AI model PRISM2 matches clinical-grade pathology tests without additional training

Scientists report today in Nature Medicine that a new AI model, PRISM2, trained on millions of pathology images and clinical reports, can detect cancer in prostate, breast, and lymph node samples with accuracy matching or exceeding commercial clinical-grade products—without any task-specific fine-tuning. The work was led by senior authors Kristen Severson and Siqi Liu at Paige and Tempus, with first author Eugene Vorontsov.

PRISM2 is a slide-level foundation model trained on 2.3 million whole-slide images and 14 million question–answer pairs derived from 700,000 pathology reports. Moving beyond models that only process image tiles or lack dialogue supervision, PRISM2 learns to align histomorphology with diagnostic reasoning through clinical dialogue. Its dual-embedding architecture produces a general-purpose "base" embedding and a "diagnostic" embedding tuned for cancer detection and subtyping.

Clinical-grade performance without retraining
Using prompt-based inference—where the model answers yes/no or multiple-choice questions—PRISM2 matched the balanced accuracy of Paige Prostate and Paige Breast, and outperformed Paige BLN, all of which are commercial products calibrated for clinical use. The model also excelled at pan-cancer detection, subtyping breast, lung, and kidney cancers, and tumor staging in lymph node resections. In head-to-head comparisons, PRISM2 embeddings outperformed those of previous slide-level models (PRISM, TITAN, COBRA, Prov-GigaPath) on most diagnostic tasks, often by a wide margin.

Broader applications: survival and biomarkers
Beyond diagnosis, PRISM2 base embeddings performed well on biomarker prediction (e.g., identifying genetic alterations from H&E slides) and survival prediction. When fine-tuned on a large survival dataset of nearly 100,000 patients, PRISM2’s survival embeddings outperformed a specialist model trained from scratch, achieving a concordance index of 0.809 for colorectal cancer recurrence-free survival vs. 0.773.

Limitations and next steps
The model’s performance on biomarker tasks was only modestly better than baselines, and its spatial reasoning is limited because it lacks position encoding. The authors note that further improvements in prognostic and biomarker prediction may require including those tasks during pretraining. Additionally, PRISM2 was trained only on slides scanned at a single center, so robustness across institutions needs more study.

Despite these caveats, PRISM2 demonstrates that language-supervised pretraining can scale to produce clinically meaningful pathology representations. The work suggests that the next frontier for computational pathology may be moving beyond diagnostic tasks to tackle biomarker and prognostic prediction at scale.