Pith. sign in

REVIEW 3 major objections 5 minor 5 cited by

PathBench: A comprehensive comparison benchmark for pathology foundation models towards precision oncology

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper benchmarks 19 pathology foundation models on 64 private, multi-center clinical tasks and finds Virchow2 and H-Optimus-1 the most effective overall.

desk verdict A large, useful PFM benchmark whose central leakage-free claim is undercut by slide-level splits in the gastric biopsy tasks; major revision needed before the rankings can be trusted. read the letter →

arxiv 2505.20202 v1 pith:OOD44IY5 submitted 2025-05-26 cs.CV

Jiabo Ma , Yingxue Xu , Fengtao Zhou , Yihui Wang , Cheng Jin , Zhengrui Guo , Jianfeng Wu , On Ki Tang
show 18 more authors
Huajun Zhou Xi Wang Luyang Luo Zhengyu Zhang Du Cai Zizhao Gao Wei Wang Yueping Liu Jiankun He Jing Cui Zhenhui Li Jing Zhang Feng Gao Xiuming Zhang Li Liang Ronald Cheong Kin Chan Zhe Wang Hao Chen
This is my paper · ORCID
classification cs.CV
keywords pathologyfoundationmodelswhole-slideimagesclinicalbenchmarkdataleakagepreventionprecisiononcologymulti-centerevaluationdiagnosisandprognosismodelleaderboard
verification ladder T0 review T1 audit T2 compute T3 formal

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 presents PathBench, a benchmark built to answer a clinical question: which pathology foundation models - AI systems pretrained on large collections of tissue slide images - actually work on real, multi-hospital cancer data, from diagnosis through prognosis. To answer it, the authors assembled 15,888 whole-slide images from 8,549 patients across 10 hospitals, spanning 64 tasks in lung, breast, gastric, colorectal, and brain cancer, and ran 19 publicly available models through a standardized evaluation pipeline. All evaluation data comes from private medical providers, which the paper argues excludes any overlap with the models' pretraining data and therefore avoids the data leakage that can inflate scores in public benchmarks. The headline result, if correct, is that Virchow2 and H-Optimus-1 are the most effective models overall, that vision-only models beat vision-language models on most clinical tasks, and that the best model differs by organ and task, so model choice should be task-specific. The paper also opens a live, community-updatable leaderboard so future models can be scored against the same leakage-controlled data.

What carries the argument

The benchmark itself is the machinery: a standardized whole-slide pipeline in which each foundation model extracts patch-level features from foreground tissue, and a train-from-scratch attention-based multiple instance learning aggregator (ABMIL) merges those features into a slide-level prediction for each task. Internal cohorts use a 7:1:2 train/validation/test split repeated 10 times; survival tasks use 5-fold cross-validation repeated 3 times; and significance is assessed with 1,000-fold bootstrapping and the Wilcoxon signed-rank test. Two design choices carry the argument: the use of exclusively private, in-house data to prevent pretraining leakage, and the aggregation of per-task ranks into an average ranking score that lets one leaderboard order models across heterogeneous clinical tasks.

What would settle it

Check the gastric biopsy cohort from Hospital H7 (2,700 slides from 1,345 patients): if even one patient has slides in both the training and test portions of the normal-versus-abnormal, intestinal-metaplasia, or subtyping splits, those tasks' reported AUCs are inflated and the paper's blanket leakage-free claim fails for that subset; the check is a lookup against the patient identifiers behind the slide labels used in the 7:1:2 splits.

Watch

Extended reading notes

Core claim

PathBench claims to be the first comprehensive benchmark for pathology foundation models that combines strictly private, multi-center clinical data with rigorous leakage prevention and coverage of the full clinical spectrum from diagnosis to prognosis. On its 64 tasks across five cancer types, evaluated with a fixed attention-based multiple-instance-learning protocol, the ranking of 19 models places Virchow2 (average rank score 5.0) and H-Optimus-1 (5.9) at the top, ahead of H-Optimus-0 (6.6), UNI2 (7.1), and mSTAR (7.4). The results further claim that the best model depends on the clinical setting: H-Optimus-1 leads in lung and colorectal cancer, Virchow2 leads in breast, gastric, and brain cancer, and UNI2 and CONCH1.5 lead in survival prognosis. The paper concludes from these results that vision-only foundation models remain more effective than vision-language models for clinical-level tasks, that larger pretraining scale still pays off, and that the field needs organ- and task-specific model selection rather than a single universal winner.

