List of Papers Browse by Subject Areas Author List
Abstract
The integration of Pathology Foundation Models (PFMs) with Multiple Instance Learning (MIL) has emerged as a prevailing paradigm for Whole Slide Image (WSI) analysis. However, due to diverse pre-training architectures and data, PFMs exhibit intrinsic heterogeneity and inconsistent performance across tasks. This limitation mirrors the challenge of clinical inter-observer variability, where relying on a single pathologist introduces the risk of subjective bias. To address this, we propose MEC, a Multi-Expert Consultation framework compatible with arbitrary weight-based MIL aggregators, designed to simulate the clinical consultation workflow. Specifically, MEC leverages diverse frozen PFMs as digital experts to independently localize task-specific discriminative regions within heterogeneous feature spaces. To broaden the overall scope of collective diagnostic evidence, we introduce a Stochastic Evidence Masking strategy that suppresses top-ranking features, thereby surfacing a wider variety of underlying cues. Finally, MEC synthesizes a robust, multi-view diagnostic decision by aggregating the collective consultation consensus with individual expert opinions. Extensive experiments demonstrate that MEC outperforms individual PFM baselines and alternative ensemble strategies, offering a practical framework to harness collective diagnostic potential without requiring additional large-scale pre-training. Code is available at https://github.com/JoeyFe/MEC.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1547_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
Link to the Dataset(s)
Camelyon17: https://camelyon17.grand-challenge.org/
TCGA-NSCLC, TCGA-GBMLGG: https://portal.gdc.cancer.gov/
BibTex
@InProceedings{CheZon_MEC_MICCAI2026,
author = { Chen, Zongyi AND Liang, Yu AND Lin, Weiping AND Chen, Rui AND Magnier, Baptiste AND Wang, Xi AND Wang, Liansheng},
title = { { MEC: A Multi-expert Consultation Framework for Synergizing Frozen Foundation Models in Whole Slide Image Analysis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16878},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper introduces a method to fuse predictions from a set of frozen pathology foundation models. Their main contribution is about isolating the most attented regions (tiles) from MIL models trained on top of each frozen expert and aggregating predictions from all of them on such top tiles. This allows to benefit from specific knowledge coming from different foundation models that might highlight different regions of interest so that the final model covers a broad range of them and does not miss any important tile that would have been overlooked by a specific expert. The proposed method is compared against individual experts and other fusion methods on 4 datasets, and evaluated in association with various MIL methods.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- S1.The need to efficiently aggregate predictions from different foundation models is well motivated and presented in the paper.
- S2.The intuition behind the specific proposed implementation, i. e. consensus between experts, is also properly motivated.
- S3.the method is clearly persented.
- S4.The method is compared against many different fusion baselines, individual foundtion models and associated with various MIL methods.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
-
W1.The fused representation presented in equation 5 is obtained by concatenating top-K embeddings from all foundation models. This might lead to a very high-dimensional representation. Can authors comment on this, and provide more insights about how the dimension how such a representation is handled. For instance, this means that the trained MLP will have a high number of parameters, which could cause overfitting.
-
W2.Fig. 3 presents a very appreciaated ablation study. However, it seems that the different components do not have a huge impact, in particular on specific datasets. Even drastic changes in hyperparameter values (e. g. mask probability) do not seem to change the performance much. This questions the necessity of the different components in the method. Could authors provide more details about these ablation results? For instance, is the proposedstochastic evidence masking really needed?
-W3.Fig. 4 seems to show that MEC leads to a broader attention covering the slide more uniformly and smoothly. However, I would be interested in authors commenting on why this is beneficial. Indeed, one could argue that a more localized attention is better from an interpretability perspective as it more clearly isolates imporant regions.
-
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I believe the proposed method is well motivated and valuable. However, as can be seen in the “Weaknesses” section, I have concerns regarding a few points, in particular ablation results, as I feel some of the components of the method might not be needed. I am looking forward to authors addressing these concerns.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The authors acknowledge that current Pathology Foundation Models (PFMs) exhibit inconsistent performance across tasks, drawing a parallel to clinical inter-observer variability. They propose MEC, a framework that leverages multiple frozen PFMs as complementary experts, arguing that their heterogeneous feature spaces encode complementary information that, when combined, yield more robust and consistent predictions. Concretely, each PFM is paired with an independent attention-based MIL head trained to produce the correct WSI label. A Stochastic Evidence Masking (SEM) strategy during training suppresses the top-Q patches by attention score with probability p, preventing any single set of features from dominating. The top-K raw patch embeddings from each PFM, selected at indices based on masked attention scores of the corresponding MIL head, are concatenated along the feature dimension to form a per-expert evidence vector, which are then concatenated across all experts into a fused representation and processed by an MLP to produce a consensus prediction. The final slide-level output is computed as the sum of this consensus logit and the average of the individual per-expert logits, a residual-style aggregation.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1) A novel and well-motivated pipeline for leveraging multiple frozen pathology foundation models for WSI classification. The combination of stochastic masking of top-attended patches with multi-expert evidence fusion is a principled design to compute the consensus vector that goes meaningfully beyond naive ensembling. Claim 1 is well supported.
