Microsoft spotlights Paige's year-old PRISM2 model trained on 2.3 million pathology slides

Microsoft Research on August 4th spotlighted PRISM2, a multimodal pathology model developed by Paige and Microsoft researchers, trained on 2,350,518 whole-slide images from 685,507 specimens and about 200,000 patients. The model, first detailed in a preprint submitted June 16, 2025, achieves 0.874 balanced accuracy for pan-cancer detection and 0.861 for rare cancers in developer tests, though it is not a new release and is part of the digital-pathology portfolio acquired by Tempus AI on August 22, 2025.

Microsoft Research https://x.com/MSFTResearch/status/2084616849052831991 on August 4th drew fresh attention to PRISM2, a multimodal pathology model that Paige and Microsoft researchers first detailed in 2025. The model processes whole-slide tissue images and answers diagnostic questions in natural language, an approach its developers say can match or exceed narrower cancer-detection systems on several benchmarks. x.com https://x.com/MSFTResearch/status/2084616849052831991 The post amplifies research that is more than a year old. The PRISM2 preprint https://arxiv.org/abs/2506.13063 was initially submitted on June 16th, 2025, and revised on October 31st. Paige announced the model on July 16th, 2025, before Tempus AI acquired Paige the following month. PRISM2 is therefore neither a new model release nor a new Microsoft product. It is part of the digital-pathology portfolio that Tempus bought when it acquired Paige on August 22nd, 2025. arxiv.org https://arxiv.org/abs/2506.13063 PRISM2 extends work associated with Paige founder Thomas J. Fuchs, a former Memorial Sloan Kettering computational-pathology researcher who used large collections of digitized tissue slides and their existing diagnostic reports to train cancer models. That method reduced reliance on pathologists manually marking individual regions of each slide, one of the constraints on earlier pathology AI systems. Fuchs and colleagues founded Paige around technology developed through MSK's pathology research program. mskcc.org https://www.mskcc.org/news/deeppet-uses-artificial-intelligence-generate-images-of-the-bodys-internal-activities?utm source=openai What PRISM2 does PRISM2 was trained on 2,350,518 hematoxylin and eosin-stained whole-slide images representing 685,507 specimens and roughly 200,000 patients. Paige paired those images with diagnostic reports and generated about 14 million question-and-answer pairs from the text. The researchers used GPT-4o to rewrite reports and generate training questions, while a 4-billion-parameter Phi-3 Mini model supplied PRISM2's language component. arxiv.org https://arxiv.org/abs/2506.13063 Training happened in two stages. Paige first aligned representations of the pathology slides with diagnostic language. It then trained the language model to interpret those representations, generate summaries and answer questions such as whether a cancer indication was present. The aim was to replace carefully engineered text prompts with the kind of direct yes-or-no questioning a pathologist or software developer could use. arxiv.org https://arxiv.org/abs/2506.13063 In the researchers' tests, that dialogue method reached 0.874 balanced accuracy for pan-cancer detection and 0.861 for rare cancers. The contrastive version of PRISM2 scored 0.719 and 0.718 on the same tasks. The paper also reported stronger results than the earlier PRISM and TITAN models for several diagnostic tasks, although biomarker prediction was one area where PRISM2's dialogue-tuned representation did not produce the best overall result. arxiv.org https://arxiv.org/abs/2506.13063 Those figures come from the developers' preprint and benchmark setup. The paper also documents errors in example conversations, including an incorrect tumor-grade prediction and a missed treatment-related finding. Its evaluations show how the architecture behaves on research datasets; they do not establish how PRISM2 would perform prospectively across hospitals, scanners and patient populations during routine clinical care. arxiv.org https://arxiv.org/abs/2506.13063 Tempus owns the route to market Paige made a PRISM2 checkpoint available through Hugging Face https://huggingface.co/paige-ai/Prism2 under a license limited to noncommercial academic research. Commercial use requires separate approval. Paige also said in its 2025 announcement that PRISM2 was available for licensing and enterprise integration, and that it had been incorporated into Alba, a beta clinical copilot that had not been commercially released at the time. huggingface.co https://huggingface.co/paige-ai/Prism2?utm source=openai The ownership changed five weeks after Paige's announcement. Tempus said the acquisition would add Paige's digital-pathology technology and data to its precision-medicine operations. A subsequent Tempus regulatory filing https://www.sec.gov/Archives/edgar/data/1717115/000119312525264652/tem-20250930.htm valued the acquisition consideration at $102.7 million, consisting of about $3 million in cash and the rest primarily in Tempus stock. investors.tempus.com https://investors.tempus.com/news-releases/news-release-details/tempus-announces-acquisition-paige?utm source=openai That transaction gives PRISM2 a potential distribution path beyond a standalone research release. Tempus can connect Paige's slide-level models with a broader collection of clinical and molecular oncology data. Microsoft, meanwhile, remains a research collaborator and infrastructure partner rather than the model's owner. The immediate evidence still comes from benchmarks and a preprint, leaving clinical deployment as the test that matters next.