Paige, a leader in next-generation AI technology, has announced the release of PRISM2, its most advanced slide-level foundation model to date. PRISM2 is designed to advance large language models (LLM) by enabling them to understand pathology images, connecting visual patterns with the clinical language used by clinicians to unlock novel capabilities.
Developed in collaboration with Microsoft Research and expanding on the original PRISM model, PRISM2 was developed by building on Virchow2, the company’s state-of-the-art pathology foundation model. PRISM2 is trained on more than 2.3 million H&E-stained whole-slide images spanning hundreds of thousands of patient cases. Each slide is paired with its corresponding clinical report, grounding the model in the language of real-world diagnosis and enhancing its ability to generate clinically relevant insights across diverse populations, tissue types, and cancers. The result is a multimodal AI model capable of capturing both fine-grained cellular detail and whole-slide context.
PRISM2 was built with direct compatibility for modern language models, including a version integrated with Microsoft Phi-3, enabling more efficient multimodal deployment. This unique design bridges vision and text, powering natural language responses in pathology and delivering performance that surpasses previous models, particularly in rare cancer types and low-sample-size scenarios.
“PRISM2 represents a defining moment in digital pathology and AI. By combining the versatility of Virchow2 with rich clinical ground truths and seamless LLM integration, we’ve created a model that doesn’t just analyse tissue, it contextualises morphological patterns with diagnostic cues,” said Siqi Liu, VP of AI Science at Paige. “This unlocks new capabilities in reporting, screening, and outcome prediction, allowing AI to become a true partner in diagnosis, research and treatment.”
Benchmarks demonstrate that PRISM2 outperforms its predecessor and other slide-level foundation models across a wide range of diagnostic and biomarker prediction tasks. It delivers strong zero-shot performance on tasks such as tumour detection and can instantly generate concise, clinician-aligned reports, streamlining workloads, reducing turnaround times, and improving consistency in diagnosis.