PCaBiop

PCaBiop is a whole-slide-image benchmark for measuring the robustness of slide-level pathology foundation models to non-biological variation. It pairs a controlled two-centre prostate-biopsy cohort, in which the contributing centre acts as a labelled confounder, with an external out-of-domain cohort for downstream shortcut-learning evaluation.

The benchmark accompanies the paper A distributional robustness margin for pathology foundation models (citation below, forthcoming). Metric implementations and evaluation code are available in the croma package.

Benchmarks

Config	Slides	Label	Centres	Role
pcabiop (default)	1,000	benign / cancer	Radboud UMC, Karolinska	in-domain robustness benchmark, 250 slides per label–centre cell
pcabiop-isup	3,000	ISUP grade group 0–5	Radboud UMC, Karolinska	finer-grained variant, 250 slides per grade–centre cell
par	324	benign / cancer	PAR (Erbil, Iraq)	external out-of-domain cohort, 162 slides per label

pcabiop is a strict subset of pcabiop-isup: the same slides with the label binarised (ISUP 0 versus ISUP ≥ 1), restricted to a cohort balanced across the two label–centre cells. The two configs are benchmark definitions over one slide pool, not disjoint datasets.

All three configs share the same five-column schema. Paths are relative to the repository root.

sample_id, image, mask, label, center

Usage

from datasets import load_dataset
from huggingface_hub import hf_hub_download, snapshot_download

# benchmark tables
pcabiop = load_dataset("waticlems/pcabiop", "pcabiop", split="train")
isup = load_dataset("waticlems/pcabiop", "pcabiop-isup", split="train")
par = load_dataset("waticlems/pcabiop", "par", split="train")

# fetch a single slide or mask by its path column
path = hf_hub_download("waticlems/pcabiop", pcabiop[0]["image"], repo_type="dataset")

# or mirror the full repository (~358 GB)
snapshot_download("waticlems/pcabiop", repo_type="dataset", local_dir="pcabiop")

Slides are pyramidal whole-slide TIFF/SVS files. Read them with a WSI library such as OpenSlide, not with a plain image reader.

Files

images/panda/<slide_id>.tiff       3,000 PANDA slides (~106 GB)
images/par/<sample_id>_leica.svs   324 PAR slides (~251 GB)
masks/panda/<slide_id>.tif         3,000 tissue masks (~0.5 GB)
masks/par/<sample_id>_leica.tif    324 tissue masks (~0.2 GB)
pcabiop.csv                        1,000 rows
pcabiop-isup.csv                   3,000 rows
par.csv                            324 rows

The masks are binary tissue-segmentation masks (tissue versus background), not Gleason annotation masks. Gleason label masks for the PANDA slides are available from the original Kaggle release.

Provenance and modifications

PANDA. The 3,000 slides originate from the public development set of the PANDA challenge (Bulten et al., 2022), contributed by Radboud University Medical Center and Karolinska Institutet, and are redistributed here under the original slide identifiers. Modifications: the tissue masks were losslessly recompressed (uncompressed tiled TIFF to zlib/deflate, pyramid levels, tiling and resolution metadata preserved, every level verified pixel-for-pixel) and renamed from <slide_id>_tissue.tif to <slide_id>.tif. ISUP grades and centre labels are taken from the PANDA release. Richer per-slide metadata (Gleason scores, scanner vendor, pixel spacing) is available from the original release.

PAR. The 324 slides are the out-of-domain cohort used in the paper, drawn from the PAR dataset (Muhammad Ali et al., 2026, BioImage Archive accession S-BIAD2323), Leica scans only. The cohort comprises all 162 benign slides and an ISUP-stratified random subsample of 162 of the 177 cancer slides (fixed seed, selection code in the croma repository). Slides and tissue masks are redistributed unmodified. Binary labels derive from rater 1 of the PAR release, except for two slides (c120a, c170b) that carry the cancer label following rater 3. Full three-rater grading is available from the PAR release.

Licence

The repository is released under CC BY-NC-SA 4.0, inherited from the PANDA development set (Bulten et al., 2022, Radboud University Medical Center and Karolinska Institutet).

Exception: everything under images/par/ and masks/par/ is redistributed from the PAR dataset under its original CC BY 4.0 licence, with attribution to Muhammad Ali et al. (2026).

Citation

If you use this dataset, please cite the accompanying paper (reference to follow), together with the sources of the underlying data:

@article{bulten2022panda,
  title={Artificial intelligence for diagnosis and {G}leason grading of prostate cancer: the {PANDA} challenge},
  author={Bulten, Wouter and Kartasalo, Kimmo and Chen, Po-Hsuan Cameron and others},
  journal={Nature Medicine},
  volume={28},
  pages={154--163},
  year={2022},
  doi={10.1038/s41591-021-01620-2},
}

@article{par2026,
  title={The {PAR} dataset: Prostate biopsy whole slide images from an underrepresented {M}iddle {E}astern population},
  author={{Muhammad Ali}, Peshawa J. and Vincent, Navin and Abdulla, Saman S. and others},
  journal={Scientific Data},
  volume={13},
  pages={1061},
  year={2026},
  doi={10.1038/s41597-026-07798-9},
}