Choosing a pathology foundation model in 2026

No single pathology foundation model wins. What benchmarks, licences, task fit and compute say about UNI, CONCH and Virchow2. Research use, not diagnosis.

There is no single best pathology foundation model. In an independent benchmark published in Nature Biomedical Engineering, Neidlinger et al. compared 19 foundation models on slide-level tasks: CONCH came out on top, with Virchow2 close behind. On gene-expression prediction, the HEST-Benchmark leaderboard, checked 17 September 2026, puts H-optimus-1 on top and CONCH further down. Choose in this order: what the licence lets you do, what the task needs, and what your compute budget can run on gigapixel slides. Then evaluate with patient-level splits, on cohorts the model never saw in pretraining.

I came to this question through my MSc thesis on predicting gene expression from pathology images. The comparisons here are published results from other groups, not my measurements.

What is a pathology foundation model?

A pathology foundation model is an image encoder pretrained on very large collections of histology slides and reused as a general-purpose feature extractor for many downstream tasks. In digital pathology, a scanner turns the glass slide into a whole-slide image (WSI), which is gigapixel-scale, so the usual pipeline has two steps. The foundation model turns small tiles of tissue stained with haematoxylin and eosin (H&E) into embedding vectors, and a lighter model learns the task from those vectors. Neidlinger et al. describe exactly this: features are extracted with each foundation model, then "these vectors were used to train a slide-level aggregator".

What the encoder gives you:

• Cheaper labels. Paired with a multiple instance learning (MIL) head, frozen tile features let a downstream model learn from one label per slide or per patient, not from hand-drawn outlines of tissue regions. Pretrained pathology features are also reported to need fewer labelled examples: the UNI paper says UNI enables "data-efficient" models.

• Features you compute once. Embeddings can be cached and reused across many experiments.

• Shared baselines. Independent groups publish comparisons of the same released encoders.

What it does not give you: a diagnosis, which these models are not for. Nor does it give you a model validated on your cohort and robust to your scanner and staining. Validation and robustness have to be earned, and the rest of this guide is about how.

The landscape in 2026: UNI, CONCH, Virchow2 and others

Vision-only and vision-language encoders

Most of the encoders below are vision-only: they learn from images alone.

• UNI (Mahmood Lab): a vision transformer (ViT-L/16) with 0.3B parameters. Its Nature Medicine paper reports pretraining on more than 100 million images from over 100,000 diagnostic H&E-stained slides. UNI2-h, a larger model from the same lab, is a ViT-H with 681M parameters.

• Virchow2 (Paige and Microsoft Research): a ViT-H/14 with 632 million parameters, trained on 3.1 million whole-slide images from Memorial Sloan Kettering Cancer Center, per its card. Its paper also describes Virchow2G Mini, a 22 million parameter distillation of a larger sibling, Virchow2G.

• H-optimus-0 and H-optimus-1 (Bioptimus): vision transformers with 1.1B parameters each; the H-optimus-0 card reports training on more than 500,000 H&E slides. H0-mini (Owkin and Bioptimus) is a ViT-B/14 distilled from H-optimus-0, listed at 85.7M parameters.

• Prov-GigaPath: a tile encoder paired with a slide encoder. Xu et al. report in Nature that it was pretrained on 1.3 billion tiles from 171,189 whole slides from Providence, a US health network.

• Phikon-v2 (Owkin): a ViT-L pretrained on public slide collections, which matters for the leakage check below.

CONCH (Mahmood Lab) is the vision-language model in this list. Its vision encoder is a ViT-B/16 with 90M parameters, paired with a text encoder, and its Nature Medicine paper reports training on over 1.17 million image–caption pairs. The text side is what vision-only encoders lack: the CONCH card lists zero-shot classification and image–text retrieval among its uses.

Licences and access terms decide first

A model you may not use for your purpose is not a candidate, however it scores. The table summarises the Hugging Face model cards as listed on 17 September 2026. Gated means you request access on Hugging Face and accept the terms before downloading.

Three points follow:

• Apache-2.0 is not a clinical licence. Prov-GigaPath's card says the model is "not intended or made available for clinical use as a medical device, clinical support, diagnostic tool". H-optimus-0's card says it "has not been reviewed, certified, or approved by any regulatory body".

• Virchow2's terms are the most explicit. Users agree not to use it to "Diagnose, cure, mitigate, treat, or prevent disease", including under Research Use Only (RUO) or Investigational Use Only (IUO) labels. Those are regulatory categories for diagnostic products, not academic research.

