AI models for Digital pathology analysis

UNI 2 / UNI2-h

Vendor: Mahmood Lab, Harvard Medical School

What it does: a tile encoder built on a very large slide collection, and the strongest general starting point for classification and retrieval across tissue types.

Virchow 2 / Virchow2G

Vendor: Paige

What it does: a tile encoder trained on a very large clinical archive, notable for holding accuracy across scanners and staining variation.

Prov-GigaPath

Vendor: Microsoft Research and Providence

What it does: works at whole-slide scale rather than tile scale, which suits slide-level predictions where the relevant signal is spread across the tissue.

CONCH

Vendor: Mahmood Lab, Harvard Medical School

What it does: pairs slide regions with text, so an archive can be searched by description instead of by a fixed label set.

Phikon-v2

Vendor: Owkin

What it does: a lighter tile encoder that runs comfortably on one mid-range card, for a first pipeline or a cost-sensitive archive pass.

CellViT++

Vendor: University Hospital Essen

What it does: detects and classifies individual nuclei rather than regions, and adds cell-level classification on top of segmentation, so a slide yields counts and densities rather than a heat map.

H-optimus-0

Vendor: Bioptimus, with Aignostics, Mayo and Charité collaborators

What it does: is a 1.1B tile encoder that led pan-cancer performance in one independent benchmark, and is Apache-licensed for the weights — an unusual combination in this category.

Best pan-cancer performance in one independent benchmark; AUROC 0.68 mean across 31 tasks in another.

H-optimus-1

Vendor: Bioptimus

What it does: is the newer version of the same encoder, for classification, grading, retrieval and prognosis work.

Task-dependent across classification, grading, retrieval, prognosis and slide-level benchmarks.

UNI

Vendor: Harvard Medical School (Mahmood lab)

What it does: is the widely cited Harvard tile encoder, strong on brain, bladder, breast and pan-cancer tasks. Gated and non-commercial, which is the constraint to check before it becomes your base layer.

Strong on brain, bladder, breast and pan-cancer tasks; AUROC 0.68 mean across 31 independent tasks. CC-BY-NC-ND, gated.

Virchow

Vendor: Paige AI, with Microsoft

What it does: is the first-generation Paige encoder, trained on a very large clinical archive and joint-best overall in independent benchmarking.

Joint-best overall in independent benchmarking, AUROC 0.71 across 31 tasks; leads on colon and prostate.

Hibou-B

Vendor: HistAI

What it does: is the Apache-licensed member of the modern encoder set — nearly the accuracy of the gated models with none of the licence friction, which often decides the choice.

AUROC 0.67 mean across 31 independent tasks; Apache-licensed, which is rare in this category.

Hibou-L

Vendor: HistAI

What it does: is the larger Hibou encoder, at similar benchmark accuracy and a non-commercial licence.

AUROC 0.67 mean across 31 independent tasks (CC-BY-NC for the L variant).

Phikon

Vendor: Owkin

What it does: is the small, fast Owkin encoder: a few points behind the leaders and cheap enough to run across a whole archive.

AUROC 0.65 mean across 31 independent tasks; small and fast.

EXAONEPath

Vendor: LG AI Research

What it does: is among the top encoders for brain tissue tasks specifically, which is the case for it over a general leader.

Among top models for brain tissue tasks (independent). Non-commercial.

Path Foundation

Vendor: Google Research

What it does: produces patch embeddings that linear probes turn into tumour detection or grading with small label budgets. Superseded by MedSigLIP, and listed here for continuity.

Efficient linear probes for tumour detection and grading with small label budgets (developer-reported). Legacy model — MedSigLIP is the current recommendation.

REMEDIS-Pathology

Vendor: Google Research

What it does: is a Google pathology representation model for classification, grading and retrieval from patch embeddings.

Task-dependent across classification, grading, retrieval and prognosis benchmarks.

Kaiko

Vendor: Kaiko

What it does: is an openly published tile encoder for pathology classification and retrieval work.

Task-dependent across pathology benchmarks.

Lunit-BT

Vendor: Lunit

What it does: is one of four Lunit encoders trained with different self-supervised objectives, which makes the set useful for ensembling rather than picking one.

Task-dependent across pathology benchmarks.

Lunit-DINO

Vendor: Lunit

What it does: is the DINO-trained member of the Lunit set, for patch embeddings and downstream prediction.

Task-dependent across pathology benchmarks.

Lunit-MoCoV2

Vendor: Lunit

What it does: is the MoCo v2 member of the Lunit set.

Task-dependent across pathology benchmarks.

Lunit-SwAV

Vendor: Lunit

What it does: is the SwAV member of the Lunit set.

Task-dependent across pathology benchmarks.

CTransPath

Vendor: Sichuan University / Tencent AI Lab

What it does: is a long-established pathology encoder, still a solid retrieval baseline and light to run.

Task-dependent across pathology classification and retrieval benchmarks.

RetCCL

Vendor: Sichuan University / Tencent AI Lab

What it does: is built for retrieval specifically — finding visually comparable regions across an archive rather than labelling them.

Task-dependent across pathology retrieval benchmarks.

PathoDuet

Vendor: Shanghai Jiao Tong University

What it does: is a pathology encoder for patch-level classification and downstream prediction.

