Pith. sign in

REVIEW 2 major objections 2 minor 41 references

Tile-level linear probing correlates strongly with slide-level performance across histopathology foundation models.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

High correlation between tile-level linear probing and slide-level performance of foundation models across 58 tasks indicates tile benchmarks can reliably shortlist strong candidates for WSI analysis.

T0 review reviewed 2026-06-27 challenge →

load-bearing objection Tile probing correlates well enough with slide results to act as a cheap filter for model selection, but the limited task and aggregator set leaves the general claim on shaky ground. the 2 major comments →

arxiv 2606.10778 v1 pith:GSTEM522 submitted 2026-06-09 cs.CV

From Patches to Patients: A study of the tile-to-slide performance transferability in Digital Pathology

Sofi\`ene Boutaj , Leo Fillioux , Maria Vakalopoulou , Stergios Christodoulidis , Pierre Marza This is my paper
classification cs.CV
keywords digital pathologyfoundation modelswhole slide imagesmultiple instance learningtile-level probingperformance transferabilityhistopathologyencoder representations
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper examines whether fast tile-level tests on foundation models can reliably predict their success on full whole-slide image tasks. Benchmarking 19 models on 42 slide-level and 16 tile-level tasks reveals high correlation between the two, with encoder representation quality as the main driver of WSI outcomes. Transferability holds steady across models and depends more on dataset size and tiles per slide than on task difficulty. Tile results can shortlist strong candidates, though full slide evaluation is still required for final clinical validation. This setup aims to cut the cost of testing every model through complete preprocessing and MIL training pipelines.

Core claim

High correlation exists between tile probing metrics and slide-level outcomes using ABMIL and Mean Pooling aggregators. Encoder representation quality is the primary determinant of WSI success. Sensitivity analyses confirm stability across models, with cohort sizes and tile counts per slide exerting more influence than average task difficulty. Tile benchmarks agree with slide benchmarks on top-performing models and serve as an efficient first filter.

What carries the argument

Correlation between tile-level linear probing performance and slide-level MIL aggregator results across multiple tasks and models.

Load-bearing premise

The selected 42 slide-level and 16 tile-level tasks together with ABMIL and Mean Pooling aggregators are representative enough of real clinical cohorts for the transferability conclusion to generalize.

What would settle it

A foundation model that ranks high on tile probing but ranks low on slide-level performance in a large new clinical cohort would show the proxy does not hold.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 2 minor

Summary. The manuscript benchmarks 19 foundation models on 42 slide-level and 16 tile-level histopathology tasks. It reports high correlations between tile-level linear probing metrics and slide-level performance under ABMIL and mean-pooling aggregators, concluding that encoder representation quality is the primary driver of WSI success and that tile-level probing can serve as an efficient proxy for initial model shortlisting, while slide-level evaluation remains necessary for final clinical validation. Sensitivity analyses examine effects of cohort size and tiles per slide but find limited influence from average task difficulty.

Significance. If the reported correlations prove robust beyond the studied tasks and aggregators, the work would offer a practical efficiency gain for foundation-model selection in digital pathology by avoiding full MIL training for every candidate encoder during initial screening. It supplies concrete empirical support for prioritizing representation quality over aggregation choice in the tested regime and identifies cohort size and tile count as more influential factors than task difficulty.

major comments (2)
  1. [Abstract] Abstract: the claim that encoder representation quality is the primary determinant of WSI success is demonstrated only for ABMIL and mean pooling; the transferability conclusion would be strengthened by showing that the same high correlations hold (or quantifying degradation) for at least one additional aggregator family such as transformer-based MIL.
  2. [Abstract] Abstract (sensitivity analyses paragraph): stability is shown with respect to cohort size and tiles per slide, yet no analysis or discussion addresses whether the chosen 42+16 tasks span clinically relevant variability (staining protocols, scanner differences, class imbalance, rare subtypes). If rank correlations materially decrease on a differently sampled task distribution, the proxy recommendation requires qualification.
minor comments (2)
  1. [Abstract] Abstract: numerical values for the reported correlations (e.g., Spearman ho or Pearson r) and their confidence intervals are omitted, making it impossible to judge effect size from the summary alone.
  2. [Abstract] The abstract states that tile benchmarks 'reliably shortlist strong candidates' but does not specify the agreement metric (top-k overlap, rank correlation of best models, etc.) used to reach this conclusion.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the two major comments point by point below and indicate planned revisions.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that encoder representation quality is the primary determinant of WSI success is demonstrated only for ABMIL and mean pooling; the transferability conclusion would be strengthened by showing that the same high correlations hold (or quantifying degradation) for at least one additional aggregator family such as transformer-based MIL.

