The Pathologist’s New Microscope: How Foundation Models Like Virchow 2, CHIEF, and Prov-GigaPath Are Quietly Transforming Cancer Diagnosis in 2026

For more than a century, the diagnosis of cancer has rested on a craft as old as modern medicine itself. A surgeon removes a piece of tissue. A laboratory stains it pink and purple. A pathologist leans over a microscope and renders judgment. The verdict, written by hand on a slip of paper and later typed into an electronic record, sets in motion every subsequent decision a cancer patient will face. Whether to operate. Whether to irradiate. Whether to begin chemotherapy. Whether to wait.

Presented By Our Partners

In 2026, that century-old craft is undergoing the most consequential transformation in its history. The microscope has not been replaced. The pathologist has not been replaced. But the visual field has been quietly extended by a new layer of intelligence that can see patterns no human eye can resolve and remember every slide it has ever encountered. The technology is called the pathology foundation model, and it is moving from research papers into hospital workflows at a pace that has surprised even its developers.

This week, the U.S. Food and Drug Administration extended its 510(k) clearance for ArteraAI Breast, a digital pathology platform that uses deep learning to predict recurrence risk in patients with early-stage hormone receptor-positive breast cancer. ArteraAI Breast is the first FDA-cleared digital pathology-based risk stratification tool for the disease. It runs on routine surgical resection samples and returns same-day prognostic information that clinicians can use alongside standard reports. The clearance arrived less than five months after the FDA cleared an enterprise digital pathology platform from Indica Labs and Leica Biosystems for in vitro diagnostic use, a regulatory inflection point that has prompted academic medical centers from Memorial Sloan Kettering to Mass General Brigham to accelerate their plans for clinical deployment of foundation models.

To understand what is happening, it helps to begin with what these models actually are.

What a Pathology Foundation Model Is

A foundation model is a single, very large neural network trained on a vast and varied corpus of data using self-supervised learning. Unlike older medical AI tools, which were trained to perform one narrowly defined task such as flagging mitotic figures or identifying lymph node metastases, a foundation model learns a general representation of the underlying visual world. Once that representation has been learned, the model can be adapted with relatively little additional data to perform almost any downstream task in the same domain.

The idea was popularized in natural language by models such as GPT and in computer vision by models such as CLIP. Its translation into medicine has been slower and more difficult, because medical data is fragmented, privately held, and exquisitely sensitive. Pathology has nevertheless emerged as the medical specialty best positioned to benefit. A single hematoxylin and eosin slide, when digitized at clinical resolution, can contain ten billion pixels. A whole-slide image is essentially a multi-gigabyte visual document. Across the world’s pathology laboratories, hundreds of millions of such documents have accumulated over the past two decades of slide digitization. That is the raw material on which the new generation of pathology foundation models has been trained.

Virchow and Virchow 2: The Memorial Sloan Kettering Engine

The most widely studied pathology foundation model is Virchow, developed by Paige in collaboration with Memorial Sloan Kettering Cancer Center and Microsoft. The original Virchow, described in Nature Medicine in 2024, was trained on approximately 1.5 million hematoxylin and eosin whole slide images drawn from 100,000 patients at Memorial Sloan Kettering. Built as a vision transformer with 632 million parameters and trained using the DINOv2 self-supervised algorithm, Virchow achieved a 0.95 area under the curve in detecting sixteen cancer types, including several rare malignancies for which conventional tissue-specific models had failed to reach clinical-grade performance.

The follow-on model, Virchow 2, was trained on more than three million slides from 800 laboratories worldwide and was designed explicitly to generalize beyond Memorial Sloan Kettering’s scanner and staining protocols. In published benchmarks released over the past twelve months, Virchow 2 has matched or exceeded the performance of bespoke clinical models on tasks ranging from biomarker quantification to outcome prediction. The strategic significance is that a single backbone, fine-tuned with a few hundred or a few thousand labeled examples, can replace dozens of one-off models, each of which would otherwise require its own development cycle, validation study, and regulatory submission.

Paige has begun licensing fine-tuned Virchow derivatives to community pathology laboratories. The company’s prostate detection product, Paige Prostate, became the first FDA-authorized AI software for prostate cancer detection back in 2021. The new generation of Virchow-derived clinical products is moving through the same regulatory pipeline.

Prov-GigaPath and the Microsoft Research Approach

In parallel, Microsoft Research and Providence Health published Prov-GigaPath in Nature in 2024, an open-access pathology foundation model trained on more than 170,000 whole slide images drawn from real-world clinical practice across Providence’s twenty-eight hospital network. Prov-GigaPath combined whole-slide context with individual image tiles using a novel hierarchical architecture, which allowed the model to learn both fine cellular detail and broader tissue-level organization.