• Most gated models check who you are. The UNI, UNI2-h, CONCH, Virchow2, H-optimus-1 and H0-mini cards require your Hugging Face email to match your institutional email before access is approved.

Terms change, so read the current card and your institution's rules before downloading. This table is orientation, not legal advice.

Pathology foundation model benchmarks: what they say and who ran them

A model's own paper evaluates it on tasks its authors chose. Neidlinger et al. put the problem plainly: when groups that publish pathology foundation models evaluate them "on tasks of their own choosing, there is a high potential for bias". Start from independent benchmarks, and check who ran each evaluation.

Slide-level tasks: CONCH first, Virchow2 close

Neidlinger et al. benchmarked 19 histopathology foundation models on weakly supervised tasks covering biomarkers, morphology and prognosis. Their data came from 13 patient cohorts with 6,818 patients and 9,528 slides from lung, colorectal, gastric and breast cancers. Downstream models were trained on The Cancer Genome Atlas (TCGA) and tested on external cohorts. These were the public CPTAC collection, which none of the benchmarked models had been trained on, and proprietary cohorts that "were never part of any foundation model training". Predictions were made per patient.

The published abstract's headline: CONCH "yielded the highest overall performance when compared with vision-only foundation models, with Virchow2 as close second". The same sentence carries the caveat that matters for rare targets: CONCH's lead "was less pronounced in low-data scenarios and low-prevalence tasks". In the full results, averaged across 31 tasks, CONCH and Virchow2 shared the highest mean AUROC of 0.71 (area under the receiver operating characteristic curve), so the gap behind the headline ranking is small.

The ranking covers only the models tested. H-optimus-1 and UNI2-h were not among them.

Ensembles and data diversity

According to the published benchmark, an ensemble combining CONCH and Virchow2 predictions outperformed individual models in 55% of tasks. The authors' explanation is that foundation models "trained on distinct cohorts learn complementary features to predict the same label".

The benchmark's findings also "suggest that data diversity outweighs data volume for foundation models", and its authors write that "bigger is not always better". The Virchow2 paper, written by the model's developers, adds a qualification. Diverse data and domain-specific methods "can outperform models that only scale in the number of parameters". Yet "on average, performance benefits from the combination of domain-specific methods, data scale, and model scale". Either way, the parameter and slide counts on a card do not tell you how a model will do on your task.

UNI vs CONCH vs Virchow2: why leaderboards disagree

Other evaluations produce other orders:

• PathBench, a preprint, collected 15,888 slides from 8,549 patients across 10 hospitals, with strict exclusion of any pretraining usage. Across the 19 pathology foundation models it evaluated, it reports Virchow2 and H-optimus-1 as "the most effective models overall".

• The HEST-Benchmark leaderboard in the HEST repository ranks encoders on gene-expression prediction by average Pearson correlation. It is maintained by the Mahmood Lab, which also develops UNI and CONCH. On the HEST leaderboard, checked 17 September 2026, H-optimus-1 leads at 0.4229, Virchow2 scores 0.4034, CONCH v1 scores 0.3696, and an ImageNet ResNet50 baseline scores 0.3252. The leaderboard also lists newer entries this guide does not profile, such as GenBio-PathFM, just behind H-optimus-1, and CONCH v1.5; read their cards before shortlisting.

These studies test different models, tasks and cohorts, so they answer different questions. UNI vs CONCH shows it: CONCH ranked above UNI overall on Neidlinger et al.'s slide-level tasks, while UNI averages above CONCH v1 on the HEST expression leaderboard. The useful output is not a winner but a shortlist of two or three encoders that do well across several evaluations, including at least one independent one, then tested on your own task.

Match the model to the task

Slide classification and biomarkers

Slide-level prediction is usually weakly supervised: one label per slide or patient, such as a biomarker status or a prognostic outcome. That is the setting Neidlinger et al. measured, so their benchmark is the most directly relevant peer-reviewed evidence cited here. PathBench, a preprint, also evaluates diagnosis and prognosis tasks on whole-slide images, and ranks Virchow2 and H-optimus-1 highest.

Two findings matter when labelled data is scarce. CONCH's lead shrank in low-data and low-prevalence settings. And when Neidlinger et al. trained downstream models on 300, 150 and 75 patients, "All models show similar performance declines with reduced training sizes". For a rare biomarker, plan the cohort first and the encoder second.