Load-bearing premise

The whole ranking rests on the assumption that no patient's slides appear in both the training and the test data; in the gastric biopsy tasks the splits are described as made on slides by label rather than on patients, so the same patient can sit on both sides of the split and inflate the reported scores.

Editorial extensions

If this is right

  • Clinicians can select a foundation model per organ: H-Optimus-1 for lung and colorectal questions, Virchow2 for breast, gastric, and brain, while UNI2 and CONCH1.5 are the prognosis specialists.
  • Vision-language models such as CONCH1.5 (rank 8) and CHIEF trail vision-only models on most clinical tasks, indicating that their multimodal pretraining has not yet translated into superior slide-level diagnosis.
  • Pretraining scale remains the strongest lever: the two top models were both trained on more than one million slides, while mSTAR shows that adding molecular and report knowledge can partly compensate for far less image data.
  • Newly released models can be added to the live leaderboard and scored on the same private, leakage-controlled data, so the benchmark can track the field as it evolves.
  • The consistent gap between internal and external hospital cohorts across tasks means a model's published internal accuracy should not be trusted without external validation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the leakage-free claim holds, rankings from benchmarks built on public slide archives may need re-examination, since some evaluated PFMs were pretrained on those same public collections, a concern this benchmark was explicitly designed to neutralize.
  • The organ-specific winner pattern suggests clinical deployment will likely involve a small portfolio of models chosen per cancer type and task, rather than one general-purpose best model.
  • A testable extension of the benchmark's own logic: re-run the gastric biopsy tasks (Hospital H7) with strict patient-level splits to verify whether the reported slide-stratified AUCs change; this would directly test the consistency of the paper's leakage-prevention claim.
  • Future multimodal models that combine pathology with genomics, named in the paper as candidates for the leaderboard once weights are public, imply the current ranking may be a snapshot of an open competition.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces PathBench, a benchmark that evaluates 19 pathology foundation models (PFMs) across 64 diagnostic, staging, biomarker, and survival tasks on 15,888 whole-slide images from 8,549 patients at 10 private hospitals. The authors report that Virchow2 and H-Optimus-1 are the most effective models overall, propose a live leaderboard, and claim that all evaluation data are private and leakage-free. The benchmark is described as addressing gaps left by existing public-dataset benchmarks, including pretraining-data overlap and narrow task coverage.

Significance. If the leakage-free claim held, PathBench would be a substantial resource: it is larger and more multi-center than most existing PFM benchmarks, covers diagnosis through prognosis, uses a fixed ABMIL aggregator with standardized preprocessing, and reports bootstrap confidence intervals, repeated splits, and Wilcoxon tests with external validation cohorts. The private-data design is a genuine attempt to avoid pretraining contamination. However, the central claim currently fails for a subset of tasks, and the headline ranking depends on an insufficiently specified aggregation. The resource is valuable in principle, but its main advertised property is not yet established.