    Authors: We agree that the experiments were restricted to ABMIL and mean pooling. These two aggregators were selected as representative of attention-based and simple pooling approaches commonly used in digital pathology. Extending the analysis to transformer-based MIL would indeed provide stronger evidence for the general claim. Given the substantial computational cost of retraining across 19 encoders and 42 tasks, we did not perform those additional runs. In the revision we will qualify the abstract claim to specify the tested aggregators and add a limitations paragraph discussing the scope and suggesting future validation on other MIL families. revision: partial

  2. Referee: [Abstract] Abstract (sensitivity analyses paragraph): stability is shown with respect to cohort size and tiles per slide, yet no analysis or discussion addresses whether the chosen 42+16 tasks span clinically relevant variability (staining protocols, scanner differences, class imbalance, rare subtypes). If rank correlations materially decrease on a differently sampled task distribution, the proxy recommendation requires qualification.

    Authors: The 42 slide-level tasks were drawn from multiple public datasets spanning different organs, cancer types, and staining conditions. We did not, however, conduct an explicit meta-analysis quantifying coverage of scanner variability, class imbalance ratios, or rare-subtype prevalence. We will revise the sensitivity-analysis discussion to describe the diversity of the task collection more precisely and to qualify the proxy recommendation by noting that the observed correlations hold within the studied distribution; broader sampling of task variability remains an important direction for future work. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical benchmarking with direct measurements

full rationale

The paper conducts an empirical study benchmarking 19 foundation models across 42 slide-level and 16 tile-level tasks using ABMIL and Mean Pooling. It reports observed correlations between tile probing metrics and slide-level outcomes, plus sensitivity analyses on cohort size and tiles per slide. No mathematical derivations, fitted parameters presented as predictions, self-citations as load-bearing premises, or ansatzes are present. The central claim follows directly from measured rank correlations on held-out tasks without reduction to inputs by construction. This is a standard self-contained empirical analysis.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Empirical study relying on standard machine learning evaluation practices for histopathology; no new free parameters, axioms beyond domain conventions, or invented entities are introduced.

axioms (1)
  • domain assumption Linear probing on tiles and MIL aggregators on slides are appropriate proxies for evaluating foundation model representation quality in histopathology.
    Invoked to justify comparing tile and slide performance metrics across models.

reviewed 2026-06-27 · how reviews work

0 comments
Cite this review

Pith. "Pith review of From Patches to Patients: A study of the tile-to-slide performance transferability in Digital Pathology." pith.science (2026). https://pith.science/paper/GSTEM522

@misc{pith2026260610778,
  author       = {Pith},
  title        = {Pith review of: From Patches to Patients: A study of the tile-to-slide performance transferability in Digital Pathology},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GSTEM522}},
  note         = {Machine review of arXiv:2606.10778}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Foundation Models (FMs) have recently redefined the state-of-the-art in histopathology by providing robust representations for whole-slide 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.

Figures

Figures reproduced from arXiv: 2606.10778 by Leo Fillioux, Maria Vakalopoulou, Pierre Marza, Sofi\`ene Boutaj, Stergios Christodoulidis.

Figure 1
Figure 1. Figure 1: Comparison of slide-level and tile-level benchmarks. (1,2) Overview of the 2 benchmark types. (3) Storage and compute require￾ments: Average storage (log scale) per dataset and compute time per dataset and per model, measured on a single NVIDIA V100 GPU. Slide-level averages were computed across 42 tasks and 19 models, while tile-level averages were computed across 16 tasks and 19 models. [25,20,22,8,40,17… view at source ↗
Figure 2
Figure 2. Figure 2: Rank correlation between slide-level and tile-level benchmarks. Comparing the rank across 19 models on a set of tile-level and slide-level tasks. Slide-level aggregation is performed via mean-pooling (left) and ABMIL (right); measured by Spearman (ρS), Pearson (ρP ), and Kendall’s τ . centrated around the identity line and strong correlations (ρS = 0.925, τ = 0.778, ρP = 0.967; permutation p = 2 × 10−4 for… view at source ↗
Figure 3
Figure 3. Figure 3: Comprehensive sensitivity analysis of tile-to-slide performance correlation on ABMIL. From left to right: (1 ) Leave-One-Model-Out sensi￾tivity distribution demonstrating a stable Spearman correlation (ρs) when in￾dividual foundation models are removed. The remaining plots show correlation trajectories when slide-level tasks are iteratively removed (smallest-to-largest vs. largest-to-smallest) based on: (2… view at source ↗
Figure 4
Figure 4. Figure 4: Micro-Level Transferability of Top-5 Histopathology Models. Distribution of model overlap between tile-level and slide-level tasks, strat￾ified by slide task clinical category. For each evalu￾ated pair of tile and slide tasks, the intersection of the top 5 highest-performing models was computed [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 1 canonical work pages

