# Virchow: A Million-Slide Digital Pathology Foundation Model

**Authors:** Eugene Vorontsov, Alican Bozkurt, Adam Casson, George Shaikovski, Michal Zelechowski, Siqi Liu, Kristen Severson, Eric Zimmermann, James Hall, Neil Tenenholtz, Nicolo Fusi, Philippe Mathieu, Alexander van Eck, Donghun Lee, Julian Viret, Eric Robert, Yi Kan Wang, Jeremy D. Kunz, Matthew C. H. Lee, Jan Bernhard  
**arXiv:** [2309.07778](https://arxiv.org/abs/2309.07778)  
**Date:** 2023-09-14  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2309.07778v5/

## Key points

- Use of AI in precision medicine and decision support systems can revolutionize cancer diagnosis and treatment
- Analysis of pathology images provides valuable insights into the disease
- Researchers have developed the Virchow model, a foundation model for computational pathology
- Virchow is a vision transformer model with 632 million parameters trained using self-supervised learning empowered by DINOv2 algorithm
- Trained on a massive dataset of 1.5 million hematoxylin and eosin stained whole slide images from various tissue and specimen types
- Virchow has demonstrated remarkable performance in multiple tasks, including pan-cancer detection system with an overall AUC score of 0.949 across 17 different cancer types and AUC score of 0.937 on 7 rare cancer types
- Virchow is the state-of-the-art model for internal and external image tile level benchmarks and slide level biomarker prediction tasks
- Scaling up data size and network architecture is important in high-impact computational pathology applications with limited training data
- Virchow captures diverse patterns in pathology images, advancing AI-enabled computational pathology
- Opens up new possibilities for precise cancer diagnosis and treatment planning, improving patient outcomes in oncology care

## Summary

The use of artificial intelligence (AI) in precision medicine and decision support systems has the potential to revolutionize cancer diagnosis and treatment. One key aspect of this is the analysis of pathology images, which can provide valuable insights into the disease. To address this challenge, a team of researchers has developed the Virchow model - a foundation model for computational pathology. With an impressive 632 million parameters, Virchow is a vision transformer model trained using self-supervised learning empowered by the DINOv2 algorithm on a massive dataset consisting of 1.5 million hematoxylin and eosin stained whole slide images from various tissue and specimen types. This dataset is significantly larger than previous works in this field, highlighting the importance of training on massive pathology image datasets. The Virchow model has demonstrated remarkable performance in multiple tasks, including enabling the development of a pan-cancer detection system with an overall specimen-level AUC (Area Under Curve) score of 0.949 across 17 different cancer types and achieving an AUC score of 0.937 on 7 rare cancer types. These results establish Virchow as the state-of-the-art model for internal and external image tile level benchmarks and slide level biomarker prediction tasks. The gains in performance achieved by Virchow highlight the importance of scaling up both data size and network architecture in high-impact computational pathology applications where limited training data is available. Overall, Virchow represents a significant advancement in AI-enabled computational pathology with its ability to capture diverse patterns in pathology images. This opens up new possibilities for precise cancer diagnosis and treatment planning, ultimately improving patient outcomes in oncology care.