A foundation model for clinical-grade computational pathology and rare cancers detection

Eugene Vorontsov; Alican Bozkurt; Adam Casson; George Shaikovski; Michal Zelechowski

doi:10.1038/s41591-024-03141-0

A foundation model for clinical-grade computational pathology and rare cancers detection

Eugene Vorontsov, Alican Bozkurt, Adam Casson, George Shaikovski, Michal Zelechowski, Kristen Severson, Eric Zimmermann, James M. Hall, Neil Tenenholtz, Nicolò Fusi, Ellen Yang, Philippe Mathieu, Alexander van Eck, Donghun Lee, Julian Viret...

Nature Portfolio (2024) • Volume 30, Issue 10, Pages 2924-2935

Method-ToolMethod-ToolPDF AvailableGrade Eligible⚠️ High Risk Flags

Overall Assessment

Strong Methodological Quality

Assessment created by PaperScorers Medical AI v0.1.0 on Dec 15, 2025

B-

71/100

Key Takeaways

•Virchow (ViT-H, 632M) trained on ~1.5M WSIs enables robust pan-cancer detection (AUC 0.95).
•Strong generalisation to rare cancers and external sites; UNI/Phikon lag.
•Pan-cancer model approaches specialist products, surpassing on some rare variants.
•Biomarker prediction from H&E competitive across 9 targets.
•Model and SDK openly released; proprietary WSI data available on request.

Conclusion

A large-scale pathology foundation model delivers near–clinical-grade performance with superior breadth, though COIs and lack of preregistration temper confidence.

Quick Actions

Read Full Paper

Quality Dimensions

Bias & Integrity

COI handling, outcome switching, selective reporting

D45

Methods Rigour

Protocol clarity, sampling, controls

B-74

External Validity

Generalisability, setting, population

B77

Transparency & Reproducibility

Data/code/materials availability, preregistration

C+68

Significance & Novelty

Contribution vs prior art; context if known

A90

Statistical Validity

Models, assumptions, multiplicity, CI priority

B78

Integrity & Transparency

Integrity checks

P-hacking risk?

?

None suspected

Signs of selective hypothesis testing or analysis choices that could inflate false positive rates.

Outcome switching?

?

None suspected

Evidence that primary outcomes were changed during or after the study, potentially distorting results.

Conflict of interest?

?

Serious concern

Authors have financial or other relationships that could bias research findings.

Data integrity issues?

?

None suspected

Concerns about the accuracy, completeness, or authenticity of the reported data.

Open science signals

Research data

?

Upon request

Raw data used in the analyses is publicly available or accessible upon request.

Analysis code

?

Open access

Code or scripts used to analyze the data are shared for reproducibility.

Study materials

?

Open access

Protocols, questionnaires, and other materials are publicly available.

Premise

Primary Research Question

Does a million-scale pathology foundation model enable robust pan-cancer detection and biomarker prediction, including rare cancers?

Hypothesis

Scaling data and model size improves generalisation and performance on rare cancers and OOD settings.

Hypothesis is Falsifiable

✓Yes

Assess via AUC, specificity at fixed sensitivity across tissues and sites.

PICO Framework

Population—Not specified

Non-interventional method-tool study.

Intervention—Not specified

No clinical intervention; index test is algorithm.

Comparison—Not specified

Comparators are other models; not a clinical comparator.

Outcome✓Yes

Diagnostic AUC/specificity; biomarker AUC are clinically pertinent.

Literature Positioning

Literature Review Balanced

Well covered

Extensive prior SSL/foundation model pathology citations (refs 35–42).

Evidence: Intro/References.

Research Gap Clearly Stated

Well covered

Need for generalisable, clinical-grade pathology FMs and rare cancer coverage.

Evidence: pp. 2926–2928

Stated Contribution Clear

Well covered

Introduces Virchow; pan-cancer detector and biomarker tasks across tissues.

Evidence: Abstract; Fig. 1

Study Provenance

Peer reviewed venueGrant funding disclosedConflicts disclosed5 affiliations listed

Authors & Affiliations

Paige AI; Microsoft Research; Memorial Sloan Kettering Cancer Center; NSW Health Pathology; University of Rochester.

Funding Statement

Acknowledgements cite NIH/NCI P30CA008748; Paige–Microsoft collaboration supported the work.

Conflicts of Interest

Multiple authors are Paige employees/equity holders; MSK holds financial/IP interests in Paige; competing interests detailed.

Peer Review

✓Nature Medicine, 2024; peer-reviewed.

Evidence: Front matter: Published online 22 July 2024.

Methodological Assessment

Diagnostic Accuracy StudySTARDDiagnostic

8 applicable checklist items.

Yes 4 · No 2 · Unclear 2

✓

Representative Spectrum of Patients

Broad, multi-tissue spectrum incl. rare cancers and external sites.

