Trained on 2.3M slides, PRISM2 generates diagnosis instead of pixels
paige and microsoft released prism2, a pathology model trained on 2.3m slides and clinical reports to generate diagnostic text. it shifts pathology from image classification to interpretable clinical dialogue.
paige and microsoft introduced prism2, a perceiver-based encoder trained jointly on tissue tiles and pathology reports to interpret whole-slide images and answer diagnostic questions. the model aggregates thousands of tile embeddings per slide from a dataset of 2.3 million slides to generate text-based conclusions. this moves beyond standard image classification by integrating clinical context into the diagnostic loop. it signals a maturation of medical ai from simple detection tools to systems capable of complex reasoning and communication.