Pathology AI Is Catching Cancers That Humans Miss

A pathologist in a U.S. community hospital sees roughly 60 to 80 slides a day. Each one is a gigapixel image — typically 100,000 by 100,000 pixels at peak zoom — that has to be scanned, triaged, and diagnosed in minutes. Miss a subtle mitotic figure or an early-stage tumor bed, and a patient loses months of treatment runway. In April 2026, a new class of AI systems called pathology foundation models has started closing that gap in ways that are hard to ignore.

Microsoft’s Prov-GigaPath, Paige’s Virchow2, and the Mahmood Lab’s UNI have become the workhorses of this shift. Trained on millions of whole-slide images and pulled into clinical workflows at Mayo Clinic, Memorial Sloan Kettering, and Charité Berlin, these models are now detecting cancers that human pathologists routinely overlook — especially the rare ones.

What Pathology Foundation Models Actually Do

Traditional computational pathology used task-specific models: one CNN for prostate detection, another for breast lymph node metastases, another for colorectal microsatellite instability. Every new indication required a fresh dataset, fresh labels, and six to twelve months of retraining.

Foundation models flip that. Pretrained on millions of unlabeled slides using self-supervised learning, they produce embeddings that work across dozens of cancer types and downstream tasks with minimal fine-tuning. Virchow2 was trained on 3.1 million whole-slide images from 225,000 patients across 45 countries. Prov-GigaPath ingested 1.3 billion image tiles from 171,189 slides and uses a novel LongNet architecture purpose-built for gigapixel scans. UNI, published in Nature Medicine, was fine-tuned on 100 million tiles and shipped open-source — which is why you now see it baked into dozens of hospital pilot programs.

The performance numbers are striking. On rare-cancer detection, clinical-grade foundation models now achieve 0.95 specimen-level AUROC across nine common and seven rare cancers. Benchmark evaluations from the Broad Institute and Microsoft Research put Virchow2 first overall on seven oncology tasks, with Prov-GigaPath and UNI close behind. In biomarker prediction — microsatellite instability, HER2 status, tumor mutational burden directly from H&E slides, no genomic test required — the gap between these models and human experts has either closed or flipped.

The Rare Cancer Problem Is Where It Shows Up First

The hardest slides are the rare ones. A community pathologist might see one case of angiosarcoma or one carcinoid tumor every two or three years. Human accuracy on rare tumor identification sits around 60 to 70 percent on first-read. Foundation models, because they pool training signal across a massive long-tail of histologies, do not have the same forgetting curve.

In a 2025 evaluation run by researchers at Memorial Sloan Kettering, Virchow flagged rare hematolymphoid malignancies that had been missed on original read. Several cases were later confirmed by molecular testing. This pattern — the AI catching cancers on re-review — is quietly becoming the dominant clinical use case. Hospitals aren’t replacing pathologists with AI. They’re running every slide through a foundation model as a second-read quality net before sign-out.

This use pattern fits neatly alongside a broader push toward AI-driven biomarker discovery in oncology, where the same H&E images are now being mined for information that used to require separate, expensive assays.

FDA Is Moving — Slowly

The regulatory picture is less dramatic. As of early 2026, the FDA has cleared only a handful of AI/ML Software-as-a-Medical-Device tools specifically for histopathology. Paige Prostate Detect remains the flagship clearance. Every cleared tool requires full pathologist oversight for the final diagnosis — the AI is never the legal decision-maker.

What has changed is the FDA’s PathologAI initiative, launched through the National Center for Toxicological Research, and the agency’s 2026 framework that redefines what counts as a breakthrough AI medical device. The new posture treats foundation-model-backed diagnostics as adaptive products rather than static ones. Expect performance monitoring, post-market surveillance, and population-drift audits to become conditions of approval.

Europe is slightly ahead in deployment. The Atlas project — a Korea-Germany-U.S. collaboration using 1.2 million whole-slide images across 490,000 cases from Mayo Clinic and Charité — is already being used for research and retrospective review at participating centers. Expect the first prospective clinical trials with Atlas to report out in late 2026.

Copilot Mode Is Winning Over Full Autonomy

One pattern has become clear to anyone inside these pilots: the clinical value is not autonomous diagnosis. It is the copilot.

A pathologist using a foundation-model copilot gets three things that matter: heatmap localization showing where on a 100,000-pixel slide to look; a ranked differential list with probabilities; and a biomarker readout that would otherwise have required a separate immunohistochemistry or molecular test. Turnaround time drops. Inter-observer agreement goes up. Missed findings — the thing that generates malpractice suits and, far worse, patient harm — drops sharply.