major comments (3)
  1. [Materials and Methods, Gastric Cancer (1)-(3); Evaluating Protocols, Data Split; Supplementary Table 2] The gastric biopsy tasks are split at the slide level, not the patient level. The H7 biopsy cohort is described in Supplementary Table 2 as 2,700 slides from 1,345 patients, so on average each patient contributes about two slides. Sections (1) and (3) state that data were label-stratified into training, validation, and test sets, and section (2) states that slides were stratified by pathological labels, with no mention of grouping by patient. In contrast, the breast and colorectal sections explicitly say case-level or case-level split to avoid leakage. Because the Data Split protocol in Evaluating Protocols specifies only a 7:1:2 ratio and not the grouping unit, the same patient's slides can appear in both training and test sets in the six biopsy tasks. This directly contradicts the paper's abstract and Introduction claims of rigorous leakage prevention and inflates the AUCs in Tables 11-12, the gastric ranking in Figure 7i, and any global ranking that includes these tasks. Please re-run the gastric biopsy experiments with patient-level splits, or clearly quantify the effect of the current slide-level protocol.
  2. [Results, Figures 1d and 2g etc.; Evaluating Protocols] The headline conclusion that Virchow2 and H-Optimus-1 are the most effective models overall rests on the average ranking score, but the manuscript never defines how this score is computed. It is not stated whether the score is a mean rank across tasks, across cohorts, or across repeated runs; whether ties are broken; or how classification AUC and survival C-index are combined. Without this definition, the central ranking claim is not reproducible and its sensitivity to task composition (e.g., 31 gastric tasks versus 3 brain tasks) cannot be assessed. Please provide the exact formula and aggregation rule, and report per-organ and per-task-type sensitivity analyses.
  3. [Abstract, Introduction, Materials and Methods, Dataset] The statement that all evaluation data comes from private medical providers, with strict exclusion of any pretraining usage, is not backed by any auditable evidence in the manuscript. There is no data card, no listing of the hospitals' cohort inclusion criteria beyond the numbers in Supplementary Table 2, no description of how pretraining overlap was checked (e.g., hashing against public pretraining corpora), and no release of code or de-identified metadata. Since the entire value of PathBench is its claimed independence from PFM pretraining data, the authors should either provide a concrete exclusion protocol and data descriptors, or soften the claim to no known overlap.
minor comments (5)
  1. [Abstract] There are several typos and spacing errors, including in-hourse for in-house, We presentPathBench missing a space, and bioposy in the Gastric Cancer methods section.
  2. [Materials and Methods, Dataset] Some table and text contents are inconsistent or unclear: for example, inlcuding appears multiple times, evaluted appears in table captions, and lebel-stratified appears in the Brain Cancer section. Please run a thorough proofreading pass.
  3. [Evaluating Protocols, Statistical Analysis] The one-sided Wilcoxon signed-rank test is applied to compare the best model with every other model, but no multiple-testing correction is described. Given 19 models and up to 64 tasks, some significant P values are expected by chance; state how multiplicity was handled or justify the exploratory nature of these tests.
  4. [Model, Table 1] The table's Pretraining Data Source column is incomplete for several models (e.g., Hibou-L and H-Optimus-0 show dashes), which makes the pretraining-overlap assessment harder to follow. Please fill these entries with the public sources or state explicitly that they are undisclosed.
  5. [Results, Discussion] The paper repeatedly refers to a GitHub-hosted automated leaderboard but does not provide a repository URL or a code/data availability statement. Please add these, including the exact submission protocol for model weights.

Circularity Check

0 steps flagged · score 2.0 of 10

No construction-level circularity: the benchmark ranking is an empirical evaluation of external models, and the inclusion of two same-group models is not load-bearing.

full rationale

PathBench is an empirical evaluation study rather than a derivation chain. The central claim that Virchow2 and H-Optimus-1 are the most effective models overall is an observed ranking produced by running 19 publicly released models through a fixed ABMIL protocol with identical hyperparameters on private downstream tasks. No quantity is defined in terms of another, no fitted parameter is renamed as a prediction, and no uniqueness theorem or ansatz is imported from the authors' prior work to force the conclusion. The two models from the same research group, GPFM and mSTAR, are evaluated alongside the others but finish mid-pack, so the self-citations are not load-bearing for the headline result. A separate methodological concern is real but not circular: the gastric biopsy tasks are split at the slide level by pathological labels, while breast and colorectal tasks explicitly use case-level splits to avoid leakage; with roughly two slides per patient, same-patient slides could appear in both training and test sets, undermining the paper's leakage-free certification and possibly inflating those AUCs. That is an evaluation-validity flaw, not a case where the result reduces to its own input by construction. Accordingly, the circularity score is low, reflecting only the minor presence of same-group models in the benchmark rather than any derivation-level circularity.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The benchmark introduces no new entities. The main load-bearing assumptions are the unverifiable pretraining-data exclusion and the validity of slide-level splits. The ABMIL hyperparameters are hand-chosen protocol settings that could affect results.