Task-dependent across pathology benchmarks.

BEPH

Vendor: Shanghai Jiao Tong University

What it does: produces patch embeddings for classification, grading and prognosis work.

Task-dependent across pathology benchmarks.

PathOrchestra

Vendor: Shanghai AI Lab

What it does: covers a wide set of pathology tasks from one model, which reduces the number of encoders a pipeline has to maintain.

Task-dependent across pathology benchmarks.

GPFM

Vendor: Smart Lab / HKUST collaborators

What it does: is a general pathology foundation model covering patch and slide-level prediction.

Task-dependent across pathology benchmarks.

HIPT

Vendor: Mahmood Lab

What it does: works up a hierarchy from patch to region to slide, which is how it reaches slide-level prediction without a separate aggregator.

Task-dependent across classification, grading and prognosis benchmarks.

CONCH1.5 / CONCHv1.5

Vendor: Harvard Medical School (Mahmood lab)

What it does: pairs slide regions with text, so an archive can be labelled and searched in words with no training. Joint-best overall in independent benchmarking and best on lung and prognostic tasks — gated and non-commercial.

Task-dependent; CONCH is joint-best overall at AUROC 0.71 across 31 tasks and best on lung and prognostic tasks (independent). CC-BY-NC-ND, gated.

TITAN

Vendor: Harvard Medical School (Mahmood lab)

What it does: predicts at whole-slide level and beats multiple-instance-learning baselines on rare disease and few-shot classification, which is where slide-level labels are scarcest.

Outperforms MIL baselines for rare-disease and few-shot slide classification (developer-reported). CC-BY-NC-ND.

MUSK

Vendor: Stanford University (Li Lab)

What it does: combines pathology images with clinical text for prognosis and immunotherapy response prediction — the furthest into outcome prediction any model here goes.

Leading multimodal pathology results including prognosis and immunotherapy response prediction (independent, Nature 2025). CC-BY-NC.

PLIP

Vendor: Stanford University

What it does: is the smallest vision-language option: weakest of the modern set on accuracy, and easy enough to run that it suits a first pipeline or a cheap archive pass.

AUROC 0.64 mean across 31 independent tasks — weakest of the modern set but very small and easy to run. Non-commercial research use.

MI-Zero

Vendor: MI-Zero authors

What it does: labels whole slides zero-shot from text prompts, with no slide-level training set required.

Task-dependent across zero-shot slide classification benchmarks.

QuiltNet

Vendor: QuiltNet authors

What it does: pairs pathology images with text for zero-shot labelling and retrieval.

Task-dependent across pathology classification and retrieval benchmarks.

PathAsst

Vendor: PathAsst authors

What it does: answers questions about a pathology image and generates descriptive text, for review and teaching rather than reporting.

Task-dependent across pathology VQA and classification benchmarks.

CHIEF

Vendor: Yu Lab / Harvard Medical School

What it does: produces slide-level embeddings, disease classes and prognosis scores from tiles or whole slides.

Task-dependent across classification, grading, retrieval and prognosis benchmarks.

PRISM

Vendor: Paige AI, with Microsoft

What it does: aggregates tile features into a slide embedding and generates report text from it.

Task-dependent across slide-level benchmarks.

COBRA

Vendor: Kather Lab

What it does: produces whole-slide embeddings for downstream slide-level prediction.

Task-dependent across slide-level benchmarks.

MADELEINE

Vendor: Mahmood Lab

What it does: learns across several stains of the same tissue, which is what makes it useful where your lab routinely runs more than H&E.

Task-dependent across slide-level benchmarks.

GigaSSL

Vendor: CBIO / Mines Paris

What it does: produces whole-slide embeddings cheaply, trained self-supervised at slide rather than tile level.

Task-dependent across slide-level benchmarks.

CellViT

Vendor: University Hospital Essen (IKIM)

What it does: detects and classifies individual nuclei at the strongest published accuracy for the task — PanNuke panoptic quality about 0.50 and detection F1 about 0.82 — which is what cell counting and density measurement actually need.

PanNuke panoptic quality about 0.50, mean detection F1 about 0.82 — state of the art for nuclei instance segmentation (independent).

HoVer-Net

Vendor: University of Warwick (TIA Centre)

What it does: is the long-standing reference implementation for nucleus instance segmentation and classification, and still a dependable baseline.

PanNuke and CoNSeP detection F1 about 0.80; long-standing reference implementation (independent).

StarDist / Cellpose 3

Vendor: MPI-CBG and community / MouseLand

What it does: segments cells across varied microscopy without tuning per dataset — the generalist option when your inputs are not all H&E.

Generalist cell segmentation, average precision above 0.80 on varied microscopy (independent).

QuPath + StarDist / Cellpose extensions

Vendor: University of Edinburgh / QuPath community

What it does: is the platform rather than the model: it reads whole-slide formats, manages annotation and hosts the segmentation models above.

Pipeline platform; accuracy depends on the embedded model.

CLAM / TRIDENT

Vendor: Harvard Medical School (Mahmood lab)

What it does: trains a slide-level classifier from slide-level labels alone, with attention heatmaps showing which regions drove the prediction — no region annotation project required.

Weakly-supervised MIL framework; typical slide-level AUROC 0.85–0.98 depending on task and encoder.