Gene expression from H&E: bulk and spatial

Gene expression is how active each gene is in a tissue sample, measured with RNA assays such as RNA sequencing (RNA-Seq). The question for this line of research is how much of that molecular signal can be read from tissue morphology in an H&E image.

Bulk expression means one expression profile per sample. Schmauch et al. report that HE2RNA "can be trained to systematically predict RNA-Seq profiles from whole-slide images alone, without expert annotation."

Spatial expression means predicting expression at locations within the tissue, learned from spatial transcriptomics. He et al. (ST-Net) used 30,612 spatially resolved gene expression measurements matched to histology from 23 patients with breast cancer, and identified over 100 genes whose expression could be predicted at a resolution of 100 µm.

HEST-1k, a dataset whose paper was a NeurIPS 2024 Spotlight, collects 1,229 spatial transcriptomic profiles, each linked to a whole-slide image. Its HEST-Benchmark defines nine tasks: predict the top 50 genes with the highest normalised variance from 112 × 112 µm H&E regions, scored by Pearson correlation.

Take the spread across tissues seriously. On the HEST leaderboard, H-optimus-1 reaches 0.6589 on skin cutaneous melanoma (SKCM) but 0.2421 on rectal adenocarcinoma (READ). The HEST authors note that performance "remains low for some tasks (e.g., READ and ccRCC)"; ccRCC is clear-cell renal cell carcinoma. Wang et al. benchmarked eleven spatial-expression methods in Nature Communications. Scored over all genes on two of their spatial transcriptomics datasets, Pearson correlation (PCC) and structural similarity (SSIM) were "low crossing all methods". Scores on selected genes were higher; the evaluation traps below return to this. Predicted expression is a research signal, not a replacement for an assay, and its strength depends on the tissue.

From tiles to slides: MIL and graphs over tissue

Attention-based multiple instance learning

MIL treats a slide as a bag of tiles with one label. Ilse, Tomczak and Welling (ICML 2018) proposed attention-based MIL (ABMIL), a "permutation-invariant aggregation operator" whose attention weights show each tile's contribution. CLAM (Lu et al., 2021) uses attention-based learning and "only requires slide-level labels".

For comparing encoders, the aggregator mattered little in Neidlinger et al.'s benchmark: ABMIL trailed a transformer aggregator by "an average AUROC difference of 0.01, leaving the overall model rankings largely unchanged". A simple ABMIL head is a reasonable, cheap probe for a shortlist.

Graphs over tissue

Permutation-invariant pooling does not, by itself, use where tiles sit relative to each other. Graph methods bring that spatial information back into the model. Patch-GCN (Chen et al., MICCAI 2021), from a paper titled "Whole Slide Images are 2D Point Clouds", is a graph convolutional network that "hierarchically aggregates instance-level histology features to model local- and global-level topological structures". The price is more design choices, such as how the graph is built and pooled, each to be validated with the same splits as everything else.

Slide encoders versus tile embeddings

Some foundation models add a slide encoder; Prov-GigaPath is one. Neidlinger et al. included four slide encoders and found they "showed no advantage over tile encoders in MIL set-ups, except in low-data scenarios". Try a slide encoder when labelled cohorts are small.

Evaluation traps that inflate results

Each trap makes a result look better than it will be on new patients, and each is documented in published work. The same measure-before-trust habit applies to language-model systems; see RAG hallucination guardrails.

Split by patient, not by tile or spot

The failure: tiles, spots or adjacent sections from one patient land in both training and test sets, so the score partly measures memory of that patient's tissue.

HEST-Benchmark uses patient-stratified splits "to avoid train/test patient-level data leakage". That means k-fold cross-validation where k is the number of patients, or k/2 for ccRCC because of its large number of patients. The splits ship as CSV files. Wang et al. kept "adjacent tissue sections from the same patient" in the same fold; their HER2+ dataset has 36 tissue sections from 8 patients, so the independent unit is the patient, not the section.

Site and scanner signatures

The failure: the model learns where a slide was prepared or scanned, and that correlates with the label. In TCGA data from over 3,000 patients with six cancer subtypes, Howard et al. found both parts of that failure. Targets such as survival, gene expression patterns and driver mutations varied substantially between submitting sites, and the site itself could be detected from the images. In their words, "Site detection remains possible despite commonly used color normalization and augmentation methods". Such signatures "can lead to overoptimistic estimates of model performance". Their fix keeps training and validation samples for a model from different sites.