free parameters (1)
  • ABMIL hyperparameters = hidden=512, dropout=0.25, lr=2e-4, 50 epochs (classification), 30 epochs (survival)
    Chosen by hand and fixed across all models; different hyperparameters per model could change the ranking.
assumptions (3)
  • domain assumption Private evaluation data are disjoint from the pretraining data of all 19 evaluated models
    Asserted but not verifiable; central to the leakage-free claim. No data provenance or overlap analysis is provided.
  • ad hoc to paper Slide-level splits are valid for tasks where a patient contributes multiple slides
    The gastric biopsy tasks split slides, implicitly assuming independence of slides from the same patient, which is questionable and likely inflates performance.
  • standard math AUC and C-index are appropriate metrics for these clinical tasks
    Standard metrics in the field; no issue.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PathBench: A comprehensive comparison benchmark for pathology foundation models towards precision oncology." pith.science (2026). https://pith.science/paper/OOD44IY5

@misc{pith2026250520202,
  author       = {Pith},
  title        = {Pith review of: PathBench: A comprehensive comparison benchmark for pathology foundation models towards precision oncology},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OOD44IY5}},
  note         = {Machine review of arXiv:2505.20202}
}
read the original abstract

The emergence of pathology foundation models has revolutionized computational histopathology, enabling highly accurate, generalized whole-slide image analysis for improved cancer diagnosis, and prognosis assessment. While these models show remarkable potential across cancer diagnostics and prognostics, their clinical translation faces critical challenges including variability in optimal model across cancer types, potential data leakage in evaluation, and lack of standardized benchmarks. Without rigorous, unbiased evaluation, even the most advanced PFMs risk remaining confined to research settings, delaying their life-saving applications. Existing benchmarking efforts remain limited by narrow cancer-type focus, potential pretraining data overlaps, or incomplete task coverage. We present PathBench, the first comprehensive benchmark addressing these gaps through: multi-center in-hourse datasets spanning common cancers with rigorous leakage prevention, evaluation across the full clinical spectrum from diagnosis to prognosis, and an automated leaderboard system for continuous model assessment. Our framework incorporates large-scale data, enabling objective comparison of PFMs while reflecting real-world clinical complexity. All evaluation data comes from private medical providers, with strict exclusion of any pretraining usage to avoid data leakage risks. We have collected 15,888 WSIs from 8,549 patients across 10 hospitals, encompassing over 64 diagnosis and prognosis tasks. Currently, our evaluation of 19 PFMs shows that Virchow2 and H-Optimus-1 are the most effective models overall. This work provides researchers with a robust platform for model development and offers clinicians actionable insights into PFM performance across diverse clinical scenarios, ultimately accelerating the translation of these transformative technologies into routine pathology practice.

Figures

Figures reproduced from arXiv: 2505.20202 by the authors.