This is the same pattern we have seen in radiology over the past eight years. The AI doesn’t replace the specialist. It gives the specialist better eyes.

The Multimodal Wave Is Already Forming

The frontier of 2026 is multimodal pathology. Instead of models that ingest only histology images, the next generation is pulling in genomic sequencing, clinical notes, radiology images, and proteomic readouts in a single pass.

Microsoft’s GigaPath team has already shown early results combining pathology foundation embeddings with matched RNA-seq data to predict treatment response. Early readouts suggest these multimodal systems outperform either modality alone on survival prediction and treatment selection by a meaningful margin — and they pair naturally with companion diagnostics AI that matches cancer patients to drugs.

For oncologists, this matters because cancer has always been a multimodal disease. A slide alone tells you the histology. A gene panel alone tells you the mutations. A clinical note alone tells you the patient’s trajectory. The first model that reads all three at once and weighs them correctly is going to set the new bar for precision oncology.

The Bottlenecks That Still Matter

None of this works without three things that remain scarce.

Slide scanning infrastructure. Roughly 10 to 15 percent of U.S. pathology departments have fully digital workflows. The rest still cut glass slides and pass them around physically. Foundation models are only as useful as the scanner pipeline feeding them, and a Philips IntelliSite or Leica Aperio scanner runs $150,000 to $300,000 before you count the storage and network costs.

Dataset diversity. The largest foundation models were trained heavily on academic medical centers in the U.S. and Europe. Performance on slides from Brazil, Nigeria, or Vietnam — prepared with different staining protocols, different scanners, different tissue fixation practices — is measurably worse. The field is still working out how much of a generalization gap actually exists and how much can be fixed with fine-tuning on local data.

Reimbursement codes. There is still no dedicated CPT code for AI-assisted pathology review in most U.S. payer contracts. Hospitals running these tools are doing so on research budgets or as loss leaders. Until CMS creates a reimbursement path, adoption will stay limited to academic centers and cancer-focused networks that can afford the capital outlay.

Who Is Actually Shipping This

The commercial field has narrowed fast. Paige built Virchow and sells an end-to-end prostate and breast cancer diagnostic suite with FDA-cleared components. PathAI focuses on clinical trial enrollment and pharma partnerships, using foundation models to find patients who match rare molecular subtypes. Proscia sells the digital slide management platform that a lot of these models plug into. Ibex Medical Analytics, based in Israel, has aggressive European deployments.

The academic open-source camp — UNI from Mass General Brigham, CHIEF from the Broad Institute — is where most of the methodological innovation still happens. Commercial products lag open-source benchmarks by roughly 6 to 12 months, which is narrowing as partnerships between health systems and vendors deepen. The result is that a pathologist at a mid-tier cancer center in 2026 can realistically get access to 90 percent of the top-benchmark performance through either a vendor contract or a research collaboration agreement, which was not true even in late 2024.

What to Watch in the Next 12 Months

Three things will decide whether pathology foundation models become standard of care or stay a specialist tool.

First, the Atlas prospective trial readouts expected in Q3 and Q4 2026. A clean positive result on treatment-relevant endpoints — not just accuracy, but measurable impact on patient outcomes — would be the first real clinical validation of a foundation model in histopathology.

Second, watch for a pan-cancer FDA clearance. Existing clearances are single-indication. A foundation model approved for multi-indication use, even with pathologist oversight, would be a structural shift.

Third, the consolidation play. Paige, PathAI, Proscia, and Ibex all have commercial deployments. Foundation-model-as-a-service is going to attract a hyperscaler — Microsoft is already halfway there with Prov-GigaPath — and that acquisition or partnership will likely reshape who controls the data flywheel for the next decade.

The Bottom Line

Pathology foundation models are not theoretical anymore. They are catching cancers humans miss, they are shrinking turnaround time on slide review, and they are being adopted as second-readers in major cancer centers. The clinical value is real, measurable, and now being published in peer-reviewed journals on a monthly basis.

What they are not, in 2026, is standalone diagnostic tools. The copilot pattern is winning, the FDA is cautious, and reimbursement lags. But for cancer patients whose slides would otherwise sit in a stack of 70 waiting for a busy pathologist at 6 p.m., having a foundation model flagging the subtle finding first is already the difference between catching it and missing it.

The question for the next twelve months is not whether pathology AI works. It is how fast health systems can afford to deploy the scanners, storage, and workflows needed to actually use it.