Paige announced the open-source release of its foundation model suite, Open Paige Foundation Model, that includes Virchow and PRISM, for computational pathology, as well as an ML SDK, to support the use of these models in the research community.
Virchow works as a million-slide foundation model for cancer, according to Paige. As detailed in a July 22 Nature Medicine article, Virchow-based pan-cancer prediction has been studied for its ability to generalize to tissue types or slides submitted from institutions not observed in the training data. Results observed build confidence that, with sufficient scale, foundation models will serve as the building blocks for the future development of a wide variety of downstream tasks, the authors wrote.
"In clinical practice, where most biopsy samples are benign, a pan-cancer detection system can prioritize cases to help reduce diagnostic turnaround," stated lead author Eugene Vorontsov and colleagues from Paige, Microsoft, NSW Health Pathology at St. George Hospital in Australia, and the University of Rochester and Memorial Sloan Kettering Cancer Center in New York. "With decreasing training data requirements, clinical-grade products for less common cancers could be developed."
Also, "biomarker prediction using routine [hematoxylin and eosin stain] H&E [whole-slide images] WSIs would increase screening rates; reduce intrusive, tissue-destructive testing; and rapidly provide the data needed to make more informed treatment decisions," Vorontsov and colleagues continued. "Virchow unlocks the ability to accurately and precisely detect unusual histological variants of cancer as well as biomarker status, something that is difficult to achieve with cancer- or biomarker-specific training due to the limited amount of associated training data."
The PRISM foundation model builds on image representations generated by Virchow and combines them into a whole-slide image (WSI) signature, according to Paige. Trained on an additional 587,000 WSIs and 195,000 clinical reports, PRISM generates written diagnostic summaries (for research use only). It can be used for standard diagnostic tasks such as cancer detection and subtyping without needing additional training, Paige said.
"Using text prompts, PRISM achieves zero-shot cancer detection and sub-typing performance approaching and surpassing that of a supervised aggregator model," a May paper posted to the ArXiv preprint server explained.
The models are a result of Paige's collaboration with Microsoft Research announced last year. Together, the foundation model represents a globally diverse dataset of 1.5 million digitized pathology slides from Memorial Sloan Kettering Cancer Center, according to Paige.
“Our goal in offering access to our technology, that underpins our clinical-grade applications, is to drive innovation and help push the boundaries of what is possible in cancer diagnostics and drug development, ultimately transforming patient care," stated Paige SVP Razik Yousfi in the company's announcement.
Virchow and PRISM are available at the URL https://huggingface.co/paige-ai, with ML SDK available at https://github.com/Paige-AI/paige-ml-sdk for the development of customized AI models, according to Paige.