Figure 1
Figure 1. The workflow and overall results of PathBench. a. The data used for the benchmark. b. The workflow of evaluating foundation models. c. The evaluated foundation models. d. The average ranking score of evaluated models. f. The average performance on the pathological diagnosis, molecular diagnosis, and prognosis tasks. 4/35 [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Performance on lung cancer data. a-b. Performance on molecular classification tasks CK7 and TTF-1. c. Classification performance for primary LUAD versus metastatic lung cancer. d. Primary site prediction for metastatic lung cancer. e-f. Performance on molecular classification tasks C-MET and NpsinA. g. Overall ranking scores of foundation models on lung cancer data. The error bar indicates the standard deviation. 5/… view at source ↗
Figure 3
Figure 3. The overall results of foundation models on the breast cancer data. a. Overall survival analysis results. b. Disease-free survival analysis results. c. Molecular subtyping results. d. TNM N Staging prediction results. The error bars represent the standard deviation. 6/35 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: The overall results of foundation models on the breast cancer data. a-e. The molecular subtyping performance. f. pTNM Staging prediction results. g. The average ranking score of various foundation models on breast cancer tasks. The error bars represent the standard dev…
Figure 5
Figure 5. Figure 5: Overall results of gastric cancer. a-b. The molecular subtyping results of HER-2 and S-100, respectively. c. Gastric cancer grading results. d. Lauren subtyping results. e. Perineural invasion detection results. The error bars represent the standard deviation. 9/35 [P…
Figure 6
Figure 6. Figure 6: Results of gastric cancer. a. Vascular invasion detection results. b. Pathological subtyping results. c. TNM N stage prediction results. d. TNM T stage prediction results. The error bars represent the standard deviation. 10/35 [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Result of gastric cancer. a. Normal and Abnormal slide classification. b. Detection of intestinal metaplasia. c. Classification of autoimmune chronic gastritis with Helicobacter pylori task. d. Detection of Helicobacter pylori-associated chronic gastritis. e-f. Detecti…
Figure 8
Figure 8. Figure 8: Overall results on colorectal cancer data. a-e. Staging tasks including TNM, CMS, TNM N, TNM T (4 classes), and TNM T (early or late stage). f-h. Survival analysis tasks including DFS, OS, and DSS. i. The average ranking score of the foundation model on the colorectal …
Figure 9
Figure 9. Figure 9: Results on the brain cancer data. a. Results of IDH Mutation prediction task. b. Performance on pathological subtyping task. c. Performance on the WHO Grading task. d. Average ranking score of various foundation models. The error bars represent the standard deviation. …

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 5 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. A Clinically Validated Foundation Model for Comprehensive Lung Pathology Interpretation

    eess.IV 2026-05 unverdicted novelty 7.0 of 10

    PulmoFoundation achieves 92.3% average AUC on 32 lung pathology tasks in prospective validation and raises pathologist accuracy from 83.8% to 91.7% in a crossover RCT.

  2. A Unified Low-level Foundation Model for Enhancing Pathology Image Quality

    cs.CV 2025-09 conditional novelty 6.0 of 10

    A prompt-guided diffusion model pretrained on 190 million pathology patches outperforms task-specific models across most restoration and virtual staining benchmarks.

  3. Towards Robust Foundation Models for Digital Pathology

    eess.IV 2025-07 conditional novelty 6.0 of 10

    PathoROB shows that all 20 evaluated pathology foundation models encode medical center information and that lower robustness correlates with larger downstream performance drops.

  4. BRIGHT: A Collaborative Generalist-Specialist Foundation Model for Breast Pathology

    cs.CV 2026-03 conditional novelty 5.0 of 10

    Fine-tuning a generalist pathology model on 51,000 breast WSIs and concatenating its features with the original model yields top-1 performance on 21 of 24 internal breast-pathology tasks, but only 5 of 10 external tasks.

  5. OTSurv: A Novel Multiple Instance Learning Framework for Survival Prediction with Heterogeneity-aware Optimal Transport

    cs.CV 2025-06 conditional novelty 4.0 of 10

    OTSurv applies progressive partial optimal transport to whole-slide survival prediction, achieving a 0.646 average C-index across six TCGA datasets.

Reference graph

Works this paper leans on

35 extracted references · 20 canonical work pages · cited by 5 Pith papers

  1. [3]

    Skrede, O.-J.et al.Deep learning for prediction of colorectal cancer outcome: a discovery and validation study.The Lancet 395, 350–360 (2020)

    work page 2020

  2. [4]

    & Kather, J

    Shmatko, A., Ghaffari Laleh, N., Gerstung, M. & Kather, J. N. Artificial intelligence in histopathology: enhancing cancer research and clinical oncology.Nat. cancer3, 1026–1038 (2022)

    work page 2022

  3. [5]

    advances10, eadi0302 (2024)

    Verma, R.et al.Sexually dimorphic computational histopathological signatures prognostic of overall survival in high-grade gliomas via deep learning.Sci. advances10, eadi0302 (2024)

    work page 2024

  4. [6]

    InProceedings of the AAAI conference on artificial intelligence, vol

    Shen, Y .et al.Explainable survival analysis with convolution-involved vision transformer. InProceedings of the AAAI conference on artificial intelligence, vol. 36, 2207–2215 (2022). 7.Oquab, M.et al.Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193(2023)

    arXiv 2022

  5. [8]