Foundation models do not remove this. Carloni et al. (MedAGI 2025 workshop at MICCAI) tested them on a multi-scanner dataset and found they "still suffer from scanner bias". A July 2026 preprint by Filiot et al. describes tile representations that "remain sensitive to scanner and staining variability".

Artefacts are a related case. Neidlinger et al. note that "models such as UNI, Hibou, Virchow and Kaiko occasionally highlighted pen marks"; there, pen marks did not skew classification because they "were equally distributed across different classes". Check whether artefacts, sites and scanners correlate with your labels.

Pretraining leakage

The failure: test slides, or slides from the same institutions, were in the foundation model's pretraining. Neidlinger et al. note that "concerns about data leakage arise when foundation models are tested on images from the same institutions where they were trained". They tested on proprietary cohorts they describe as never part of any foundation model training, plus CPTAC, which none of the models they benchmarked had been trained on.

Do the same check. The Phikon-v2 card lists TCGA, CPTAC and GTEx among its pretraining sources (Phikon-v2 was not in that benchmark), and the H0-mini card lists TCGA slides. For those models, a test set drawn from those collections cannot be assumed unseen.

Averages over chosen genes

The failure: a correlation averaged over selected genes is compared with one over all genes. On their HER2+ dataset, Wang et al. found most methods scored higher (in correlation or SSIM) on highly variable or spatially variable genes than on all genes. Scores on HEST-Benchmark's top 50 highly variable genes are therefore not comparable with an all-gene average. Report which genes you scored and how you chose them.

How much compute do whole-slide images need?

Xu et al. note that "a standard gigapixel slide may comprise tens of thousands of image tiles", and every tile you keep passes through the encoder. Three quantities set the budget:

• Encoder size, a first-order driver of cost per tile.

• Tile count, set by tissue area and the magnification you tile at.

• Embedding size, which sets storage and the downstream model's input width.

For the encoders above, the sizes below come from their model cards unless noted. "Concatenated" means the class token joined with the mean patch token, an option the H0-mini and Virchow2 cards show.

Storage is easy to estimate before you commit: for float32 embeddings, bytes ≈ tiles × embedding dimension × 4.

On HEST-Benchmark, bigger encoders tend to score higher, with diminishing returns: the HEST paper reports that model size "strongly correlates with average performance, but the gains grow logarithmically with the number of trainable parameters". Distillation is the middle ground. H0-mini's own preprint reports "nearly comparable performance to large FMs at a significantly reduced inference cost", and on the HEST leaderboard it averages 0.3958 against 0.4150 for H-optimus-0, its teacher. When comparing across embedding sizes, note that HEST reduces every embedding to 256 principal components first, because ridge regression "may unfairly penalize models with larger embedding dimensions".

The pattern that keeps budgets sane: extract embeddings once, cache them, and train many light heads on the cache. For scale, the HEST paper's experiments "were run on a single NVIDIA 3090 GPU". Latency is a separate budget: if every new slide must be embedded at inference time, encoder size and tile count set the wait, not only the bill.

Where my thesis fits

My MSc thesis at Isfahan University of Technology, completed in September 2025, asked whether gene expression can be predicted from digital pathology images. Gene-expression assays are slow and costly; whole-slide images are fast and cheap, but gigapixel-scale. The thesis worked under a hard constraint on compute and latency, and its approach combined graph neural networks, vision foundation models, adapted large-language-model techniques, and statistical and probabilistic modelling. A paper based on it is in preparation.

That constraint is why compute is a selection criterion here, not an afterthought. This guide does not say which foundation models the thesis used and reports no thesis results: the comparisons above are published results from other research groups, not my measurements. The medical imaging case study describes the thesis, and the research page lists it.

Checklist for choosing a pathology foundation model

• Licence and access. The card allows your use, and your account meets its access conditions.

• Task type. Slide label, biomarker, prognosis, bulk expression or spatial expression.

• Shortlist. Encoders that do well in more than one evaluation, including an independent one, not only in their own papers.

• Pretraining overlap. Each card's pretraining sources checked against your test cohorts.

• Splits. Patient-level at minimum, site-level where possible, an external cohort if you can get one.

• Confounders. Site, scanner, stain and artefacts checked against the label.

• Metric and gene set. For expression, the metric and the genes scored, both reported.

• Compute. Extraction time measured on your hardware, embeddings cached, a distilled model tried.

• Ensembles last. Models combined only after each has been measured alone.

If you are scoping a computer-vision system that has to fit a hardware ceiling, the services page describes how I work.

