PRISM

Pathology Research Insights from Slide Models

AI estimates from the slide image, released only after a person at the lab has looked.

MSI from H&E today; breast, lung and more on the roadmap
Researchers watch the run live, then get maps and scores
Citable methods, and questions answered by AI or the lab

PRISM · MSI from H&E

Every tile casts a vote, and you can see where

MSI-high colon cancers respond differently to treatment, and the standard tests take extra tissue and time. Our model estimates MSI status from the routine H&E slide, for research, and shows which parts of the tissue drove the estimate.

AI map of a public colon slide, mostly red: leaning MSI-high — **1.00** leaning MSI-high · published test: MSI-high · 5,996 tiles

AI map of a public colon slide with blue across the tumour: leaning MSS — **0.00** leaning MSS · published test: MSS · 6,000 tiles

pushes toward MSI-highpushes toward MSSPublic TCGA slides TCGA-AA-3710 and TCGA-NH-A50T, each scored by models that never saw its hospital.

A tile of colon tissue, 224 pixels square

Cut into tiles

Each slide becomes thousands of small squares of tissue, 224 pixels at about 20x. Background is skipped.

Fingerprinted

An advanced pathology foundation model reads each tile and captures what the tissue looks like as a compact numeric fingerprint.

Weighed and voted

Our AI attention model, trained on public slides, learns which tiles matter. Their pushes add up to the score.

Questions get answered, by the AI or by the lab

Researchers can ask about their released results in plain English. Gemini answers from the released scores, the model card and the methods, nothing else. Anything only the lab can see goes to the lab with one button, and the lab's reply arrives in the same thread.

What does borderline mean for my slides?

AIBorderline is a score from 0.40 to 0.60, where the model is close to undecided. Two of your slides fall there, so treat them as uncertain rather than as a lean either way.

Ask the lab

From the labOn slide 4 the red and blue are mixed through the tumour. We would confirm it with mismatch-repair staining.

Illustration. The lab reads every question.

Evidence so far

Tested the hard way: on hospitals the model never saw

0.81AUC on unseen hospitals (95% range 0.73 to 0.89)

Colonfirst tissue; breast and lung next

22hospitals, each held out in turn

543labelled TCGA colorectal patients available for the next phase

Phase 1 · done

Feasibility

MSI on public colon slides, tested hospital by hospital, and the full loop from request to released map running end to end.

Phase 2 · next

Scale and validate

More public slides and hospitals, external validation on an independent cohort, and new tissues, starting with breast and lung.

Phase 3

Each lab's own slides

Local validation on a partner lab's scanner and stain, using its own mismatch-repair results as labels. A model that gets better with every lab that joins.

These are research results on public data. The AI estimate is not a validated clinical test and is not used for patient care. Demonstrations use made-up (synthetic) cases.

See it on your own workflow

We will walk you through custody, research access and the lab AI with synthetic data, and talk about what your lab needs.

LabsTrack every specimen; open your archive to research on your terms.

ResearchersFind tissue, request it, and get AI estimates you can cite.

PartnersHelp validate the next models on your own slides.

Write to

renee.hoyle@labrunner.app

Open email

Tell us about your lab and what you would like to see.