One Model Reads an Entire Pathology Slide: PRISM2 Cancer Detection Catches Up With Clinical-Grade Specialized Systems

The challenge of digital pathology has never been merely teaching computers to recognize a few cells. It is about reading a whole-slide image containing billions of pixels as a complete case, as a pathologist would. A study published in *Nature Medicine* shows that the multimodal foundation model PRISM2 can now directly answer cancer-related questions and, in several tests, has caught up with clinical-grade systems developed for specific cancer types, providing more concrete evidence that “one model can support multiple pathology tasks.”

PRISM2 was jointly developed by Paige and Microsoft Research and trained on more than 2.3 million hematoxylin-and-eosin-stained whole-slide images, covering approximately 685,000 specimen reports. The research team converted the reports into approximately 14 million clinical question-and-answer pairs, first aligning image representations generated by Virchow2 with diagnostic text and then conducting clinical dialogue training using the Phi-3 Mini language model. As a result, the model does more than output image features: it can also answer yes-or-no, multiple-choice, and open-ended questions, as well as generate reports.

Without additional training for individual cancer-detection tasks, PRISM2 used yes-or-no questions to achieve balanced accuracy matching Paige’s specialized clinical products in prostate and breast cancer tests, and it outperformed the comparison system in detecting cancer in breast lymph nodes. These tests used evaluation data the model had not seen during training. However, the researchers specifically avoided describing this as strict “zero-shot” capability because the relevant disease concepts had appeared in the training reports. A more accurate description is that the model can perform prompt-based inference without requiring a separate specialized classifier for each task.

PRISM2’s applications are not limited to screening for cancer. Its slide representations were used for cancer subtyping, grading, lymph node staging, predictions of certain biomarkers, and survival prediction. Across multiple linear-probe evaluations, its performance was not significantly worse than that of existing pathology foundation models. After further fine-tuning for survival prediction, it also outperformed specialized models trained from scratch. In the future, this architecture could help triage cases, prepopulate structured reports, or flag cases requiring further immunohistochemistry and sequencing.

However, matching the level of clinical-grade products in benchmark tests does not mean that the model itself has obtained the same clinical qualifications. The study conducted a retrospective analysis using de-identified data, and the slides seen by both PRISM2 and the underlying image model were scanned by the same cancer center. Differences among hospitals in staining, scanners, case mix, and workflows could still cause performance shifts. It lacks complete global spatial reasoning and uses only a single resolution, which may limit lesion counting, size measurement, and certain staging tasks.

The language output requires even greater caution. Spot checks in the study showed that the report-derived training questions and answers themselves contained noise. When generating diagnostic summaries, the model still produced fabricated findings, omissions, and grading errors, and certain report fields had to be calibrated before interpretation could be improved. Paige has made PRISM2 available for licensing and enterprise integration and has incorporated it into Alba, a clinical assistant that remains in testing, but the company also explicitly states that its generated content cannot replace medical diagnosis or treatment.

The real next threshold is to demonstrate, within multicenter prospective clinical workflows, that the model can maintain safe, stable, and traceable performance, while clarifying human review, responsibility for errors, and its regulatory status. PRISM2’s significance does not lie in declaring that general-purpose pathology AI is already mature, but in showing that combining large-scale images with diagnostic language may indeed narrow the gap between general-purpose foundation models and specialized clinical tools.