sofieneb/uni2h_features

UNI2-h Patch Features for TCGA and CPTAC

Pre-extracted patch-level embeddings from the UNI2-h pathology foundation model for 11,759 whole-slide images (WSIs): 9,838 from TCGA (32 projects) and 1,921 from CPTAC (9 cohorts).

Features were extracted with TRIDENT. They are meant for weakly supervised slide-level tasks, such as multiple-instance learning (MIL) for classification, survival or biomarker prediction, without having to download or process the raw WSIs.

These features were produced for the study From Patches to Patients: A study of the tile-to-slide performance transferability in Digital Pathology. If you use them, please cite it (see Citation).

Extraction settings

Setting	Value
Pipeline	TRIDENT (run_batch_of_slides.py --task all: tissue segmentation, patching, feature extraction)
Patch encoder	uni_v2 (MahmoodLab/UNI2-h)
Embedding dimension	1536
Patch size	256 x 256 px
Magnification	20X
Patch overlap	none
Unit	one .h5 file per WSI

Slides that failed during processing were skipped (--skip_errors), so a small number of slides from the original cohorts may be missing.

Repository structure

TCGA/
  TCGA-ACC/
    TCGA-XX-XXXX-01Z-00-DX1.<uuid>.h5
    ...
  TCGA-BLCA/
  ...
CPTAC/
  CPTAC_BRCA/
  ...

Each file is named after the source slide, so TCGA files keep the full slide barcode and UUID.

Cohorts

TCGA (9,838 slides)

Project	Slides	Project	Slides	Project	Slides
TCGA-ACC	227	TCGA-BLCA	457	TCGA-BRCA	1,133
TCGA-CESC	279	TCGA-CHOL	38	TCGA-COAD	458
TCGA-DLBC	44	TCGA-ESCA	158	TCGA-GBM	195
TCGA-HNSC	443	TCGA-KICH	77	TCGA-KIRC	519
TCGA-KIRP	298	TCGA-LGG	491	TCGA-LIHC	359
TCGA-LUAD	541	TCGA-LUSC	512	TCGA-MESO	75
TCGA-OV	83	TCGA-PAAD	176	TCGA-PCPG	176
TCGA-PRAD	401	TCGA-READ	166	TCGA-SARC	253
TCGA-SKCM	432	TCGA-STAD	438	TCGA-TGCT	149
TCGA-THCA	504	TCGA-THYM	117	TCGA-UCEC	502
TCGA-UCS	57	TCGA-UVM	80

CPTAC (1,921 slides)

The CPTAC slides were selected to match those available in the Patho-Bench library, so these features can be used directly with its task splits.

Cohort	Slides	Cohort	Slides	Cohort	Slides
CPTAC_BRCA	112	CPTAC_CCRCC	244	CPTAC_COAD	98
CPTAC_GBM	243	CPTAC_HNSCC	259	CPTAC_LSCC	304
CPTAC_LUAD	326	CPTAC_PDA	242	CPTAC_UCEC	93

File format

Each .h5 file contains:

• features: float array of shape (N, 1536), one UNI2-h embedding per tissue patch
• coords: int array of shape (N, 2), the (x, y) top-left coordinate of each patch in level-0 pixel space of the original WSI

Extraction metadata, such as patch size and magnification, is stored as HDF5 attributes.

Download

The dataset is public, so no login is needed. Install the client (hf_transfer is optional but makes downloads faster):

pip install -U huggingface_hub hf_transfer
export HF_HUB_ENABLE_HF_TRANSFER=1   # optional, faster downloads

You can download a single cohort, several cohorts, a whole source, a single slide, or everything. Pick the folder(s) you need with --include (CLI) or allow_patterns (Python).

Command line

# One cohort (e.g. TCGA-BRCA)
hf download sofieneb/uni2h_features --repo-type dataset \
  --include "TCGA/TCGA-BRCA/*" --local-dir ./uni2h_features

# Several cohorts (e.g. NSCLC = LUAD + LUSC)
hf download sofieneb/uni2h_features --repo-type dataset \
  --include "TCGA/TCGA-LUAD/*" "TCGA/TCGA-LUSC/*" --local-dir ./uni2h_features

# One CPTAC cohort
hf download sofieneb/uni2h_features --repo-type dataset \
  --include "CPTAC/CPTAC_GBM/*" --local-dir ./uni2h_features

# All of CPTAC
hf download sofieneb/uni2h_features --repo-type dataset \
  --include "CPTAC/*" --local-dir ./uni2h_features

# Everything
hf download sofieneb/uni2h_features --repo-type dataset \
  --local-dir ./uni2h_features

With older versions of huggingface_hub, replace hf download with huggingface-cli download.