    & Chen, H

    Zhou, F. & Chen, H. Cross-modal translation and alignment for survival analysis. InProceedings of the IEEE/CVF International Conference on Computer Vision, 21485–21494 (2023)

    work page 2023

  6. [10]

    InInternational conference on machine learning, 8748–8763 (PmLR, 2021)

    Radford, A.et al.Learning transferable visual models from natural language supervision. InInternational conference on machine learning, 8748–8763 (PmLR, 2021). 11.Yu, J.et al.Coca: Contrastive captioners are image-text foundation models.arXiv preprint arXiv:2205.01917(2022)

    arXiv 2022

  7. [13]

    Bilal, M.et al.Foundation models in computational pathology: A review of challenges, opportunities, and impact.arXiv preprint arXiv:2502.08333(2025)

    arXiv 2025

  8. [14]

    Wang, X.et al.Foundation model for predicting prognosis and adjuvant therapy benefit from digital pathology in gi cancers.J. Clin. Oncol.JCO–24 (2025)

    work page 2025

Show all 35 references
  1. [15]

    Fournier, L. M.et al.Extended pre-training of histopathology foundation models uncovers co-existing breast cancer archetypes characterized by rna splicing or tgf-βdysregulation.bioRxiv2025–03 (2025). 16.Kondepudi, A.et al.Foundation models for fast, label-free detection of gli...

    2025

  2. [17]

    Mulliqi, N.et al.Foundation models–a panacea for artificial intelligence in pathology?arXiv preprint arXiv:2502.21264 (2025)

    2025

  3. [18]

    arXiv preprint arXiv:2408.15823(2024)

    Neidlinger, P.et al.Benchmarking foundation models as feature extractors for weakly-supervised computational pathology. arXiv preprint arXiv:2408.15823(2024)

    2024 arXiv

  4. [19]

    & Kwak, J

    Lee, J., Lim, J., Byeon, K. & Kwak, J. T. Benchmarking pathology foundation models: Adaptation strategies and scenarios. Comput. Biol. Medicine190, 110031 (2025)

    2025

  5. [20]

    A.et al.How good is my histopathology vision-language foundation model? a holistic benchmark.arXiv preprint arXiv:2503.12990(2025)

    Majzoub, R. A.et al.How good is my histopathology vision-language foundation model? a holistic benchmark.arXiv preprint arXiv:2503.12990(2025)

    2025 arXiv

  6. [21]

    Gustafsson, F. K. & Rantalainen, M. Evaluating computational pathology foundation models for prostate cancer grading under distribution shifts.arXiv preprint arXiv:2410.06723(2024). 21/35

    2024 arXiv

  7. [22]

    NPJ Precis

    Breen, J.et al.A comprehensive evaluation of histopathology foundation models for ovarian cancer subtype classification. NPJ Precis. Oncol.9, 33 (2025)

    2025

  8. [23]

    Commun.16, 3640 (2025)

    Campanella, G.et al.A clinical benchmark of public self-supervised pathology foundation models.Nat. Commun.16, 3640 (2025)

    2025

  9. [24]

    & Chen, W

    Cao, W., Qin, K., Li, F. & Chen, W. Comparative study of cancer profiles between 2020 and 2022 using global cancer statistics (globocan).J. Natl. Cancer Cent.4, 128–134 (2024)

    2024

  10. [25]

    & Sun, J

    He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, 770–778 (2016)

    2016

  11. [26]

    image analysis81, 102559 (2022)

    Wang, X.et al.Transformer-based unsupervised contrastive learning for histopathological image classification.Med. image analysis81, 102559 (2022)

    2022

  12. [27]

    Filiot, A.et al.Scaling self-supervised learning for histopathology with masked image modeling.medRxiv2023–07 (2023)

    2023

  13. [28]

    J.et al.Towards a general-purpose foundation model for computational pathology.Nat

    Chen, R. J.et al.Towards a general-purpose foundation model for computational pathology.Nat. Medicine30, 850–862 (2024)

