PRISM2 Ties Pathology Slides to Diagnostic Reasoning Through Clinical Dialogue

A new end-to-end multimodal pathology foundation model, PRISM2, is trained on millions of whole-slide images and pathology reports with clinical-dialogue supervision to align histomorphology with diagnostic reasoning. It uses contrastive and autoregressive objectives to generate diagnostic embeddings that support prompt-based inference and transfer to downstream tasks, achieving clinical-grade performance on cancer detection benchmarks and enabling tasks such as cancer grading, survival and biomarker prediction with minimal additional training.