Teaching AI to speak the language of pathology

A study published in Nature Medicine describes PRISM2, a pathology foundation model developed by researchers from Microsoft Research and Paige, now part of Tempus. The model is trained on both tissue images and language derived from real pathology reports, aiming to learn from large text data. In testing, PRISM2 matched or exceeded the performance of specialized cancer-detection systems for prostate cancer, breast cancer, and breast lymph node metastasis detection, without creating separate models for each task. The full model weights are publicly available on Hugging Face. Pathology is central to many cancer diagnoses, and most current AI systems are single-purpose; PRISM2 supports a broader range of tasks via a single model that responds to prompts. The model pairs pathology images with report-derived information, creating millions of question-answer examples to connect visual findings with diagnostic language. Unlike earlier foundation models focused on visual representations, PRISM2 links visual patterns with clinical language from the outset.