    2024

  14. [29]

    medicine30, 2924–2935 (2024)

    V orontsov, E.et al.A foundation model for clinical-grade computational pathology and rare cancers detection.Nat. medicine30, 2924–2935 (2024). 30.Xu, H.et al.A whole-slide foundation model for digital pathology from real-world data.Nature630, 181–188 (2024)

    2024

  15. [31]

    & Ivanova, E

    Nechaev, D., Pchelnikov, A. & Ivanova, E. Hibou: A family of foundational vision transformers for pathology.arXiv preprint arXiv:2406.05074(2024)

    2024 arXiv

  16. [32]

    Ma, J.et al.Towards a generalizable pathology foundation model via unified knowledge distillation.Nat. Biomed. Eng. (2025)

    2025

  17. [33]

    34.Saillard, C.et al.H-optimus-0 (2024)

    Zimmermann, E.et al.Virchow2: Scaling self-supervised mixed magnification models in pathology.arXiv preprint arXiv:2408.00738(2024). 34.Saillard, C.et al.H-optimus-0 (2024)

    2024 arXiv

  18. [35]

    & Saillard, C

    Filiot, A., Jacob, P., Mac Kain, A. & Saillard, C. Phikon-v2, a large and public feature extractor for biomarker prediction. arXiv preprint arXiv:2409.09173(2024). 36.Saillard, C.et al.H-optimus-1 (2024)

    2024 arXiv

  19. [37]

    Huang, Z., Bianchi, F., Yuksekgonul, M., Montine, T. J. & Zou, J. A visual–language foundation model for pathology image analysis using medical twitter.Nat. medicine29, 2307–2316 (2023). 38.Lu, M. Y .et al.A visual-language foundation model for computational pathology.Nat. Med...

    2023

  20. [39]

    40.Xiang, J.et al.A vision–language foundation model for precision oncology.Nature1–10 (2025)

    Wang, X.et al.A pathology foundation model for cancer diagnosis and prognosis prediction.Nature634, 970–978 (2024). 40.Xiang, J.et al.A vision–language foundation model for precision oncology.Nature1–10 (2025)

    2024

  21. [41]

    Xu, Y .et al.A multimodal knowledge-enhanced whole-slide pathology foundation model.arXiv preprint arXiv:2407.15362 (2024)

    2024 arXiv

  22. [42]

    Chen, X., Xie, S. & He, K. An empirical study of training self-supervised vision transformers. InProceedings of the IEEE/CVF international conference on computer vision, 9640–9649 (2021)

    2021

  23. [43]

    InInternational Conference on Learning Representations

    Dosovitskiy, A.et al.An image is worth 16x16 words: Transformers for image recognition at scale. InInternational Conference on Learning Representations. 44.Oquab, M.et al.Dinov2: Learning robust visual features without supervision.Transactions on Mach. Learn. Res

  24. [45]

    neural information processing systems35, 25278–25294 (2022)

    Schuhmann, C.et al.Laion-5b: An open large-scale dataset for training next generation image-text models.Adv. neural information processing systems35, 25278–25294 (2022)

    2022

  25. [46]

    InInternational conference on machine learning, 8748–8763 (PmLR, 2021)

    Radford, A.et al.Learning transferable visual models from natural language supervision. InInternational conference on machine learning, 8748–8763 (PmLR, 2021). 47.Yu, J.et al.Coca: Contrastive captioners are image-text foundation models.Transactions on Mach. Learn. Res.. 48.Va...

    2025 arXiv

  26. [50]

    & Welling, M

    Ilse, M., Tomczak, J. & Welling, M. Attention-based deep multiple instance learning. InInternational conference on machine learning, 2127–2136 (PMLR, 2018)

    2018

  27. [51]

    Individual comparisons by ranking methods

    Wilcoxon, F. Individual comparisons by ranking methods. InBreakthroughs in statistics: Methodology and distribution, 196–202 (Springer, 1992). Supplementary information Table 2.The in-house data used for the benchmark. Organ Source Cohort Case Number Slides Number Lung H1 Meta...

    1992

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.