Python

from huggingface_hub import snapshot_download, hf_hub_download

REPO = "sofieneb/uni2h_features"

# One cohort
snapshot_download(REPO, repo_type="dataset",
                  allow_patterns=["TCGA/TCGA-BRCA/*"],
                  local_dir="./uni2h_features")

# Several cohorts, e.g. RCC subtypes (KIRC + KIRP + KICH)
snapshot_download(REPO, repo_type="dataset",
                  allow_patterns=["TCGA/TCGA-KIRC/*", "TCGA/TCGA-KIRP/*", "TCGA/TCGA-KICH/*"],
                  local_dir="./uni2h_features")

# A single slide
hf_hub_download(REPO, repo_type="dataset",
                filename="TCGA/TCGA-BRCA/<slide_id>.h5",
                local_dir="./uni2h_features")

To see which cohorts and files are available before downloading:

from huggingface_hub import HfApi
from collections import Counter

files = HfApi().list_repo_files("sofieneb/uni2h_features", repo_type="dataset")
print(Counter("/".join(f.split("/")[:2]) for f in files if f.endswith(".h5")))

Downloads resume automatically if interrupted. Just rerun the same command.

Usage

Load a slide's features:

import h5py
import torch

with h5py.File("uni2h_features/TCGA/TCGA-BRCA/<slide_id>.h5", "r") as f:
    feats = torch.from_numpy(f["features"][:])   # (N, 1536)
    coords = f["coords"][:]                      # (N, 2)

Build a slide table for one cohort:

from pathlib import Path
import pandas as pd

root = Path("uni2h_features/TCGA/TCGA-BRCA")
df = pd.DataFrame({"path": sorted(root.glob("*.h5"))})
df["slide_id"] = df["path"].map(lambda p: p.stem)
df["case_id"] = df["slide_id"].str.slice(0, 12)   # TCGA-XX-XXXX, for patient-level splits

Important notes

• Features are per slide, not per patient. Some patients have several slides. Split train and test sets by patient (the TCGA case ID) to avoid data leakage.
• No labels are included. Clinical, molecular and survival labels must be taken from the GDC (TCGA) and PDC/TCIA (CPTAC), or from Patho-Bench for the CPTAC tasks.
• Features depend on the preprocessing. Tissue segmentation, patch size and magnification all affect the embeddings. Features from a different pipeline or encoder are not interchangeable with these.

Source data and licensing

• TCGA slides: NCI Genomic Data Commons
• CPTAC slides: The Cancer Imaging Archive (TCIA)
• UNI2-h weights are released by the Mahmood Lab under CC BY-NC-ND 4.0, for non-commercial academic research only. These features are derived from that model, and this dataset uses the same license. Use of the features must also follow the TCGA and CPTAC/TCIA data use policies.

Citation

If you use these features, please cite:

@article{boutaj2026patches,
  title   = {From Patches to Patients: A study of the tile-to-slide performance transferability in Digital Pathology},
  author  = {Boutaj, Sofi{\`e}ne and Fillioux, Leo and Christodoulidis, Stergios and Marza, Pierre and others},
  journal = {arXiv preprint arXiv:2606.10778},
  year    = {2026}
}

Please also cite the encoder, the extraction pipeline and the source datasets:

• UNI / UNI2-h: Chen, R. J. et al. Towards a general-purpose foundation model for computational pathology. Nature Medicine (2024).
• TRIDENT: Zhang, A. et al. Accelerating Data Processing and Benchmarking of AI Models for Pathology. arXiv:2502.06750 (2025).
• TCGA: The Cancer Genome Atlas Research Network, https://www.cancer.gov/tcga
• CPTAC: National Cancer Institute Clinical Proteomic Tumor Analysis Consortium, https://proteomics.cancer.gov/programs/cptac