What distinguished the Prov-GigaPath effort was the deliberate choice to use real-world, decentralized clinical data rather than curated academic data. The investigators argued that this approach would produce a model better adapted to the variability of slides processed in the community settings where most patients are actually treated. The model has since been integrated into Providence’s pathology workflow and made openly available to researchers, an unusual move in a field dominated by proprietary commercial efforts.

CHIEF: The Harvard Approach to Pan-Cancer Intelligence

The third major pathology foundation model is CHIEF, the Clinical Histopathology Imaging Evaluation Foundation, developed at Harvard Medical School and the Brigham and Women’s Hospital by the laboratory of Faisal Mahmood. CHIEF was pretrained on 44 terabytes of pathology imaging data and validated on 19,491 whole slide images drawn from 32 independent datasets across multiple cancer types.

Featured Partner

Invest in the Infrastructure Behind Modern Medicine

As healthcare expands beyond hospital walls, the buildings and campuses supporting that shift are generating compelling returns for investors who move early. The Healthcare Real Estate Fund offers qualified investors direct access to a curated portfolio of medical office, outpatient, and specialty care facilities.

Learn More →

The Harvard team designed CHIEF to perform four conceptually distinct tasks within a single model. It identifies the presence and likely tissue of origin of a tumor. It characterizes molecular profiles such as gene expression and mutation status directly from the histological image. It predicts patient prognosis. And it provides interpretability outputs that allow a pathologist to understand which regions of the slide drove the model’s conclusions. CHIEF and its successor models have become benchmarks against which other pathology foundation models are now measured.

A separate Harvard effort produced PathChat, a multimodal pathology assistant that allows clinicians to query whole slide images in natural language. A pathologist can upload a slide and ask, in plain English, whether the lesion shows evidence of perineural invasion or whether the tumor margin appears clean. PathChat returns a structured answer with the relevant regions of the slide highlighted. The model is being piloted at several academic medical centers as a teaching and second-opinion tool, not yet as a primary diagnostic instrument.

UNI and the University-Led Models

Beyond the three flagship efforts, a growing constellation of academic models has entered the field. UNI, developed at Harvard with public benchmarks, has become a widely cited baseline. H-Optimus-1, released by Bioptimus in 2024, was trained on more than 500,000 slides from more than 7,000 patients and is openly available. UNI2, an updated successor, has been integrated into several open clinical benchmarks. In Asia, BEPH and Phikon have emerged as strong contenders, validating that the foundation model paradigm is not the exclusive province of American or European institutions.

This proliferation matters for a practical reason. The choice of foundation model can substantially affect the performance of downstream clinical applications. A recent clinical benchmark published in Nature Communications in 2025 compared more than a dozen public foundation models across cancer detection, biomarker prediction, and outcome tasks, and found significant variation across models and tasks. The community is now converging on a small number of preferred backbones, and the choices being made in 2026 will shape the architecture of computational pathology for the next decade.

The ArteraAI Breast Story: From Foundation Model to Clinical Decision

The recent FDA clearance of ArteraAI Breast illustrates how foundation models are translating into bedside decisions. The product takes a digitized whole slide image of a routine breast cancer surgical sample and applies a deep learning model trained on tens of thousands of cases with long-term outcome data. The model produces a personalized recurrence risk score that complements the established Oncotype DX and MammaPrint molecular assays.

The clinical question that ArteraAI Breast is designed to answer is one of the most consequential in early-stage breast cancer care. For a woman with hormone receptor-positive, HER2-negative disease, the central decision after surgery is whether chemotherapy will meaningfully reduce her risk of recurrence. Molecular assays such as Oncotype DX have answered this question well for many years, but they require shipping a tissue block to a central laboratory, waiting up to two weeks for results, and paying several thousand dollars per test. ArteraAI Breast returns a result in hours and runs on the same slides the pathologist is already reviewing.

If the prognostic information from ArteraAI Breast tracks the molecular assays in the real world, the implication is that a meaningful share of breast cancer patients could receive comparable decision support at a fraction of the cost and a fraction of the wait. ArteraAI’s earlier products in prostate cancer have already been adopted by several large oncology networks. The breast clearance opens a far larger market and signals to other developers that the FDA is willing to clear digital pathology biomarkers built on deep learning when the underlying validation is sufficient.

