MUSK (Stanford, vision-language pathology)

MUSK is a vision-language pathology model from Stanford that uses masked multimodal pretraining to place image tokens and text tokens in one shared space, so the same network can read a slide and the words written about it. It was pretrained on 50M pathology images and 1B pathology-related text tokens, and the Nature 2025 paper reported that it predicted immunotherapy response and prognosis across cancers better than models built on images or text alone. The intended users are researchers and, in time, tumour boards deciding who will benefit from checkpoint inhibitors. All validation so far is retrospective, so whether these predictions change treatment decisions or outcomes has not been tested prospectively. For a newcomer: it reads slides and notes together to predict who will respond to immunotherapy, and that prediction has not yet been tested in a trial.