zhihuanglab/VISTA-PATH_data

VISTA-PATH training data (release subset)

Processed image/mask pairs used to train VISTA-PATH, an interactive foundation model for histopathology image segmentation.

• 📎 Paper: https://www.arxiv.org/abs/2601.16451
• 💻 Code: https://github.com/zhihuanglab/VISTA-PATH
• 🤗 Model: https://huggingface.co/zhihuanglab/VISTA-PATH

This repository holds the train split of 9 of the datasets used in the paper, already tiled into patches and paired with label masks. It is a subset of the full training corpus used in the paper.

Contents

Dataset	Image–mask pairs	Shards (images + masks)	Size
AGGC	90,978	68 + 1	135.9 GiB
BRCA	7,959	8 + 1	14.9 GiB
CoCaHis	302	1 + 1	0.4 GiB
DRYAD	3,343	3 + 1	5.3 GiB
DigestPath	4,949	3 + 1	5.9 GiB
Histo-Seg	1,869	1 + 1	2.0 GiB
TIGER_wsibulk	47,669	35 + 1	69.1 GiB
TIGER_wsirois	1,225	2 + 1	2.5 GiB
WSSS4LUAD	540	1 + 1	1.1 GiB

158,834 image–mask pairs, 131 shards, 237.2 GiB in total. Every image has exactly one mask.

Layout

The data is stored as plain tar shards of roughly 2 GB rather than as loose files — half a million individual PNGs run into the Hub's request limits — so the repository looks like this:

<dataset>/images_train_000.tar
<dataset>/images_train_001.tar
...
<dataset>/masks_train_000.tar
manifest.json

manifest.json lists every shard with its file count and byte size.

Usage

Download one dataset, or drop allow_patterns to fetch everything:

from huggingface_hub import snapshot_download

snapshot_download("zhihuanglab/VISTA-PATH_data", repo_type="dataset",
                  local_dir="./data", allow_patterns="WSSS4LUAD/*")

Each shard carries its own <dataset>/ prefix inside the archive, so extract from the directory the shards were downloaded into:

cd data
for f in WSSS4LUAD/*.tar; do tar -xf "$f"; done

That restores the original tree:

data/WSSS4LUAD/images/train/000-0-0.png
data/WSSS4LUAD/masks/train/000-0-0.png

Format

• Images — RGB PNG patches extracted from the source whole-slide images or ROIs.
• Masks — single-channel uint8 PNG label maps, the same height and width as the image they belong to and carrying the same filename. Pixel values are class indices, not colours; 0 marks background or unannotated pixels.

Licensing

Each constituent dataset remains under the terms of its original provider, and those terms differ between sources. Consult the licence of the specific source dataset before using or redistributing the corresponding subset. The VISTA-PATH model and code are licensed separately, under the Penn Academic Software License.

Citation

@misc{liang2026vistapath,
  title        = {VISTA-PATH: An interactive foundation model for pathology image segmentation and quantitative analysis in computational pathology},
  author       = {Liang, Peixian and Li, Songhao and Koga, Shunsuke and Li, Yutong and Alipour, Zahra and Tang, Yucheng and Xu, Daguang and Huang, Zhi},
  year         = {2026},
  eprint       = {2601.16451},
  archivePrefix= {arXiv},
  primaryClass = {cs.CV},
  doi          = {10.48550/arXiv.2601.16451},
}