Evidence: Fig. 2a–f

✓

Index Test Replicable

Model weights on HF; SDK available; architecture/hparams detailed.

Evidence: Code availability; Methods Virchow architecture

✓

Reference Standard Appropriate

Specimen labels from synoptic/diagnostic reports; curated sets by pathologists.

Evidence: Methods Label extraction; Clinical evaluation datasets

?

Verification Bias Minimised

Retrospective labels; not clear all positives verified uniformly.

Evidence: Methods Label extraction

?

Blinding/Independence Adequate

No blinding between index test and reference assessment reported.

✗

Thresholds Pre-specified

Operating points tuned on a dataset; not pre-specified a priori.

Evidence: p. 2929–2930 (Fig. 3e text)

✗

2×2 Table Derivable

Counts at thresholds not given; only AUC/specificity at fixed sensitivity.

Evidence: Fig. 2b–d; Ext. Data Fig. 3

✓

AUC/ROC with CI Reported

AUCs with 95% CIs (DeLong); multiplicity corrected (Holm).

Evidence: Fig. 2; Methods Statistical analysis

Modeling / SimulationModel ReportingModeling

5 applicable checklist items.

Yes 3 · No 2

✓

Structure Transparent

ViT-H/14 (632M), DINOv2 student–teacher; hyperparameters stated.

Evidence: Methods Virchow architecture; Ext. Data Fig. 2

✗

Uncertainty Propagated

No model uncertainty propagation; only CI via resampling.

Evidence: Methods Statistical analysis

✗

Sensitivity Analyses (Global)

Scaling analyses only; no global sensitivity exploration.

Evidence: Fig. 2g

✓

External Validation Performed

OOD: external institutions and unobserved tissues evaluated.

Evidence: Fig. 2e–f

✓

Code/Model Available

Weights on Hugging Face; SDK on GitHub.

Evidence: Code availability

Modules without applicable checklist items (8)

These designs were fully marked as not applicable but remain available for reference.

Abstract

The analysis of histopathology images with artificial intelligence aims to enable clinical decision support systems and precision medicine. The success of such applications depends on the ability to model the diverse patterns observed in pathology images. To this end, we present Virchow, the largest foundation model for computational pathology to date. In addition to the evaluation of biomarker prediction and cell identification, we demonstrate that a large foundation model enables pan-cancer detection, achieving 0.95 specimen-level area under the (receiver operating characteristic) curve across nine common and seven rare cancers. Furthermore, we show that with less training data, the pan-cancer detector built on Virchow can achieve similar performance to tissue-specific clinical-grade models in production and outperform them on some rare variants of cancer. Virchow’s performance gains highlight the value of a foundation model and open possibilities for many high-impact applications with limited amounts of labeled training data.

Quick Actions

Read Full Paper

Study Overview

Study Design

Method Tool

Primary classification used to evaluate this research.

Population

Retrospective H&E WSIs from MSKCC and external sites; multi-tissue (16 cancer types); foundation model training and slide-level evaluations.

Brief summary of who or what was studied.

Sample Size

6,142

Number of participants included in the study analyses.

Pre-registration

Not registered

Indicates whether the protocol or trial was registered before data collection.

Primary Outcome

Unclear whether the primary outcome was pre-specified.

Effect Size

AUC: AUC 0.950 (specimen-level pan-cancer)

Reviewer Notes

Method-tool: self-supervised ViT-H trained on ~1.5M WSIs with retrospective diagnostic accuracy and biomarker prediction evaluations.

Publication Details

DOI

10.1038/s41591-024-03141-0

Published

July 22, 2024

Citations

218 citations

FWCI

138.61

External Resources

DOI Link OpenAlex

Disclaimer: This assessment is generated by AI and should not be the sole basis for clinical or research decisions. Always review the original paper and consult with domain experts.

Suggested Papers

D+

Radiopharmaceuticals and their applications in medicine

Siqi Zhang et al.•2025

F

The Annals of Internal Medicine

LB Wright et al.•2025

No Report

Preventive Effects of a Fermented Dairy Product against Alzheimer’s Disease and Identification of a Novel Oleamide with Enhanced Microglial Phagocytosis and Anti-Inflammatory Activity

Yasuhisa Ano et al.•2015

Suggested Papers

From Our Blog

Publication Bias: The Science We Never See

Journals love 'breakthroughs' and hate 'failures'. Why the 'File Drawer Problem' means half of science is missing.

Subgroup Analysis: Finding Patterns in Noise

The drug didn't work. But it worked for 'Left-handed Geminis.' This is how you find fake results.

Attrition Bias: The Vanishing Patients

If the patients who fail the treatment leave the study, the treatment looks perfect. This is attrition bias.