences in their architectures, self-supervised objectives, and pretraining data give rise to distinct representations of the same slide. We argue that our alignment objectives (DML and Gramian) succeed in distinguishing useful diversity from spurious disagreement since DICE variants outperform both deep ensembles (one model repeated five times) and late fusion (five-model ensembles without the alignment losses) in classification accuracy and uncertainty estimation. We respectfully draw the reviewer’s attention to the fact that this strong empirical performance (Table 1, Figure 2, and Figure 3) shows preservation of expert complementarity.

AuthorsOpenReview ↗

Our Rebuttal to Reviewer 3679 (W4 & W5 & Q2 & Q3)

W4: Clarification on the Gramian term

The motivation for the Gramian term is described in Section 3.2: it introduces slide-level representation alignment alongside the prediction-level alignment provided by DML. DML operates on class probabilities and does not directly constrain the geometry of the underlying slide embeddings. The Gramian term addresses this additional level of interaction by penalizing the volume jointly spanned by the projected representations of the same slide.

While the advantage of DICE was clear on CAMELYON16 and PANDA classification and calibration performance and globally in terms of UQ, we observe lower performance on CAMELYON17. One possible explanation for the degradation on CAMELYON17 is sensitivity to the strength and scheduling of Gramian regularization. Our hyperparameter search considered linear and cosine schedules within a prescribed weight range, but did not systematically explore a broader range of regularization strengths and scheduling configurations for each dataset. These choices determine how strongly and at what stage representation alignment influences task learning.

W5: Extent of the theoretical analysis

We would like to clarify the scope of Proposition 2 and its relationship to our uncertainty evaluation. The uncertainty measure used throughout our experiments is total predictive uncertainty, measured by the entropy of the ensemble’s averaged predictive distribution. We do not evaluate residual disagreement as a standalone uncertainty measure. Proposition 2 establishes that the DML objective controls an upper bound on epistemic uncertainty. Its contribution is therefore to formalize how prediction-level coupling relates to the epistemic component of the total uncertainty used in our framework.

Q2: Ablation on the DML weights

Thank you for this suggestion. We ablate the DML weight at 0, 0.5, 1, and 2 and report classification, calibration, and error detection performance in the table below for PANDA. We observe that along all these axes, our framework stays robust in the evaluated hyperparameter range.

MetricDML weight 0DML weight 0.5DML weight 1DML weight 2
Macro-F1 ↑72.33 ± 1.5273.37 ± 1.1973.62 ± 1.1173.15 ± 1.02
Error-detection AUROC ↑80.42 ± 0.7881.03 ± 0.8380.59 ± 1.1481.21 ± 0.39
AURC ↓7.68 ± 0.727.23 ± 0.437.32 ± 0.687.26 ± 0.30
NLL ×100 ↓55.26 ± 3.0353.61 ± 2.3054.59 ± 2.9755.19 ± 1.16
Brier ×100 ↓5.08 ± 0.254.92 ± 0.174.99 ± 0.245.02 ± 0.09

Q3: Heterogeneous late-fusion baseline

Yes it does. We agree that including this baseline in the uncertainty analysis is essential to assess the contribution of DML beyond heterogeneous ensembling. We have therefore added Late Fusion to Figures 2 and 3 in the main paper and Figures 6, 7, and 8 in the appendix. We highlighted these changes in blue in the revised manuscript and include the results of Late Fusion, DICE (w/o reg), and DICE in the tables below. These results clearly demonstrate the benefit of our DML loss in uncertainty analysis.

Figure 2, left — PANDA retained error rate (%) ↓

Method0% deferred20%40%60%80%
Late Fusion20.6514.047.642.020.50
DICE (w/o reg)20.6514.297.392.140.76
DICE18.4411.716.131.390.50

Figure 2, right, and Figure 3(b) — uncertainty separation ↑

MethodFigure 2: PANDA, Cliff’s ΔFigure 3(b): C16 → C17, Cliff’s Δ
Late Fusion0.630.65
DICE (w/o reg)0.630.66
DICE0.650.64

All listed separation comparisons have p < 10⁻⁴.

