Skip to content

Virchow / Virchow-2

Family: path-fm Modality: H&E whole-slide image tiles (224×224 at 0.5 µm/px for Virchow; mixed magnifications 5×/10×/20×/40× plus IHC for Virchow-2) Released: Virchow: Nature Medicine, October 2024 (Vorontsov et al.); Virchow-2: arXiv August 2024 (Zimmermann et al., updated 2025) License: Virchow: Apache-2.0 weights (academic, non-clinical use). Virchow-2: CC-BY-NC-ND-4.0, non-commercial academic research only; institutional email required Code/checkpoint: huggingface.co/paige-ai/Virchow; huggingface.co/paige-ai/Virchow2 Also surfaced in: single-cell-spatial-omics, virtual-cells

Virchow is a tile-level pathology foundation model from Paige and Memorial Sloan Kettering Cancer Center, designed as a general-purpose H&E feature extractor that supports cancer detection, subtyping, biomarker prediction, and rare-cancer triage with a single frozen backbone. Virchow-2 keeps the same 632M-parameter ViT-H/14 backbone but doubles the training corpus to 3.1 million slides, adds IHC and mixed magnifications, and retrains with a pathology-tuned DINOv2 recipe. The line is one of the three pathology FMs (with UNI and Prov-GigaPath) most often used as a drop-in patch encoder in 2025-2026 translational pathology pipelines.

Architecture & training

Both Virchow and Virchow-2 use a ViT-H/14 backbone: 632M parameters, 32 layers, 1280-dim embeddings, 16 attention heads, SwiGLU activation, LayerScale enabled, 224×224 input. The final tile embedding is a 2560-dim vector formed by concatenating the class token (1280-d) with the mean-pooled patch tokens (1280-d). Virchow-2 adds 4 register tokens.

Virchow (v1) was pretrained with standard DINOv2 self-supervision on 1.5 million H&E whole-slide images from Memorial Sloan Kettering spanning ~100,000 patients, sampled at 0.5 µm/px (20× magnification). The model is released under Apache-2.0 (with a non-clinical-use disclaimer).

Virchow-2 scales the training corpus to 3.1 million WSIs covering ~225,000 patients and nearly 200 tissue types, drawn from MSK plus external consult cases — ~15% of slides and ~57% of patients come from non-MSK institutions submitting challenging cases worldwide. The recipe is a pathology-modified DINOv2: the KoLeo regularizer is replaced with a kernel density estimator (more stable on near-duplicate histology tiles), the standard crop-and-resize augmentation is replaced with an "extended context translation" that preserves magnification semantics, and tiles are sampled at 5×, 10×, 20×, and 40× to make the encoder magnification-aware. Virchow-2 ships as three variants: Virchow-2 (632M), Virchow-2G (1.9B), and Virchow-2G Mini (22M, distilled). The Virchow-2 weights carry the more restrictive CC-BY-NC-ND-4.0 license.

Neither model includes a slide-level head — both are tile encoders intended to be paired with attention MIL (abMIL/ABMIL), TransMIL, or a linear probe downstream.

Headline benchmarks

Pan-cancer detection (Virchow Nature Medicine 2024). A weakly-supervised aggregator on top of Virchow tile features reached 0.949 specimen-level AUC across 17 cancer types and 0.937 AUC on 7 rare cancer subtypes. On the same evaluation, Virchow embeddings beat UNI (0.940), Phikon (0.932), and CTransPath (0.907) on overall pan-cancer AUC.

Tile-level tasks (Virchow-2, arXiv 2024). Across 12 tile-level benchmarks, Virchow-2 lifts the weighted average F1 from 0.944 (Virchow) to 0.966, with the pathology-tuned DINOv2 recipe contributing most of the gain. Virchow-2G (1.9B) is reported as state-of-the-art across the same suite.

Independent comparisons. Public clinical benchmarks of pathology FMs (e.g. the 2025 ArteraAI evaluation and the PMC clinical benchmark PMC12003829) consistently place Virchow / Virchow-2 in the top tier alongside UNI2-h and Prov-GigaPath, with model ordering shifting by task — Virchow-2 tends to win on rare-cancer detection and IHC-aware tasks where mixed magnifications matter.

Why it matters for AACR / clinical pathology

  • Rare-cancer triage. Virchow is the first pathology FM whose flagship paper foregrounds rare-cancer detection (7 rare tumors at 0.937 AUC), directly relevant to AACR streams on under-served cancers and pediatric oncology.
  • Magnification-aware features. Virchow-2's mixed-magnification training is the closest current answer to the spatial-pathology question of how to share features across 10×/20×/40× tiles without retraining — useful for pipelines that combine tile-level cell morphology with slide-level architecture.
  • Industrial provenance. Both Virchow checkpoints come from Paige, whose Paige Prostate Detect is FDA-cleared, so the backbone has a clearer regulatory lineage than purely academic FMs.

Known limitations

  • License asymmetry. Virchow-2 (the better model) is CC-BY-NC-ND-4.0 and gated, so commercial / translational groups face a real adoption barrier and often default to Virchow-1 or UNI2-h.
  • No slide-level encoder. Unlike Prov-GigaPath, neither Virchow release includes a slide-level transformer — downstream MIL aggregation is the user's problem.
  • Training data concentration. Despite Virchow-2's global consult cases, the majority of slides still originate from a single institution (MSK), and the paper does not release fairness analyses by site or demographic.
  • No vision-language head. Virchow is vision-only; it does not support text queries or report generation natively (cf. PathChat, CONCH).

Sources

  • Vorontsov E. et al. (2024). A foundation model for clinical-grade computational pathology and rare cancers detection. Nature Medicine 30:2924-2935. doi:10.1038/s41591-024-03141-0
  • Vorontsov E. et al. (2023). Virchow: A Million-Slide Digital Pathology Foundation Model. arXiv:2309.07778 (preprint).
  • Zimmermann E. et al. (2024). Virchow2: Scaling Self-Supervised Mixed Magnification Models in Pathology. arXiv:2408.00738.
  • HuggingFace: paige-ai/Virchow (Apache-2.0), paige-ai/Virchow2 (CC-BY-NC-ND-4.0).
  • Paige.ai. The Virchow Foundation Model, Explained. paige.ai/blog.

Use in the AACR 2026 corpus

Mention scan pending — populated by the alias-matching pipeline used for UNI, CHIEF, and Prov-GigaPath.