Microsoft Introduces Pathology AI Model PRISM2 with Promising Benchmark Results

Microsoft Research and Paige have introduced PRISM2, a multimodal foundation model for computational pathology tasks. This model is trained on 2.3 million digitised pathology slides and 14 million question-and-answer pairs derived from approximately 685,000 pathology reports.

The researchers used diagnostic language as a training signal to help the model connect microscopic tissue patterns with pathology report terminology. PRISM2's architecture produces representations that can support prompt-based predictions and further development for diagnostic, biomarker, and survival-analysis tasks.

In tests against three Paige products, PRISM2 matched or outperformed them on various benchmarks, including detecting prostate cancer, breast cancer, and breast lymph node metastases. However, the model still requires task-specific adaptation or fine-tuning in some applications.