Abstract

Foundation Models (FMs) have recently redefined the stateof-the-art in histopathology by providing robust representations for wholeslide image (WSI) analysis. However, selecting the optimal foundation model (FM) for a specific clinical cohort currently requires multiple preprocessing steps, followed by computationally expensive feature extraction and the training of a Multiple Instance Learning (MIL) aggregator for every model. In this work, we investigate whether efficient tile-level linear probing can serve as a reliable proxy for slide-level performance, reducing the need to run full slide-level pipelines for every candidate encoder. We benchmark 19 state-of-the-art FMs on 42 slide-level and 16 tile-level tasks, comparing tile probing metrics against slide-level outcomes using ABMIL and Mean Pooling aggregations. We observe a high correlation between tile and slide performance across varying task difficulties, indicating that encoder representation quality is the primary determinant of WSI success. Sensitivity analyses show that transferability is stable across models and is more influenced by cohort sizes and numbers of tiles per slide than by average task difficulty. We also measure the agreement in best performing models between tile and slide-level tasks, showing tile benchmarks reliably shortlist strong candidates. Overall, our study indicates that tile-level benchmarking provides an efficient and practical first step for narrowing down candidate models, while slide-level evaluation remains essential for final validation on clinical tasks.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2673_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{BouSof_From_MICCAI2026,
        author = { Boutaj, Sofiène AND Fillioux, Leo AND Vakalopoulou, Maria AND Christodoulidis, Stergios AND Marza, Pierre},
        title = { { From Patches to Patients: A Study of the Tile-to-Slide Performance Transferability in Digital Pathology } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16891},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper presents a large-scale benchmarking study investigating whether tile-level performance correlates with slide-level performance in digital pathology. It evaluates 19 state-of-the-art foundation models across 42 slide-level tasks and 16 tile-level tasks, demonstrating that tile-level benchmarking serves as a strong proxy for slide-level model selection.

  • 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 provides the first large-scale tile-to-slide benchmarking study and quantifies the overall tile-to-slide transferability. 2.The experimental evaluation is thorough. The authors benchmark 19 open-source pathology foundation models, covering 16 tile-level tasks and 42 slide-level tasks. 3.The paper shows that tile-level benchmarking is a strong proxy for slide-level model selection in digital pathology, reducing the need to run full slide-level pipelines and providing an effective way to select candidate models.

  • 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 the paper presents extensive statistical analyses demonstrating consistency between tile-level and slide-level tasks, it does not provide sufficient insight into the underlying causes of this consistency, such as the relationships between features. 2.The slide-level evaluation uses two aggregation strategies: mean pooling with linear probing and attention-based MIL (ABMIL). Potential aggregation-specific bias should be discussed.

  • 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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The work is interesting and the results are promising. However, the issues raised in the weaknesses need to be discussed.

  • 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 paper asks whether tile-level linear probing can reliably predict slide-level model rankings, benchmarking 19 pathology foundation models across 16 tile tasks and 42 slide tasks. The main finding is that rankings correlate strongly (ρS ≈ 0.92 for mean pooling, ρS ≈ 0.81 for ABMIL).

  • 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.

    The scale of the study is genuinely impressive (15,000+ V100 GPU-hours) and the result is reproducible on public datasets.

    The leave-one-model-out analysis is a good robustness check.

    The practical motivation is real — FM selection is a genuine bottleneck.

    The finding that top/bottom of the leaderboard is stable while mid-ranks scramble under ABMIL is a concrete and useful observation.

  • 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.Limited novelty and somewhat expected findings The central result — that models with better tile representations also produce better slide features — is fairly intuitive, since slide-level predictions are directly computed from aggregated tile embeddings. The contribution would be stronger if the paper explained when and why the correlation breaks down, rather than primarily confirming that it holds. The ABMIL gap (ρS drops from 0.92 to 0.81) is the most interesting finding but is underexplored. 2.Weak practical payoff given the stated conclusion The paper concludes that tile benchmarks are useful for “shortlisting” but that “slide-level validation remains essential.” If practitioners still need to run full slide pipelines for final selection, the actual computational savings are unclear. The paper doesn’t quantify how many models can be pruned by tile-level screening, or what the expected error rate of that pruning is in practice. 3.Missing random baseline for top-5 overlap (Fig. 4) The micro-level analysis reports top-5 overlaps but never compares against a random baseline. With 19 models and top-5 selection, the expected random overlap is non-trivial. Without this, it is impossible to assess whether the reported 82–100% “at least one shared model” probability is meaningfully above chance. 4.Tile tasks may not be representative All 16 tile-level tasks come from a single benchmark (THUNDER). The slide tasks span 10 anatomical sites and 4 clinical categories. The tile tasks are skewed toward local morphology (tumor detection, histologic pattern recognition), which may structurally favor models that also do well on aggregated slide tasks. The transferability claim rests heavily on this benchmark design choice, which isn’t critically examined. 5.ABMIL — the more clinically relevant aggregator — shows weaker transferability, and this deserves deeper analysis ABMIL is generally preferred in practice over mean pooling because it learns task-specific attention. Yet this is precisely where tile-level correlation is weakest. The paper attributes this to “additional variability introduced by learning,” but doesn’t investigate which types of tasks or which models are most affected, which would be the actionable insight.