  1. [1]

    arXiv (2024)

    Aben, N., de Jong, E.D., Gatopoulos, I., et al.: Towards large-scale training of pathology foundation models. arXiv (2024)

    2024

  2. [2]

    Scientific Reports (2025)

    Alfasly, S., Alabtah, G., Hemati, S., et al.: Validation of histopathology foundation models through whole slide image retrieval. Scientific Reports (2025)

    2025

  3. [3]

    Mayo Clinic Proceedings: Digital Health (2024)

    Alfasly, S., Nejat, P., Hemati, S., et al.: Foundation models for histopathol- ogy—fanfare or flair. Mayo Clinic Proceedings: Digital Health (2024)

    2024

  4. [4]

    Database (2022)

    Brancati, N., Anniciello, A.M., Pati, P., et al.: Bracs: A dataset for breast carci- noma subtyping in h&e histology images. Database (2022)

    2022

  5. [5]

    NPJ Precision Oncology (2025)

    Breen, J., Allen, K., Zucker, K., et al.: A comprehensive evaluation of histopathol- ogy foundation models for ovarian cancer subtype classification. NPJ Precision Oncology (2025)

    2025

  6. [6]

    Nature Communications (2025)

    Campanella, G., Chen, S., Singh, M., et al.: A clinical benchmark of public self- supervised pathology foundation models. Nature Communications (2025)

    2025

  7. [7]

    Nature Medicine (2024)

    Chen, R.J., Ding, T., Lu, M.Y., et al.: Towards a general-purpose foundation model for computational pathology. Nature Medicine (2024)

    2024

  8. [8]

    arXiv (2024)

    Ding,T.,Wagner,S.J.,Song,A.H.,etal.:Multimodalwholeslidefoundationmodel for pathology. arXiv (2024)

    2024

  9. [9]

    Journal of proteome research (2015)

    Edwards, N.J., Oberti, M., Thangudu, R.R., et al.: The cptac data portal: a re- source for cancer proteomics research. Journal of proteome research (2015)

    2015

  10. [10]

    Jama318(22), 2199–2210 (2017) 10 S

    Ehteshami Bejnordi, B., Veta, M., Johannes van Diest, P., et al.: Diagnostic as- sessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. Jama318(22), 2199–2210 (2017) 10 S. Boutaj et al

    2017

  11. [11]

    In: MICCAI (2025)

    Filiot, A., Dop, N., Tchita, O., et al.: Distilling foundation models for robust and efficient models in digital pathology. In: MICCAI (2025)

    2025

  12. [12]

    medRxiv (2023)

    Filiot, A., Ghermi, R., Olivier, A., et al.: Scaling self-supervised learning for histopathology with masked image modeling. medRxiv (2023)

    2023

  13. [13]

    arXiv (2024)

    Filiot, A., Jacob, P., Mac Kain, A., et al.: Phikon-v2, a large and public feature extractor for biomarker prediction. arXiv (2024)

    2024

  14. [14]

    In: MIDL (2024)

    Gatopoulos, I., Känzig, N., Moser, R., et al.: eva: Evaluation framework for pathol- ogy foundation models. In: MIDL (2024)

    2024

  15. [15]

    arXiv (2024)

    Gustafsson, F.K., Rantalainen, M.: Evaluating computational pathology founda- tion models for prostate cancer grading under distribution shifts. arXiv (2024)

    2024

  16. [16]

    Nature medicine (2023)

    Huang, Z., Bianchi, F., Yuksekgonul, M., et al.: A visual–language foundation model for pathology image analysis using medical twitter. Nature medicine (2023)

    2023

  17. [17]

    NeurIPS (2023)

    Ikezogwo, W., Seyfioglu, S., Ghezloo, F., et al.: Quilt-1m: One million image-text pairs for histopathology. NeurIPS (2023)

    2023

  18. [18]

    In: ICML (2018)

    Ilse,M.,Tomczak,J.,Welling,M.:Attention-baseddeepmultipleinstancelearning. In: ICML (2018)

    2018

  19. [19]

    In: CVPR (2023)

    Kang, M., Song, H., Park, S., et al.: Benchmarking self-supervised learning on diverse pathology datasets. In: CVPR (2023)

    2023

  20. [20]

    arXiv (2025)

    Karasikov, M.,vanDoorn, J.,Känzig,N.,etal.:Trainingstate-of-the-artpathology foundation models with orders of magnitude less data. arXiv (2025)

