Tempus AI’s PRISM2, a multimodal pathology foundation model built with Microsoft researchers, can handle clinically important diagnostic tasks from simple prompts, according to a study published in Nature Medicine.
On diagnostic, biomarker and outcome benchmarks, PRISM2 came out even with or ahead of rival slide-level models, per Tempus. Additional tuning for long-term prognosis pushed it past standalone systems. Its colorectal cancer recurrence-free survival results stood out.
The model turns routine hematoxylin and eosin (H&E) slides into biological insight. It pairs large vision models trained on pathology images with large language models. Clinical dialogue training aligns whole-slide images with the language of diagnosis.
Razik Yousfi, Tempus’s senior vice president and general manager of AI products, called the release a leap in scale and multimodal capability. Complex research tasks need no specialized fine-tuning, he said. That makes it a strong addition to precision oncology work.
The study adds to Tempus’s push to build out its proprietary Pathology Foundation Models as it expands precision oncology offerings.
