These embeddings were computed from public pathology datasets, some of which (BRACS, SPIDER-Breast) are licensed CC BY-NC 4.0. They are provided for non-commercial research only. They are not a medical device and must not be used for diagnosis or patient care. By requesting access you agree to use them for non-commercial research, to cite the original datasets and models, and to comply with the upstream terms listed in RELEASE_POLICY.md.

PRISM embeddings

Frozen tile-level embeddings from pathology foundation models (FMs), released with PRISM: Pathology Reliability In Scarce-label Medicine. A Tile-level Benchmark of Pathology Foundation Models (Alicem Koyun, Westcliff University; NeurIPS 2026 GenAI4Health workshop). They let anyone rerun PRISM's linear-probe protocol (label fractions 1–100%, three seeds, AUROC, raw and temperature-scaled ECE, Brier score, Composite Reliability Index (CRI), and the cross-dataset and Camelyon17 hospital-transfer grids) without a GPU.

Code: the prism_bench package and extraction notebooks are in the PRISM code repository (link in the paper). Upstream licences decide which (model, dataset) pairs are included here.

What is released

Embeddings are released for 5 of the 6 main datasets and for Camelyon17, for the 5 of 8 FMs whose licences allow redistribution of model outputs:

	PCam	CRC	MHIST	BRACS	LungHist700	SPIDER-Breast	Camelyon17
CLIP	✓	✓	checksums	✓	✓	✓	✓
PLIP	✓	✓	checksums	✓	✓	✓	✓
GigaPath	✓	✓	checksums	✓	✓	✓	✓
H-Optimus-0	✓	✓	checksums	✓	✓	✓	✓
MIDNIGHT	✓	✓	checksums	✓	✓	✓	✓
UNI, CONCH, VIRCHOW2	checksums	checksums	checksums	checksums	checksums	checksums	checksums

Withheld cells:

MHIST (all 8 models): the Dataset Research Use Agreement forbids distributing any portion of MHIST and creating derivative works from it. Only file-level SHA256 values and summary fingerprints are published. No embeddings, labels or image names are published.
UNI, CONCH, VIRCHOW2: the CC BY-NC-ND 4.0 terms of use define "datasets created from the model" as derivatives.

For these cells, checksums.csv gives the SHA256 and numerical fingerprints of the files used in the paper. code/ gives the extraction recipe, and splits/ gives the split indices. With these, users who hold the upstream licences can regenerate the files and check them. RELEASE_POLICY.md has the full licence review and sources.

Rows per split and embedding dimension of the released .npy files, taken from embedding_inventory.csv (Camelyon17: summed over the 5 hospitals, stored split before re-splitting).

dataset	model	train	val	test	dim
bracs	clip	3,657	312	570	768
bracs	gigapath	3,657	312	570	1536
bracs	h_optimus_0	3,657	312	570	1536
bracs	midnight	3,657	312	570	1536
bracs	plip	3,657	312	570	768
camelyon17	clip	61,261	7,344	31,395	768
camelyon17	gigapath	61,261	7,344	31,395	1536
camelyon17	h_optimus_0	61,261	7,344	31,395	1536
camelyon17	midnight	61,261	7,344	31,395	1536
camelyon17	plip	61,261	7,344	31,395	768
crc	clip	70,000	15,000	15,000	768
crc	gigapath	70,000	15,000	15,000	1536
crc	h_optimus_0	70,000	15,000	15,000	1536
crc	midnight	70,000	15,000	15,000	1536
crc	plip	70,000	15,000	15,000	768
lunghist700	clip	483	103	105	768
lunghist700	gigapath	483	103	105	1536
lunghist700	h_optimus_0	483	103	105	1536
lunghist700	midnight	483	103	105	1536
lunghist700	plip	483	103	105	768
pcam	clip	262,144	32,768	32,768	768
pcam	gigapath	262,144	32,768	32,768	1536
pcam	h_optimus_0	262,144	32,768	32,768	1536
pcam	midnight	262,144	32,768	32,768	1536
pcam	plip	262,144	32,768	32,768	768
spider_breast	clip	68,730	12,128	12,034	768
spider_breast	gigapath	68,730	12,128	12,034	1536
spider_breast	h_optimus_0	68,730	12,128	12,034	1536
spider_breast	midnight	68,730	12,128	12,034	1536
spider_breast	plip	68,730	12,128	12,034	768

