PRISM2 outperformed or matched existing slide-level foundation models across a comprehensive set of diagnostic, biomarker and patient outcome prediction tasks. Additionally, when further tuned to specifically predict long-term outcomes, PRISM2 outperformed standalone models, including achieving high performance predicting colorectal cancer recurrence-free survival.

"PRISM2 represents a true leap forward both in terms of scale and multi-modal AI capabilities," said Razik Yousfi, Senior Vice President and General Manager of AI Products at Tempus. "Our Pathology Foundation Models allow us to have a detailed understanding of tissue, unlocking clinical-grade precision and novel research applications capable of predicting patient outcomes and biomarker status. PRISM2 builds on this by aligning whole-slide pathology images with the language of clinical diagnosis through clinical dialogue training. It can seamlessly handle complex diagnostic and prognostic research tasks without requiring specialized fine-tuning, offering a powerful tool to advance precision oncology."

As part of Tempus’ proprietary Pathology Foundation Models, PRISM2 turns routine hematoxylin and eosin (H&E) slides into deep biological insights. By combining large vision models built from pathology images with large language models, it unlocks diagnostic-grade precision in research involving cancer detection, biomarker identification and prognosis prediction.

PRISM2 was trained on a diverse set of 2.3 million whole-slide images and 14 million diagnostic question−answer pairs derived from nearly 700,000 pathology reports, making it the largest multimodal slide-level pathology datasets to date.