Microsoft, Paige Unveil PRISM2 Model Matching Cancer AI Across 3 Detection Tasks
Updated
Updated · Microsoft · Aug 4
Microsoft, Paige Unveil PRISM2 Model Matching Cancer AI Across 3 Detection Tasks
3 articles · Updated · Microsoft · Aug 4
Summary
PRISM2 matched or beat specialized pathology AI on benchmark tests for prostate cancer, breast cancer and breast lymph node metastasis without building a separate model for each task.
Microsoft Research and Paige trained the foundation model on tissue images plus language from real pathology reports, generating millions of question-and-answer pairs that link visual patterns to diagnostic terms.
The system can work from images alone or from images and text, letting researchers query it with prompts instead of relying only on task-specific software.
Full PRISM2 weights are publicly available on Hugging Face for research use, aiming to speed development of broader pathology tools as cancer diagnosis data volumes grow.
PRISM2 beat specialized cancer detectors in benchmarks, but can a research-only pathology foundation model stay accurate across hospitals, scanners, and stains?
What happens when a pathology AI learns from 2.3 million slides and millions of report-based Q&As—better diagnosis support, or new risks hidden in the data?