AI in Pathology: What’s Actually Changed, and What Comes Next
For years, AI in pathology was mostly a research story: strong benchmarks and promising papers. Over the past two years, AI-based pathology tools have earned FDA clearance, been incorporated into NCCN guidelines, and begun appearing in real clinical workflows at cancer centers across the country. This isn’t a “someday” technology anymore. It’s being ordered, reimbursed, and used to help make treatment decisions right now.
For researchers, pathologists, and lab directors trying to keep up, here’s a straightforward look at where things actually stand: which tools are cleared, what foundation models are doing under the hood, how deployments are playing out in practice, and why the specimens behind all of this matter as much as the algorithms do.
FDA-Cleared Diagnostic Tools Have Moved From Rare to Routine: Pathology AI News
As recently as mid-2025, a review in The Lancet Digital Health on artificial intelligence pathology pointed out that despite more than 1,200 AI-enabled devices holding FDA clearance across all of medicine, only two, Paige Prostate and Galen/Ibex Second Read, had been cleared specifically for histopathology. Paige Prostate, cleared back in 2021, was the first FDA-approved AI application for histopathology, and for years, it was basically alone out there.
That’s changed fast. In August 2025, the FDA granted de novo authorization to ArteraAI Prostate, the first AI-powered digital pathology tool cleared to give both prognostic and predictive information in localized prostate cancer. It looks at a biopsy whole-slide image alongside clinical variables to estimate ten-year metastasis risk, prostate cancer-specific mortality, and how likely a patient is to benefit from adding hormone therapy to radiation. It’s now in NCCN Prostate guidelines (category 2A), and Medicare has covered it since January 2024.
Then, in May 2026, the same company, Artera, got FDA clearance for ArteraAI Breast, the first and only FDA-cleared digital pathology tool for breast cancer risk stratification, a good sign of where AI in pathology is headed next. It’s built for early-stage, hormone receptor-positive, HER2-negative invasive breast cancer, and it uses digitized histopathology images plus clinical variables to generate a risk score that helps sort patients into risk groups. That’s meant to help with one of oncology’s genuinely hard calls: whether a given patient is likely to benefit from chemotherapy.
Foundation Models Are Changing What “AI Pathology” Even Means
Behind the headline clearances, something quieter, and arguably bigger, has been happening in computational pathology research: the rise of foundation models for whole-slide image analysis.
Take Virchow2 as an example. Paige and Memorial Sloan Kettering built it by showing an AI model roughly 1.5 million tissue slide images. Once trained, it could detect cancer across 17 different types with strong accuracy, even on rare cancers where AI usually struggles due to a lack of data.
Other labs have built similar tools. CONCH and TITAN paired tissue images with the pathologist’s written notes so the AI learns from both at once.
The bigger shift: instead of building a separate, narrow tool for every task, researchers are now building a few large models that broadly understand tissue, then adapting them as needed.
Real-World Deployments Are Starting to Change Clinical Workflows
Clearances and benchmarks are one thing, but the more interesting story for practicing labs is what’s actually happening on the ground.
Under current CLIA guidance, remote digital review of whole-slide images is now allowed under a lab’s primary CLIA certificate when certain conditions are met, enabling pathology groups to share subspecialty expertise across sites without physically shipping glass slides.
For oncology programs treating localized prostate cancer, this is already showing up as real workflow changes: determining whether ArteraAI ordering is enabled and integrating it into multidisciplinary treatment discussions as an AI-enabled biomarker that complements, not replaces, clinicopathologic risk assessment and genomic testing. That “adds to, doesn’t replace” framing comes up again and again in how these tools are actually being used, and it’s a pretty different tone than the “AI will replace pathologists” talk that dominated a few years ago.
The future of pathology that’s actually taking shape looks less like automation replacing expertise and more like a layered system: foundation models and FDA-cleared tools handling pattern recognition and triage at scale, while pathologists spend their time on the cases and calls that genuinely need a person.
What This Means for Researchers Working With Tissue Samples
Every one of these advances, from FDA-cleared risk-stratification tools to million-slide foundation models, depends on one thing upstream: a supply of high-quality, well-characterized tissue.
Foundation models like Virchow2 and UNI are only as good as the whole-slide images they’re trained on, which is a lesson computational pathology research keeps circling back to. A model trained mostly on common cancer types and typical cases will struggle exactly where Virchow2’s rare-cancer benchmark shows the gap still exists: on tissue types and disease states that don’t appear often in any one hospital’s archive. Building and validating models that actually generalize means obtaining specimens from a genuinely wide range of tissue types, tumor stages, and disease states, not just the common cases in one biobank.
Validation is just as demanding as training. Before a diagnostic AI tool can get FDA clearance, or before a research team can trust a foundation model’s output on a new task, its predictions have to be checked against specimens with confirmed, well-documented ground truth: known diagnosis, tumor type, stage, and grade. Specimens with poor or missing annotations don’t just fail to help. They actively add noise that can hide a model’s real weaknesses, whether anyone notices or not.
Why Specimen Quality Is the Foundation Under the Foundation Models
Formalin-fixed, paraffin-embedded tissue is still the standard for AI digital pathology work because it holds up during long-term storage, which is exactly the stability whole-slide imaging depends on. But FFPE quality isn’t the same everywhere. Fixation time, processing steps, and storage conditions all affect tissue preservation, and inconsistent quality upstream becomes noisy, unreliable data downstream that no model can fully correct.
That’s why the current wave of pathology AI news keeps circling back to specimen sourcing, not just model design. That’s what we provide through our digital pathology services: ethically sourced, U.S.-procured FFPE and fresh-frozen tissue, handled to ISBER best practices and shipped with diagnosis, tumor type, TNM staging, and grade data when available, the same principles we cover in more depth in what digital and computational pathology actually involves.
Whether you’re building a slide library to train a model or need specimens for a validation cohort, the tissue behind the pixels is what makes the result trustworthy.
Keeping Up With AI in Pathology
For anyone tracking digital pathology AI news regularly, this field moves fast; new clearances, new foundation models, and new deployment data show up every few months, so treat any single snapshot, including this one, as a point-in-time picture rather than the final word. What’s clear is the direction: AI in pathology has moved from research promise to regulated clinical tool, and the models driving that shift need exactly the kind of well-characterized tissue that responsible biobanking is built to provide. We’ll keep covering the AI pathology news that matters most to researchers as it happens.
If your research or validation work depends on reliable, well-documented FFPE specimens, place an order with Superior BioDiagnostics and let our team help you source the tissue your next computational pathology project actually needs.