A groundbreaking collaboration between Paige and Microsoft has yielded PRISM2, a sophisticated AI model poised to revolutionize digital pathology. This innovative system goes beyond traditional image classification by interpreting whole-slide images and generating natural language answers to diagnostic questions, a significant leap from simply identifying pixels.
PRISM2’s architecture is built upon a perceiver-based encoder, meticulously trained on a vast dataset comprising both tissue tiles and clinical dialogue extracted from pathology reports. This dual-pronged training approach allows the model to consolidate thousands of tile embeddings per slide into a singular, comprehensive representation. The subsequent text generation capability directly addresses diagnostic inquiries, offering a more intuitive and informative output for pathologists.
The sheer scale of the training data underscores the robustness of PRISM2. The model has been trained on an impressive 2.3 million whole-slide images. Crucially, the clinical dialogue supervision is derived from 685,507 pathology reports meticulously collected by Memorial Sloan Kettering Cancer Center (MSKCC) during routine patient care. This real-world data was then ingeniously converted into question-and-answer pairs by the advanced GPT-4o model, ensuring the dialogue component reflects authentic clinical reasoning.
Architecture and Embedding Design: A Two-Stage Approach
PRISM2’s operational framework is elegantly divided into two distinct phases
Original article, Author: Samuel Thompson. If you wish to reprint this article, please indicate the source:https://aicnbc.com/24455.html