2) Thorough evaluation across 4 datasets, 3 pathology foundation models, 5 single-PFM MIL competitors, and 6 ensemble methods, providing a comprehensive empirical picture.
3) Consistent performance gains over competitors across all datasets, with standard deviation reported throughout. The gains are largest and most consistent on the STS fine-grained sarcoma subtyping and CAMELYON17 tasks. Claim 3 is well supported.
4) Performance is relatively stable across key hyperparameters — mask number, mask probability, and number of top-K patches — suggesting the framework is not sensitive to exact tuning and would generalize reliably in practice.
5) The framework is attention-based MIL-agnostic, demonstrated to work consistently across ABMIL, ABMIL-GATE, and CLAM backbones without changes to the consensus mechanism.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1) While performance gains are consistent, they appear to be incremental and within standard deviation to other ensemble methods, as seen in Table 2.A statistical significance test would substantially strengthen the performance claims across the board.
2) The order of dimensionality of the fused representation Vfused and the architecture of the subsequent MLP are never disclosed. With K=50-75 patches per PFM, 3 PFMs, and feature dimensions of 1024+ per model, Vfused could reach on the order of 150,000-230,000 dimensions, implying a potentially enormous and underspecified MLP. The parameter count, depth, regularization strategy, and computational cost of this component should be reported, both for reproducibility and to assess whether the gains could be partly attributed to a substantially larger number of learnable parameters compared to the baselines. Also, in Table 2, this computational cost in terms of parameter count and FLOPs as a surrogate for complexity should be reported. Finally, in the ``Comparison with Ensemble Methods’’ section, you should disclose the inference time for all methods, so that the reported MEC’s inference time is properly framed.
3) Claim 2 is weakly/partially supported; namely, that the paper reveals the inherent complementarity of distinct PFMs in ROIs mining. The Analysis of Attention Regions section is currently qualitative, in 2 samples, and unconvincing as standalone evidence for a main paper claim. More critically, a quantitative measure, such as overlap or IoU or Accuracy between MEC attention maps and pathologist-annotated ROIs in multiple slides, compared to a single-PFM competitor like ABMIL and one Ensemble method, would directly validate the claim that MEC produces more diagnostically coherent attention. Given that ROI annotations are available for the STS cases, as evidenced in Figure 4, this experiment is straightforward and would substantially strengthen the interpretability argument.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
Minor weaknesses: 1) Figure 2 has some weird artifact in the offline wsi preprocess box. Otherwise this is a clean and well structured plot. 2) The attention colorbar in Figure 4 is missing.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
MEC is a well-motivated and carefully evaluated framework that addresses a genuine and timely problem in computational pathology. The core idea, treating heterogeneous PFMs as complementary experts and fusing their evidence through task-specific attention mining, goes meaningfully beyond naive ensembling. The evaluation is thorough, variance is reported, ablations isolate individual components, and attention visualizations support the qualitative claims. The main concerns are the scale of the undisclosed MLP architecture that handles the v_fused representation and the typical space v_fused lives in, statistical testing of results against benchmarks and a stronger/clearer quantitative explanation on the complementarity of ROI mining from different PFMs. These are addressable issues rather than fundamental flaws. If these are addressed, I will raise my score.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This paper proposes MEC, a multi-expert framework for whole-slide image analysis that treats multiple frozen pathology foundation models as separate experts. MEC mines task-specific evidence through an MIL branch for each expert, expands the evidence via stochastic evidence masking, and then forms a consensus prediction using top-K critical instances and residual logit aggregation. The evaluation is conducted on four WSI datasets, and the authors report consistent improvements over single-PFM baselines as well as several ensemble baselines.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper addresses an important and relevant problem. It starts from a reasonable observation that pathology foundation models may exhibit heterogeneous biases and task-dependent performance, making multi-model integration a significant question for computational pathology. The motivation of “multi-expert consultation” is also natural and easy to understand.