Figure 3(a) — PANDA classification with rejection

SplitMethodF1 before rejection (%) ↑F1 after rejection (%) ↑F1 gain (pp) ↑
ValidationLate Fusion73.89 ± 1.4281.17 ± 1.99+7.28
ValidationDICE (w/o reg)73.54 ± 1.6680.19 ± 3.23+6.65
ValidationDICE77.58 ± 1.8486.23 ± 4.10+8.65
TestLate Fusion73.41 ± 0.8179.43 ± 1.13+6.03
TestDICE (w/o reg)73.62 ± 1.1179.41 ± 1.70+5.78
TestDICE76.64 ± 1.0286.49 ± 2.23+9.85
AuthorsOpenReview ↗

Our Rebuttal to Reviewer 3679 (Q4 & Q5)

Q4: Role of the Gramian regularization

DICE with Gramian regularization improves classification and calibration performance on PANDA and CAMELYON16 relative to DICE without this regularization, while providing better or comparable uncertainty quality. Across these two datasets, it also generally outperforms the other evaluated models. However, these gains do not consistently extend to CAMELYON17. One possible explanation is the interaction between representation alignment and dataset heterogeneity. CAMELYON17 includes slides from five medical centers, compared with two for CAMELYON16, introducing greater variation in tissue preparation, staining, and scanning. This motivates examining whether the strength and scheduling of alignment require different settings across datasets.

Q5: Confidently wrong consensus

We acknowledge that confidently incorrect consensus is a failure mode that DICE’s uncertainty scores may not detect. However, this failure mode is not specific to our method and exists for all the other reported baselines. In fact, as shown in Figure 2 (right), most UQ baselines assign uncertainty values to incorrect predictions that are as low as those observed for the most confident correct predictions. Neither DICE variant exhibits similarly extreme low-uncertainty errors in these plots. This observation suggests that DICE variants are less prone to such overconfident errors in the evaluated settings.

Reviewer_pDf2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DICE, a framework for uncertainty estimation using multiple frozen pathology foundation models. The authors combine deep mutual learning and Gramian regularization to align expert predictions and representations. They also use ensemble attention scores for patch-level lesion localization. Experiments on PANDA, CAMELYON16, and CAMELYON17 evaluate classification, calibration, uncertainty estimation, and localization.

Critical strengths

The paper addresses an important problem in computational pathology, where different foundation models have different strengths and reliable uncertainty estimates are needed. Using several pretrained models to identify uncertain predictions is a reasonable approach.

The paper evaluates the method on multiple WSI datasets and considers classification, calibration, and uncertainty-based rejection. The results on PANDA are promising, and the cross-cohort experiments provide useful evidence about uncertainty estimation under distribution shift.

Critical weaknesses

The theoretical analysis does not fully support the uncertainty estimation claims. Reducing disagreement between experts does not necessarily improve uncertainty estimation, since experts may agree on incorrect predictions.

The uncertainty experiments lack a direct comparison with heterogeneous late fusion using the same five PFMs. Without this baseline, it is difficult to determine whether the improvement comes from DML or simply combining different foundation models.

The role of Gramian regularization is not sufficiently justified. Minimizing the Gram determinant does not guarantee meaningful alignment between expert representations, and its contribution to uncertainty estimation remains unclear.

The lesion localization method mainly averages attention scores from different experts. Its connection to the proposed uncertainty learning framework is unclear, and the results do not consistently outperform simpler fusion methods.

Questions for authors

Please refer to the Weaknesses.

3 comments
AuthorsOpenReview ↗

Our Rebuttal to Reviewer pDf2 (W1 & W2)

We thank the reviewer for finding our approach reasonable and our experiments extensive. We address their remaining comments below, which we hope will improve their evaluation of our paper.

W1: Relationship between the theoretical analysis and uncertainty estimation

We would like to clarify the scope of the theoretical result. Proposition 2 states that the aggregate DML objective upper-bounds the ensemble disagreement (epistemic term) defined in Section 3.3. Our claim concerns this mathematical relationship between the training objective and the uncertainty decomposition.

The observation that experts may agree on an incorrect prediction is compatible with this result: the proposition bounds disagreement, rather than prediction error. Such a case therefore does not contradict the bound. Likewise, we do not claim that consensus guarantees correctness or that the bound alone establishes improved uncertainty calibration.

Sections 4.2.1 and 4.2.2, on the other hand, empirically show that the predictive entropy (total uncertainty) of the ensemble trained under the DICE objective better identifies failure-prone cases and generalizes across cohorts. Proposition 2 therefore connects the epistemic component of the reported uncertainty to the training-time DML loss. However, we do not claim that our theoretical bound is responsible for the empirical success of our method.