The Quiet Revolution Beneath the Glass

These developments matter beyond the immediate clinical utility because they represent a structural change in how diagnostic medicine produces knowledge. For most of the twentieth century, the development of a new diagnostic tool required a long sequence of bespoke human effort. Pathologists defined a feature of interest. Researchers built a classifier to detect it. The classifier was validated on a specific tissue type, scanner, and patient population. The result was a narrow tool that worked well in the laboratory where it was developed and often struggled to generalize.

The foundation model paradigm inverts that sequence. The expensive work of training a general visual representation is done once, on the largest possible corpus of data. The remaining work of adapting the model to a specific clinical question becomes far cheaper, faster, and more reproducible. A single foundation model can support hundreds of downstream applications across dozens of cancer types, and each new application benefits from improvements made to the underlying model.

The practical effect for pathologists is that the rate at which new computational tools become available is accelerating. The practical effect for patients is that those tools are beginning to be deployed in clinics rather than published in journals.

What Still Stands Between Promise and Practice

The technology is not without serious obstacles. Pathology foundation models have been shown to inherit biases from the populations on which they were trained, and the laboratories that have contributed the most data to date are concentrated in a small number of well-funded academic medical centers. Performance on slides from community laboratories with different scanners and staining protocols has been variable. The field has responded with deliberate efforts to expand training data, federated learning approaches that allow models to learn from data they never directly see, and benchmarks such as PANDA-PLUS-Bench that test models against realistic clinical variability.

Regulatory questions remain complex. The FDA has cleared a small number of pathology AI products as software as a medical device, but the question of how to regulate a general purpose foundation model that supports many downstream applications has not been fully resolved. Most clinical deployments today rely on fine-tuned, locked versions of the underlying foundation model, evaluated against the same standards as any other clinical decision support tool. The regulatory architecture for continuously learning models remains an open area of policy work.

There is also the question of workflow. A pathologist who is asked to review the output of a foundation model in addition to the slide itself faces a new cognitive task that has its own learning curve. Studies of radiology AI have shown that the way in which model outputs are presented can substantially affect how clinicians interpret them. The pathology community is investing in human factors research and in the design of decision support interfaces that surface model conclusions without overwhelming the clinician.

Finally, there is the question of equity. The cost of integrating a foundation model platform into a pathology laboratory is not trivial. Smaller community hospitals, which already operate on thin margins, may struggle to adopt the technology even when its clinical value is clear. If foundation models become the standard of care at academic centers without becoming accessible elsewhere, the result could be a widening of the diagnostic quality gap rather than a narrowing of it. Some of the field’s most thoughtful leaders are explicit that the equity question is unresolved.

What This Means For You

If you or a family member is facing a cancer diagnosis in 2026, several practical implications follow from the developments described above. First, ask whether the laboratory that processed your tissue uses a digital pathology platform. Digital slides can be reviewed remotely by second opinions, archived efficiently, and analyzed by AI tools. Tissue blocks are still the legal record, but the digitized image is increasingly where additional analysis happens.

Second, ask whether AI-based prognostic or biomarker tools are available for your specific cancer type. ArteraAI Breast and ArteraAI Prostate are the most advanced examples in the United States, but other tools are reaching the market rapidly. The conversation to have with your oncologist is whether the additional information from such a tool would change your treatment plan. If the answer is yes, the result is often worth the additional step.

Third, if you are participating in research or clinical trials, ask whether your slides can be shared for computational analysis. The pathology foundation models being trained today depend on access to large and diverse slide collections, and patient participation in such research is the engine that will make these models better, fairer, and more clinically useful over time.

Fourth, recognize that the pathologist’s role is not being eliminated. It is being augmented. The pathologists who will be most valuable to your care in 2026 and beyond are those who understand how to use the new tools well, how to recognize when the tools are wrong, and how to translate the resulting information into a clear plan you can act on.

Finally, keep perspective. The fundamental fact of cancer care remains the same. A tumor is biology unfolding in a particular person, and the work of diagnosis is to understand that biology well enough to interrupt it. Foundation models extend what the human eye can see and what the human memory can retrieve. They do not change the underlying problem, and they do not change the centrality of the relationship between patient and physician. They are, in the end, a new and powerful instrument added to a long lineage of instruments that have made the practice of medicine more accurate, more efficient, and more humane.

The microscope is still on the pathologist’s desk. But beside it now sits something new, and what it can see is changing the future of cancer care.

Free Daily Briefing

The Latest Longevity Science.
Delivered Every Morning.

Join researchers, physicians, and health professionals getting daily breakthroughs in AI-driven medicine, epigenetics, and longevity research.

Support the research that powers this editorial

Subscribe Free

No spam. Unsubscribe anytime. We respect your inbox.