Frequently asked questions

Which pathology foundation model is best?

None is best across tasks. In Neidlinger et al.'s independent benchmark of 19 models on slide-level tasks, CONCH ranked first and Virchow2 a close second; PathBench, a preprint, favours Virchow2 and H-optimus-1; and the HEST leaderboard, checked 17 September 2026, puts H-optimus-1 first on gene expression. Shortlist from these, then test on your own task.

Can you predict gene expression from an H&E slide?

Partly, for some genes and tissues. He et al. (ST-Net) found over 100 genes whose expression could be predicted from breast cancer histology, and HE2RNA predicted bulk RNA-Seq profiles from whole-slide images. On HEST-Benchmark, the top encoder's correlation varies widely by cancer type. It is a research signal, not a substitute for a molecular assay in patient care.

Can I use UNI, CONCH or Virchow2 in a commercial product?

Not under their public terms. Their Hugging Face cards list CC-BY-NC-ND 4.0, limit use to non-commercial academic research, prohibit commercial use without prior approval and ask commercial entities to contact the authors. Of the other models covered here, H-optimus-0 and Prov-GigaPath carry Apache-2.0, but Prov-GigaPath's card puts any deployed use out of scope and H-optimus-0's card requires regulatory compliance and independent validation for healthcare use. Terms change, so check the current card; this is not legal advice.

Can these models read a pathology report or diagnose cancer?

No. They are research models that turn image tiles, and in CONCH's case text, into feature vectors. Virchow2's terms forbid using it to diagnose disease, Prov-GigaPath's card says it is not intended for clinical use, and the H-optimus cards state that no regulator has reviewed or approved them. For questions about a pathology report, ask your pathologist or doctor.

Sources

• Neidlinger et al., Benchmarking foundation models as feature extractors for weakly supervised computational pathology, Nature Biomedical Engineering 10, 1113–1123 (2026; online 2025) (arXiv preprint)

• Chen et al., Towards a general-purpose foundation model for computational pathology (UNI), Nature Medicine, 2024

• Lu et al., A visual-language foundation model for computational pathology (CONCH), Nature Medicine, 2024

• Zimmermann et al., Virchow2: Scaling Self-Supervised Mixed Magnification Models in Pathology, arXiv, 2024

• Xu et al., A whole-slide foundation model for digital pathology from real-world data (Prov-GigaPath), Nature, 2024

• Filiot et al., Distilling foundation models for robust and efficient models in digital pathology (H0-mini), arXiv preprint, 2025

• Hugging Face model cards, checked 17 September 2026: UNI, UNI2-h, CONCH, Virchow2, H-optimus-0, H-optimus-1, H0-mini, Prov-GigaPath, Phikon-v2

• Prov-GigaPath repository README, checked 17 September 2026

• Ma et al., PathBench: A comprehensive comparison benchmark for pathology foundation models towards precision oncology, arXiv preprint, 2025

• Jaume et al., HEST-1k: A Dataset for Spatial Transcriptomics and Histology Image Analysis, NeurIPS 2024 Spotlight

• HEST repository and HEST-Benchmark leaderboard, checked 17 September 2026

• He et al., Integrating spatial gene expression and breast tumour morphology via deep learning (ST-Net), Nature Biomedical Engineering, 2020

• Schmauch et al., A deep learning model to predict RNA-Seq expression of tumours from whole slide images (HE2RNA), Nature Communications, 2020

• Wang et al., Benchmarking the translational potential of spatial gene expression prediction from histology, Nature Communications, 2025

• Howard et al., The impact of site-specific digital histology signatures on deep learning model accuracy and bias, Nature Communications, 2021

• Carloni et al., Pathology Foundation Models are Scanner Sensitive: Benchmark and Mitigation with Contrastive ScanGen Loss, MedAGI 2025 workshop at MICCAI

• Filiot et al., Robustifying pathology foundation models via fine-tuning, arXiv preprint, 2026

• Ilse, Tomczak and Welling, Attention-based Deep Multiple Instance Learning, ICML 2018

• Lu et al., Data-efficient and weakly supervised computational pathology on whole-slide images (CLAM), Nature Biomedical Engineering, 2021

• Chen et al., Whole Slide Images are 2D Point Clouds: Context-Aware Survival Prediction using Patch-based Graph Convolutional Networks (Patch-GCN), MICCAI 2021

• Medical imaging — MRI denoising and gene expression from pathology, case study on this site