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

Why it matters

PRISM2 shows pathology AI moving toward general models that accept clinical questions, while Tempus ownership gives Paige's research a route into a larger oncology platform.

Microsoft Research 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)

The post amplifies research that is more than a year old. The PRISM2 preprint 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)

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)

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)

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)

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)

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)

Tempus owns the route to market

Paige made a PRISM2 checkpoint available through Hugging Face 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)

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 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)

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.