2.The method design is clean and pragmatic. Compared with approaches that rely on additional large-scale pretraining or heavy feature alignment, MEC is built on frozen PFMs, MIL branches, and a lightweight consensus head, which gives it practical value. I also appreciate that the paper does not simply combine logits, but first performs evidence mining for each expert before consensus fusion.
3.The evaluation is relatively strong for a methodological paper. It includes four datasets, three PFMs, five MIL backbones for single-model baselines, comparisons with several ensemble strategies, an ablation study, and heatmap visualizations. This experimental coverage makes the paper’s claims more credible than submissions that evaluate on only one or two datasets.
4.The empirical improvements appear consistent. From Table 2, MEC achieves the best F1 score across all four datasets (68.44, 86.05, 93.91, and 68.72), with improvements of approximately +3.9, +1.3, +0.7, and +3.0 points over the best non-MEC method on each dataset. These are not large gains for every task, but they are consistent and meaningful, especially on Camelyon17 and STS.
5.The ablation study is helpful in clarifying the contribution of each component, and the heatmaps in Figure 4 provide qualitative evidence that MEC tends to produce more comprehensive attention than single-PFM approaches.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.In my view, the novelty is moderate rather than high. In principle, the paper combines components that are already fairly familiar: multi-model integration, per-expert branches, masking/dropout-style regularization, top-K instance mining, concat+MLP fusion, and residual logit ensembling. In the broader literature, multi-foundation-model integration in pathology has already been explored through unified distillation (GPFM), consensus/divergence alignment (FM2), and slide-level ensemble pretraining (ELF). Thus, the contribution of MEC feels more like strong task-specific engineering than a genuinely new conceptual paradigm.
2.The claim that MEC is “compatible with arbitrary weight-based MIL aggregators” is not fully validated. Although the single-PFM baselines include ABMIL, CLAM, TransMIL, RRTMIL, and TDAMIL, MEC itself is only demonstrated with ABMIL, ABMIL-G, and CLAM. I would be more comfortable if this claim were softened, or if MEC were also evaluated with TransMIL, RRTMIL, and TDAMIL.
3.The comparison set still feels incomplete for a paper that positions itself as a multi-FM method. Table 2 compares against Mean, Concat, Logit, Vote, COBRA, and GPFM, but I would like to see comparisons with highly relevant prior work on multi-foundation-model integration, especially FM2 and ELF. Without these comparisons, the positioning against the state of the art is not fully convincing.
4.Not all comparisons are entirely apples-to-apples. The footnote in Table 2 states that COBRA follows its recommended linear probing setup, while the other methods use ABMIL as the backbone. This is understandable from the perspective of following each method’s original setting, but it still makes the performance gap somewhat harder to interpret.
5.The superiority claims would be stronger with statistical significance testing. The paper reports mean ± standard deviation from 5-fold cross-validation, but I did not see formal significance tests or p-values.
6.The evidence for interpretability/localization remains mostly qualitative. Figure 4 is interesting, but there is no quantitative evaluation of localization or ROI alignment.
7.There are some polish issues that should be cleaned up. At least one entry in Table 1 looks suspicious: 70.73±70.73 for RRTMIL+CONCHv1.5 on STS. This value looks like a typo or formatting error and should be checked.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
My recommendation is mainly based on the combination of an important problem, a practical method design, and broad experimental validation with consistently positive results. I do not assign a higher score because the conceptual novelty is still moderate, the comparison set for multi-FM integration is not fully complete, the claim of being MIL-agnostic is somewhat stronger than the evidence presented, and the statistical and quantitative analyses could be improved. Therefore, for me this paper is closer to a 4 than a 5 that it is worth accepting if the venue values empirical breadth and practical usefulness, but it is not a particularly outstanding paper in terms of methodological novelty.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
N/A
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
MEC ensembles frozen pathology foundation models by mining top-attention-tiles per MIL model. Reviewers broadly find the multi-expert motivation clear, intuition sound, empirical coverage strong across datasets/backbones/enemies, and gains generally consistent albeit sometimes modest. Concerns raised include very high fused dimensionality/overfitting risks and need for clarification on projection and computational cost (R1), ambiguous necessity of the different components (R1), missing statistical tests (R2, R3), undisclosed MLP geometry and fairness on parameter/FLOPs vs baselines plus timing disclosure (R2), moderate novelty (R3), and missing comparisons (R3). The AC reads the paper and reviews and agrees that the paper is interesting and recommends acceptance. However, the authors need to include all the raised points from the reviewers in their camera-ready version.
