Earlier this year, Paige raised the curtain and spotlighted new artificial intelligence models for digital cancer pathology. Now, the company is offering a peek behind the curtain by making versions of two of its programs open-source.
Aimed at oncology researchers and the makers of cancer drugs and diagnostics, the pair of foundation models were developed through a collaboration with Microsoft Research.
They include Virchow, which made its debut in January, carrying training from more than a million digitized pathology slides gathered from about 100,000 patients treated through Paige’s progenitor, Memorial Sloan Kettering Cancer Center.
The second, PRISM, builds upon the data generated by Virchow, combining tiled image representations into whole-slide signatures that can be wielded to help create written diagnostic descriptions. PRISM, short for Pathology Report and Image Summarization Model, was unveiled this past May and includes training from an additional 587,000 biopsy slides.
More advanced versions of Virchow, PRISM and other models from the former Fierce Medtech Fierce 15 winner will continue to be offered through commercial licenses.
“The introduction of our open-source models unlocks new possibilities to further advance research for better patient care and accelerate clinical discovery,” Paige’s senior vice president of technology, Razik Yousfi, said in a statement.
“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,” Yousfi said.
Last week, Virchow’s foundation model was the subject of a publication in Nature Medicine, where it was described as being capable of clinical-grade detection for 16 different cancers at once—comprising nine common and seven rare tumors—while also performing on par with more tissue-specific AI models.
In the study, Paige’s program—named for the father of modern pathology, Rudolf Virchow—showed it could spot cancers of the breast, prostate and lung in addition to the less common tumors of the stomach, cervix and bone, among others, all with a single computational model.
According to Paige, foundation models offer advantages compared to traditional AI development—as long as you have enough data. While task-specific models need to be trained with specific information, foundation models can be adapted to a broad range of queries, saving time and resources.