    2025

  21. [21]

    Computers in Biology and Medicine (2025)

    Lee, J., Lim, J., Byeon, K., et al.: Benchmarking pathology foundation models: Adaptation strategies and scenarios. Computers in Biology and Medicine (2025)

    2025

  22. [22]

    Nature Medicine (2024)

    Lu, M.Y., Chen, B., Williamson, D.F., et al.: A visual-language foundation model for computational pathology. Nature Medicine (2024)

    2024

  23. [23]

    arXiv (2025)

    Majzoub, R.A., Malik, H., Naseer, M., et al.: How good is my histopathology vision-language foundation model? a holistic benchmark. arXiv (2025)

    2025

  24. [24]

    Neural Information Processing Systems (NeurIPS) D&B Track (2025)

    Marza, P., Fillioux, L., Boutaj, S., et al.: THUNDER: Tile-level histopathology im- age understanding benchmark. Neural Information Processing Systems (NeurIPS) D&B Track (2025)

    2025

  25. [25]

    arXiv (2024)

    Nechaev, D., Pchelnikov, A., Ivanova, E.: Hibou: A family of foundational vision transformers for pathology. arXiv (2024)

    2024

  26. [26]

    arXiv (2024)

    Neidlinger, P., El Nahhas, O.S., Muti, H.S., et al.: Benchmarking foundation mod- els as feature extractors for weakly-supervised computational pathology. arXiv (2024)

    2024

  27. [27]

    arXiv (2023)

    Oquab, M., Darcet, T., Moutakanni, T., et al.: Dinov2: Learning robust visual features without supervision. arXiv (2023)

    2023

  28. [28]

    In: ICML (2021)

    Radford, A., Kim, J.W., Hallacy, C., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)

    2021

  29. [29]

    Saillard, C., Jenatton, R., Llinares-López, F.o.: H-optimus-0 (2024),https:// github.com/bioptimus/releases/tree/main/models/h-optimus/v0

    2024

  30. [30]

    arXiv (2024)

    Shaikovski, G., Casson, A., Severson, K., et al.: Prism: A multi-modal generative foundation model for slide-level histopathology. arXiv (2024)

    2024

  31. [31]

    Shao, D., Chen, R.J., Song, A.H., et al.: Do multiple instance learning models transfer? arXiv preprint arXiv:2506.09022 (2025)

    work page arXiv 2025

  32. [32]

    Contemporary Oncology (2015)

    Tomczak, K., Czerwińska, P., Wiznerowicz, M.: Review the cancer genome atlas (tcga): an immeasurable source of knowledge. Contemporary Oncology (2015)

    2015

  33. [33]

    arXiv (2025)

    Vaidya, A., Zhang, A., Jaume, G., et al.: Molecular-driven foundation model for oncologic pathology. arXiv (2025)

    2025

  34. [34]

    Nature medicine (2024) From Patches to Patients: Tile-to-slide transferability in Digital Pathology 11

    Vorontsov,E., Bozkurt, A., Casson,A., et al.: Afoundation modelfor clinical-grade computational pathology and rare cancers detection. Nature medicine (2024) From Patches to Patients: Tile-to-slide transferability in Digital Pathology 11

    2024

  35. [35]

    Nature (2024)

    Wang, X., Zhao, J., Marostica, E., et al.: A pathology foundation model for cancer diagnosis and prognosis prediction. Nature (2024)

    2024

  36. [36]

    arXiv (2023)

    Wölflein, G., Ferber, D., Meneghetti, A.R., et al.: Benchmarking pathology feature extractors for whole slide image classification. arXiv (2023)

    2023

  37. [37]

    Nature (2025)

    Xiang, J., Wang, X., Zhang, X., et al.: A vision–language foundation model for precision oncology. Nature (2025)

    2025

  38. [38]

    Nature (2024)

    Xu, H., Usuyama, N., Bagga, J., et al.: A whole-slide foundation model for digital pathology from real-world data. Nature (2024)

    2024

  39. [39]

    arXiv (2025)

    Zhang, A., Jaume, G., Vaidya, A., et al.: Accelerating data processing and bench- marking of ai models for pathology. arXiv (2025)

    2025

  40. [40]

    arXiv (2024)

    Zhou, X., Sun, L., He, D., et al.: A knowledge-enhanced pathology vision-language foundation model for cancer diagnosis. arXiv (2024)

    2024

  41. [41]

    arXiv (2024)

    Zimmermann, E., Vorontsov, E., Viret, J., et al.: Virchow2: Scaling self-supervised mixed magnification models in pathology. arXiv (2024)

    2024

This paper was first reviewed by grok-4.3 on June 27, 2026.