Finally, we acknowledge that confidently incorrect consensus is a failure mode that DICE’s uncertainty scores may not detect. However, this failure mode is not specific to our method and exists for all the other reported baselines. In fact, as shown in Figure 2 (right), baseline methods assign misclassified slides uncertainty values comparable to those of their most certain correctly classified slides. This pattern appears less pronounced for both DICE variants, suggesting our method’s superiority.

W2: Heterogeneous late-fusion baseline

The matched heterogeneous late-fusion baseline is defined in Appendix A.3.3 and evaluated in Table 1. It uses the same five frozen PFMs as DICE, with independently trained MIL heads whose predictive probabilities are averaged at inference. We agree that including this baseline in the uncertainty analysis is essential to assess the contribution of DML beyond heterogeneous ensembling. We have therefore added “Late Fusion” to Figures 2 and 3 in the main paper and Figures 6, 7, and 8 in the appendix. We highlighted these changes in blue in the revised manuscript and include the results of Late Fusion, DICE (w/o reg), and DICE in the tables below. These results clearly demonstrate the benefit of our DML loss in uncertainty analysis.

Figure 2, left — PANDA retained error rate (%) ↓

Method0% deferred20%40%60%80%
Late Fusion20.6514.047.642.020.50
DICE (w/o reg)20.6514.297.392.140.76
DICE18.4411.716.131.390.50

Figure 2, right, and Figure 3(b) — uncertainty separation ↑

MethodFigure 2: PANDA, Cliff’s ΔFigure 3(b): C16 → C17, Cliff’s Δ
Late Fusion0.630.65
DICE (w/o reg)0.630.66
DICE0.650.64

All listed separation comparisons have p < 10⁻⁴.

Figure 3(a) — PANDA classification with rejection

SplitMethodF1 before rejection (%) ↑F1 after rejection (%) ↑F1 gain (pp) ↑
ValidationLate Fusion73.89 ± 1.4281.17 ± 1.99+7.28
ValidationDICE (w/o reg)73.54 ± 1.6680.19 ± 3.23+6.65
ValidationDICE77.58 ± 1.8486.23 ± 4.10+8.65
TestLate Fusion73.41 ± 0.8179.43 ± 1.13+6.03
TestDICE (w/o reg)73.62 ± 1.1179.41 ± 1.70+5.78
TestDICE76.64 ± 1.0286.49 ± 2.23+9.85
AuthorsOpenReview ↗

Our Rebuttal to Reviewer pDf2 (W3 & W4)

W3: Role of the Gramian regularization

The motivation for the Gramian term is to introduce representation-level coupling alongside the prediction-level coupling provided by the DML term. DML acts on class probabilities and therefore does not directly constrain the geometry of the slide embeddings. After projecting the heterogeneous embeddings into a common-dimensional space, the Gramian term penalizes the volume they jointly span to align representations of the same slide across experts. The supervised objectives guide the learned representations toward the diagnostic task.

This is particularly relevant to our setting because all experts process the same slide. Their representations provide then different views of the same underlying tissue, shaped by differences in architecture, pretraining objectives, and pretraining data, so are already grounded in the same semantic information. Gramian-based alignment was originally proposed for multimodal representation learning. This provides a rationale for adapting it to heterogeneous PFMs, where the representations share both the input slide and the downstream task.

The relationship between the Gramian term and the uncertainty estimates is provided empirically in Figure 2 and Figure 3(a), where DICE outperforms all the UQ baselines, including DICE (w/o reg).

W4: Clarification on lesion localization

As stated in Section 4.2.3, the localization experiment is a supplementary analysis to examine whether the learned attention maps identify diagnostically relevant regions. It serves as a desirable byproduct of our ensemble learning framework, since it allows us to obtain unsupervised localization of tumorous regions. However, we do not claim any relationship between this analysis and uncertainty estimation. We clarify this in lines 126-127 and 420 of the revised manuscript. It is correct that the localization score is the mean of the expert attention weights. The same aggregation is used for the heterogeneous late-fusion and the homogeneous late-fusion (Deep Ensemble) baseline, so the comparison evaluates the attention maps produced by the respective training procedures under a shared aggregation rule.

We thank Reviewer pDf2 once again for the time and effort they put into our work. As the discussion period approaches its end, we would kindly like to ask if there are any further comments or concerns that we could address.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.