Cancer3.AI › AI in Oncology › AI Models › Virchow
Foundation model Gated weights (licence click-through) Research use only pathology

Virchow Virchow (v1)

632-million-parameter vision transformer pretrained on 1.5 million whole-slide images from about 100,000 patients — the largest pathology pretraining set at its release — and used to build a pan-cancer detection model covering 17 cancer types, including rare ones.

🤗 paige-ai/Virchow Paper DOI { } JSON
7 616 downloads / month ♥ 76 updated 2024-09-03 gated

At a glance

DeveloperPaige / Memorial Sloan Kettering Cancer Center
VersionVirchow (v1)
Released2024-07-22
Licencecc-by-nc-nd-4.0
AvailabilityGated weights (licence click-through)
KindFoundation model
Regulatory statusResearch use only
Parameters632M

What it does

Tile-level embeddings (class token plus mean of patch tokens) for slide-level aggregation; Paige used them for a clinical-grade pan-cancer detector and for biomarker prediction.

Notable uses in oncology

Pan-cancer detection AUC 0.95 overall in the paper, with strong performance on rare cancers where task-specific models lack data.

Tasks, data types and cancers

TaskFeature extraction (embeddings)ClassificationDetection
Data typeHistopathology (WSI)
Clinical settingDiagnosisBasic research
CancerPan-cancer
Input224×224 tile at 20×
Output2,560-dim embedding (concatenated class token and mean patch token)

Architecture

FamilyVision Transformer
BackboneViT-H/14
Parameters632M
Pre-trainingDINOv2 self-supervised

Training data

1.5 million H&E whole-slide images from about 100,000 patients at Memorial Sloan Kettering Cancer Center, spanning 17 tissue types.

Training set size1.5M WSIs
InstitutionsMemorial Sloan Kettering Cancer Center (New York, USA)
Populations / geographysingle tertiary cancer centre

Evaluation

Benchmark / datasetMetricValueExternal validationSource
Pan-cancer detection — 17 cancer types (9 common, 8 rare), specimen level, trained on Virchow tile embeddings AUROC 0.95 (Vorontsov E et al., Nat Med 2024; PMID 39039250) no Source

How to run

# generic timm loader — check the model card for the exact init args and image normalisation
import timm, torch
from huggingface_hub import login
login()  # gated repos: accept the licence on huggingface.co first
model = timm.create_model('hf-hub:paige-ai/Virchow', pretrained=True)
model.eval()
cfg = timm.data.resolve_data_config({}, model=model)
transform = timm.data.create_transform(**cfg)
# emb = model(transform(tile).unsqueeze(0))  # 1 x D tile embedding
Librarytimm

Gated on Hugging Face under a non-commercial licence; Virchow2 (mixed magnification, 3.1M slides) is the successor and usually the better default.

Regulatory status and intended use

Regulatory statusResearch use only
Intended useResearch release of the encoder; Paige's cleared products are separate.

Source →

Regulatory status is quoted from the source linked above and can change. Research-use-only models must not be used for clinical decisions.

Limitations and bias

  • Single-institution pretraining; scanner (Leica Aperio) dominance.
  • Non-commercial licence.
  • Superseded by Virchow2 for most uses.

Sources

  1. Vorontsov E et al. A foundation model for clinical-grade computational pathology and rare cancers detection. Nat Med 2024
  2. Hugging Face — paige-ai/Virchow

This page is educational — it is not medical advice and does not replace consultation with an oncologist. Diagnostic and treatment decisions are made solely by specialist physicians.