Microsoft and Paige Launch PRISM2: Pathology AI Trained on 2.3M Images.

PRISM2 Pathology Model Matches Clinical-Grade Products at 0.967 Pan-Cancer AUC.

PRISM2: Teaching Pathology AI to Talk Like a Pathologist:

Paige and Microsoft's new model reads whole-slide images through clinical dialogue, not just pixel classification — and it matches clinical-grade products on their own tests.

2.3M: Whole-slide images used in training

0.967: AUC on pan-cancer detection, diagnostic embedding

685,507: Pathology reports converted into training dialogue

1: A Pathology Model Built to Answer Questions, Not Just Classify Pixels:

Most pathology AI has been trained to sort tissue into categories. PRISM2 was trained to talk about what it sees.

PRISM2, built by Paige and Microsoft, reads whole-slide pathology images through a perceiver-based encoder trained jointly on tissue tiles and clinical dialogue pulled from real pathology reports. Instead of stopping at pixel-level classification, the model aggregates thousands of tile embeddings from a single slide into one unified representation, then generates text that directly answers diagnostic questions.

The training set behind it is enormous: 2.3 million whole-slide images, paired with dialogue supervision drawn from 685,507 pathology reports collected during routine care at Memorial Sloan Kettering Cancer Center. GPT-4o converted those reports into question-and-answer pairs, effectively teaching the model to communicate findings the way a pathologist would document them.

2: Inside the Two-Stage Architecture:

PRISM2 separates “seeing” from “explaining” into two distinct training stages — and that split shapes everything downstream.

Stage one trains the slide encoder itself, teaching it to condense tile-level features into a single slide-level vector that correlates with the language used in real reports.

Stage two freezes that encoder completely and shifts all further training onto the language model, fine-tuning it on clinical dialogue so it absorbs pathology reporting conventions rather than encoder mechanics.

Notably, that second stage relies entirely on single-turn dialogue — there's no multi-turn conversation history in the training signal, which limits the interactive back-and-forth a deployed system could support without further engineering.

At the center of stage one, a perceiver-based encoder aggregates Virchow2 tile embeddings using two loss functions running in parallel. BioGPT text embeddings drive a contrastive objective that pulls slide representations toward matching report language and away from mismatched ones. Phi-3 Mini runs an autoregressive objective alongside it, pushing the encoder's output to support direct text generation rather than similarity scoring alone.

It's a deliberate bet: contrastive training alone tends to produce embeddings good at retrieval but weak at generation, while autoregressive training alone risks overfitting to surface text patterns instead of transferable visual features.

● Base embeddings — pulled straight from the slide encoder, best suited for biomarker prediction.

● Diagnostic embeddings — pulled from the 4-billion-parameter language model's hidden state, tuned for cancer detection and subtyping.

● A third, separately fine-tuned embedding built specifically for survival prediction tasks.

3: How PRISM2 Stacks Up Against Clinical-Grade Tools:

The benchmark numbers put PRISM2 in direct competition with products already calibrated for clinical use.

On prostate and breast cancer detection, PRISM2 matches or exceeds the balanced accuracy of clinical-grade products, tested on those same products' own evaluation datasets. On breast lymph node classification, it outperformed Paige BLN outright, without any additional training on that specific task. Earlier foundation models in the comparison — PRISM and TITAN — both fell short of product-level performance, with the gap widening further on breast lymph node testing.

On pan-cancer detection, PRISM2's diagnostic embedding reached 0.967 AUC, ahead of its own base embedding at 0.956, and ahead of PRISM (0.947) and TITAN (0.931). Rare cancer detection told a slightly different story: the diagnostic embedding's score dropped to 0.957 AUC, which researchers attribute to sparser training examples for those tissue types.

Under linear probing — the cleanest test of representation quality, since it freezes the encoder and checks whether a simple classifier can still extract the right signal — PRISM2's embeddings never statistically underperformed any prior foundation model across the diagnostic benchmarks tested.

— Key finding, PRISM2 benchmark evaluation

Survival prediction followed a similar pattern. Researchers compiled more than 225,000 cases tracking overall survival across nearly 100,000 patients, then pitted a fine-tuned PRISM2 slide encoder against a survival specialist model trained from scratch on identical data — and PRISM2 won. The widest margin came on MSK colorectal cancer recurrence-free survival, where PRISM2 posted a 0.809 concordance index against 0.773 for the specialist model.

Base embeddings, without any survival-specific fine-tuning, held their own too, and on biomarker tasks outside the report-dialogue training distribution, base embeddings actually outperformed diagnostic ones — averaging 0.854 AUC on MSK tasks and 0.784 on TCGA tasks.

An ablation study isolated exactly what the dialogue supervision contributed. Adding dialogue templates to the original PRISM baseline lifted prompt-based inference from roughly 0.498 balanced accuracy to 0.653 — and the underlying question-answering dataset is 3.5 times larger than the PRISM subset it builds on. Researchers attribute about half of PRISM2's diagnostic improvement to that scale increase alone, separate from any architectural change.

4: Where the Model Still Has Limits:

A published model is not the same as a deployment-ready one, and PRISM2's own evaluation is candid about the gap.

A pathologist manually reviewed 50 held-out specimens across 10 tissue types, checking both the generated training text and PRISM2's own outputs. Ground-truth question errors landed around 3 percent for open-ended and multiple-choice formats.

Diagnostic summaries fared worse, at an 8 percent error rate, and complementary yes/no questions performed worst of all — 18 percent were irrelevant or inaccurate. PRISM2's own answers carried a 7 to 11 percent error rate in that same review, with hallucination and omission as the dominant failure modes rather than outright contradiction of the slide.

● No position encoding across tiles — the model can't reason about where structures sit relative to each other on a slide.

● Every training and test scan ran at a single fixed resolution, leaving variable-magnification performance untested.

● All training and evaluation slides came from one institution's scanning pipeline, so external validation is still needed before deployment.

● Model weights are public on Hugging Face, but the full training and inference pipeline still depends on proprietary Paige and Microsoft infrastructure.

For any team evaluating PRISM2 for their own use, the practical takeaway is straightforward: test embedding transfer against your own scanner output before assuming it will match the results reported on the MSK-trained baseline.

The Real Lesson From PRISM2 Isn't Just for Pathology Labs.

PRISM2's biggest gain didn't come from a bigger model — it came from grounding AI in the language professionals already use to describe their work, then validating every claim against real-world review.

That's the same principle behind Otherworlds AI's Agent+ Business AI Platform: AI that understands your business in your own operational language, checked against real outcomes, not just built to look impressive in a demo.

Whether it's diagnostic dialogue or customer conversations, the businesses that win are the ones that ground AI in domain expertise instead of bolting it on.

See how Agent+ grounds AI in your business at otherworldsai.com