  • 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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    See the above major weaknesses

  • 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

    While multiple pathology foundation models have been developed, it is computationally expensive to test all of them on a specific cohort and slide-level task to select the optimal foundation model. The authors of the paper ask a question on whether tile-level performance can be transferred to slide-level performance, because a tile-level task would be computationally more efficient than a slide-level task. The authors comprehensively benchmark 19 foundation models on 42 tasks and show that there is a strong correlation between tile-level and slide-level predictions. Since comparing foundation models for a slide-level task would require lots of storage and compute resource, it would offer an efficient way to select a few foundation model candidates by conducting a tile-level task which is generally lighter.

  • 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.

    The major strength of the paper is the comprehensiveness of the study design and analysis. The study is conducted on 42 clinical tasks using 19 foundation models. Rank correlation between slide-level and tile-level benchmarks in Figure 2 clearly shows that tile-level tasks can potentially replace slide-level tasks for selecting potential foundation models. Additional sensitivity analysis indicates that the transferability can be influenced by cohort sizes and number of tiles per slide which is important to understand when using the proposed method.

  • 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.

    I wonder if tile-level analysis is really more efficient. One reason why MIL is powerful is because MIL only requires slide-level labels without tile-level labels where tile-level labels can be extremely tedious and laborious. However, if tile-level analysis is needed to select foundation models, one has to additionally label tiles in a cohort. While the proposed method can be computationally efficient, by considering laborious effort of labeling tiles, the authors must comment whether it is really efficient overall.

  • 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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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 like the thorough analysis and the proposal to alternatively use tile-level performance to determine which foundation models to test for a specific cohort and task. One doubt I have is whether it is really efficient overall after considering an additional tile-level annotation process.

  • 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

We sincerely thank the reviewers for their constructive feedback and for recognizing the scale, comprehensiveness, and practical motivation of our study (15,000+ GPU hours, 19 foundation models, 16 tile-level tasks, 42 slide-level tasks). We are glad that they found the evaluation robust and the findings highly relevant for addressing the bottleneck of Foundation Model (FM) selection in digital pathology. Below, we address your concerns by sections. All discussions in this rebuttal will be incorporated into the final manuscript.

I) Underlying causes of tile-to-slide transferability & the impact of aggregation (ABMIL vs mean-pooling) To explain the showcased tile-to-slide consistency, one should note that slide-level predictions are generally performed by aggregating tile representations. As a result, informative tile-level features from a powerful encoder will likely lead to high slide-level performance, although other factors such as spatial contextualization or slide-level aggregation strategies may play a role depending on the task. Mean pooling acts as a direct evaluation of the frozen FM’s representation spaces, which is why it correlates so highly (Spearman r=0.92) with tile-level linear probing. Conversely, ABMIL introduces learned task-specific parameters that reshape this feature space to optimize for the slide-level label prediction. This added flexibility shifts the evaluation further away from the FM’s initial representation space, introducing aggregation-specific variability. However, we emphasize that, despite this variability, the Spermann correlation remains strong (r=0.81, with permutation p-values=0.0004). We included ABMIL as a slide-level aggregation strategy in our evaluation because it is a commonly-used and robust method that often matches the performance of more complex MIL architectures, as shown by Shao et al. (“Do Multiple Instance Learning Models Transfer?”, ICML 2025).

II) Practical payoff: computational vs. annotation efficiency Reviewers raised valid questions regarding the true efficiency of tile-level benchmarks, as they require laborious manual annotations, and that slide-level validation is still ultimately required. We agree that human annotation is expensive, but it is a one-time, fixed cost (which is already absorbed by leveraging public benchmarks like THUNDER). In contrast, evaluating new FMs or conducting hyperparameter searches represents a recurring, iterative hardware cost. Our results demonstrate that running a fast tile-level evaluation is an effective proxy for quantifying the incremental improvement of a newly pre-trained FM before committing vast compute resources to full slide-level MIL evaluation. While fine-grained final selection still benefits from slide-level validation, tile-level benchmarking effectively filters underperforming models, saving large amounts of computational resources in the development cycle, making it complementary to the slide-level benchmark.

III) Dataset representativeness and potential bias We acknowledge that our 16 tile-level tasks (derived from THUNDER) may be skewed toward local morphology. In fact, our micro-level transferability analysis (Fig. 4) reflects this limitation: overlap is lower for immune profiling compared to molecular or mutation tasks, likely because immune profiling relies on an understanding of more complex spatial tissue architecture. To maintain objectivity, we utilized a broad set of publicly available tile-level datasets. To verify that our findings are not driven by the specific set of tile-level datasets used in the study, we performed a leave-one-out sensitivity analysis, finding that the slide-tile Spearman correlation remained stable regardless of which tile dataset was excluded. We hope this highlights the community’s need for more diverse, context-aware tile-level datasets, which we believe will only strengthen these transferability correlations in the future and thus lead to more accessible and faster benchmarks.




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.

    This paper presents an in-depth study on the transferability of pathology foundation models from tile-level modeling to slide-level modeling. The overall motivation and the observation of a high correlation between tile-level and slide-level performance are interesting. All three reviewers recognize the contribution and potential impact of the work.

    On the other hand, I also agree with Reviewer #3 that this observation does not necessarily imply that tile-level tasks can be effectively used to select the best foundation models for WSI-level tasks, as it still higly relies on manual annotations. Nevertheless, the current observation itself, supported by large-scale evaluations across 42 tasks and 19 foundation models, is impactful enogh for discussion at MICCAI. I therefore rate this paper as an early accept.

    The theoretical value of the paper would be further strengthened if the authors could provide a more thorough discussion of why ABMIL shows a weaker correlation than mean pooling.



back to top