One AI model, many cancer tests: meet PRISM2, the pathology system that reads images and reports together

Key points

• Microsoft Research and Paige, now part of Tempus, published details of PRISM2 in Nature Medicine in 2025.
• PRISM2 matched or beat specialist AI tools at detecting prostate cancer, breast cancer, and breast lymph node metastasis in benchmark tests.
• The model learned from millions of paired tissue images and real pathology report excerpts, linking visual patterns to clinical language.
• Full model weights are publicly available on Hugging Face for non-commercial research use.

When a patient is suspected of having cancer, one of the most important steps is a biopsy: a small sample of tissue is removed, prepared on a glass slide, and examined under a microscope by a pathologist, a doctor who specialises in diagnosing disease from tissue. The pathologist writes a detailed report, and that report shapes almost every treatment decision that follows.

AI has been creeping into this process for years, mostly in the form of narrow, single-purpose tools trained to flag one specific cancer type. Think of them as very fast, very focused assistants that can do exactly one job and nothing else. Build a new job, start over.

PRISM2 is an attempt to change that.

What did the researchers actually build?

PRISM2 is a foundation model, a large AI system trained broadly so it can be adapted to many tasks, similar in concept to the technology behind ChatGPT but designed for medical images rather than general conversation.

Researchers at Microsoft Research and Paige (a pathology AI company that has since joined Tempus) trained the model on tissue slide images paired with language drawn from real pathology reports. From that pairing, they generated millions of question-and-answer examples, teaching the model to connect what it sees in a tissue image with the words a pathologist would use to describe it.

The result is a model that accepts images alone, or images plus typed questions and prompts, letting researchers interact with it in plain language rather than through rigid, pre-programmed instructions.

How well did it perform?

On standard benchmark tests, it held up against specialist tools. In prostate cancer detection, breast cancer detection, and detection of breast cancer that had spread to nearby lymph nodes, PRISM2 matched or exceeded the performance of purpose-built systems. Crucially, that was without training a separate model for each disease.

The paper appeared in Nature Medicine. As first reported by Microsoft AI, the full model weights are publicly available on Hugging Face for research use.

What does this mean for patients?

Nothing changes at the clinic today. PRISM2 is a research tool, not a product cleared for clinical use. Pathologists are still the ones reading slides and writing reports.

What it does offer is a faster on-ramp for future research. Teams working on rare cancers or diseases with limited training data could potentially adapt a single shared foundation instead of building a new system from zero. That lowers the cost and time of developing the next generation of diagnostic aids.

For patients, the practical consequence, if the research matures, is that AI assistance could reach more types of cancer, in more hospitals, faster than the current one-disease-at-a-time approach allows.

Common questions

Is PRISM2 being used to diagnose real patients?

No. PRISM2 is a research model released for academic and non-commercial use. It has not been approved by any regulator for clinical diagnosis.

Why does training on text reports matter for an image-reading model?

Pathology is as much about language as images. Pathologists describe what they see in precise, standardised words. Training the model on both images and those descriptions lets it learn the link between visual patterns and clinical meaning, making it more flexible than a model trained on images alone.

Who can access PRISM2?

Researchers can download the model weights from Hugging Face free of charge for non-commercial use.