srikarym/PixCell-training-data

PixCell — Training Data

The 1024×1024 H&E patches used to train PixCell, a generative foundation model for digital histopathology, from public sources and an in-house Stony Brook University cohort.

29,755,727 patches from 66,354 whole-slide images (~6.1 TB), extracted at 20× magnification. Images are the original JPEG files (quality 70) stored byte-for-byte, not re-encoded.

config	source	slides	patches
tcga_diagnostic	TCGA FFPE diagnostic slides, 32 cancer types	11,713	9,180,591
tcga_fresh_frozen	TCGA fresh-frozen tissue slides, 32 cancer types	18,192	3,680,187
cptac	CPTAC (TCIA), 11 cancer types	5,802	2,196,291
gtex	GTEx normal tissue, 40 tissue types	25,266	7,870,674
ptrc_hgsoc	PTRC-HGSOC (TCIA), ovarian cancer	346	144,134
sbu_olympus	Stony Brook University (in-house), 46 organ/tissue types	5,035	6,683,850

The PixCell paper's full training set also includes NLST slides, which are not part of this release. The SBU slides are not publicly available, so sbu_olympus rows have no WSI coordinates or scanner metadata (those fields are null).

Usage

from datasets import load_dataset

# stream everything, or pick one config: "tcga_diagnostic", "gtex", ...
ds = load_dataset("srikarym/PixCell-training-data", "tcga_diagnostic", split="train", streaming=True)
ex = next(iter(ds))
ex["image"]  # PIL.Image, 1024x1024 RGB

To download a single source without datasets:

hf download srikarym/PixCell-training-data --repo-type dataset --include "data/cptac/*" --local-dir PixCell-training-data

Rows are grouped by subtype and slide. Shuffle (ds.shuffle(buffer_size=...)) before training.

Fields

field	type	description
image	Image	JPEG patch; image.path is the original relative path <src>/<subtype>/single_1024/<slide_id>/<col>_<row>.jpeg
source	string	config name
subtype	string	cancer type (TCGA/CPTAC) or organ/tissue type (GTEx/SBU)
slide_id	string	slide identifier (WSI filename up to the first .)
col, row	int32	DeepZoom tile address at 20×
x, y, w, h	int32	patch region in level-0 pixels of the WSI; the patch is this region resized to 1024×1024
objective_power	float32	scanner objective power from the WSI metadata (null if absent; such slides were tiled as 20×)
mpp_x	float32	microns per pixel at level 0, from the WSI metadata
in_test_set	bool	patch belongs to the subset used for FID evaluation in the paper (it is also part of the training data)

metadata/slides.parquet has one row per slide: source, subtype, slide_id, wsi_file, num_patches, objective_power, mpp_x, width, height. Use wsi_file to locate the original slide in its source archive (null for SBU).

Extraction

Patches were tiled with the DeepZoom tiler from DSMIL (deepzoom_tiler.py --tile_size 1024: 20× base magnification, no overlap, background edge-threshold 15). To re-crop a patch from the original WSI:

import openslide
slide = openslide.OpenSlide(wsi_path)
patch = slide.read_region((x, y), 0, (w, h)).convert("RGB").resize((1024, 1024))

Using with the PixCell training code

The PixCell code expects JPEG folders plus a patch_names_all.hdf5 index. scripts/to_pixcell_layout.py rebuilds that layout from the downloaded shards:

python scripts/to_pixcell_layout.py --hf_dir PixCell-training-data --out /path/to/pixcell_data

The VAE latents and UNI-2h embeddings used for training are not included. They can be re-extracted from the images.

License and attribution

This dataset is released under CC-BY-4.0. The images are derived from public data, and use is also subject to each source's terms. Please cite and acknowledge the original sources:

• TCGA: The results here are in whole or part based upon data generated by the TCGA Research Network: https://www.cancer.gov/tcga.
• CPTAC: National Cancer Institute Clinical Proteomic Tumor Analysis Consortium (CPTAC); cite the individual CPTAC pathology collections on TCIA.
• GTEx: The Genotype-Tissue Expression (GTEx) Project was supported by the Common Fund of the Office of the Director of the National Institutes of Health, and by NCI, NHGRI, NHLBI, NIDA, NIMH, and NINDS.
• PTRC-HGSOC: cite the collection as listed on its TCIA page (CC BY 4.0).
• TCIA: Clark K, et al. The Cancer Imaging Archive (TCIA): Maintaining and Operating a Public Information Repository. J Digit Imaging. 2013;26(6):1045–1057.

Citation

@article{yellapragada2025pixcell,
  title={PixCell: A generative foundation model for digital histopathology images},
  author={Yellapragada, Srikar and Graikos, Alexandros and Li, Zilinghan and Triaridis, Kostas and Belagali, Varun and Kapse, Saarthak and Nandi, Tarak Nath and Madduri, Ravi K and Prasanna, Prateek and Kurc, Tahsin and others},
  journal={arXiv preprint arXiv:2506.05127},
  year={2025}
}