AI Model Learns Pathology Through Images and Language

Researchers from Microsoft Research and Paige developed PRISM2, a pathology AI foundation model trained on both tissue images and real pathology reports. Unlike single-task systems, it matched or exceeded specialized cancer-detection tools for prostate cancer, breast cancer, and lymph node metastasis without building separate models for each task. The model works by pairing pathology images with diagnostic language, creating millions of question-and-answer examples that link visual findings to clinical terminology. Users can interact with it through prompts, making it easier to develop future pathology tools without starting from scratch. Full model weights are publicly available on Hugging Face.