Repository layout

README.md                  this card
RELEASE_POLICY.md          licence review: classification and sources
checksums.csv              every inventory file (released and withheld): sha256, shape, dtype, fingerprints
embeddings/{model}/{dataset}/{train,val,test}_{features,labels}.npy             canonical files (main datasets)
embeddings_camelyon17/{model}/hospital_{0..4}/{train,val,test}_{features,labels}.npy
data/{dataset}/{model}/{train,validation,test}.parquet                          the same values, for `datasets.load_dataset`
splits/{dataset}.json                                  split indices for the main datasets (see code/make_split_indices.py)
splits/camelyon17/{splits.json,resplit.json,...}       Camelyon17 sampling and balanced re-split
code/notebook_map.csv, code/extract_embeddings.py, code/make_split_indices.py

Model keys: clip, plip, gigapath, h_optimus_0, midnight (withheld: conch, uni, virchow2). Dataset keys: pcam, crc, bracs, lunghist700, spider_breast (withheld: mhist), and camelyon17.

Schema

.npy (canonical; the SHA256 values in checksums.csv refer to these files). {split}_features.npy is a float32 array of shape (n, d); row i is the embedding of tile i. {split}_labels.npy is an integer array of shape (n,). Split names are train, val and test.

Parquet configs (one config per dataset; every released model sits in the same table):

column	type	meaning
`model`	string	model key
`dataset`	string	dataset key
`row`	int64	row index into the canonical `{split}_features.npy` (main datasets) or into concat(train, val, test) of the stored hospital files (Camelyon17)
`label`	int64	class id (conventions below)
`embedding`	list	the feature vector; its length depends on the model
`hospital`	int64	Camelyon17 only, 0–4 (WILDS numbering)
`orig_split`	string	Camelyon17 only: which stored file the row comes from

from datasets import load_dataset
# every released model for one dataset (large for PCam):
ds = load_dataset("alicemkyn/PRISM-embeddings", "bracs")
# one model only:
ds = load_dataset("alicemkyn/PRISM-embeddings", data_files={
    "train": "data/bracs/h_optimus_0/train.parquet"})          # path pattern: data/{dataset}/{model}/{split}.parquet
# or the canonical numpy files:
from huggingface_hub import hf_hub_download
import numpy as np
X = np.load(hf_hub_download("alicemkyn/PRISM-embeddings", "embeddings/gigapath/bracs/train_features.npy", repo_type="dataset"))

Normalisation. Measured on the stored files (mean_l2_norm in checksums.csv): every file under embeddings/ (the six main datasets, all eight models) is L2-normalised (unit norm). Under embeddings_camelyon17/ the features are the raw model outputs (mean norm 9–53 depending on the model); CONCH is the exception with unit norm. PRISM L2-normalises at load time (X / ‖X‖₂), which is a no-op for already-normalised files, so apply the same step to reproduce the paper's numbers.

Splits and label conventions

Test sets are fixed per dataset and identical across models. Every extraction used DataLoader(shuffle=False), so row order is the dataset-construction order below. splits/{dataset}.json records the indices and, where the licence allows, the file names.

dataset	split rule	labels
PCam	official train / valid / test H5 files (torchvision `PCAM`)	0 = normal, 1 = tumour
CRC	`ImageFolder(NCT-CRC-HE-100K)`; `random_split` 70/15/15, `torch.Generator().manual_seed(42)`	0–8 in alphabetical folder order: ADI, BACK, DEB, LYM, MUC, MUS, NORM, STR, TUM
BRACS	official RoI train / val / test folders (`latest_version`)	0–6 in alphabetical folder order (N, PB, UDH, FEA, ADH, DCIS, IC); the exact folder names are in `splits/bracs.json`
LungHist700	`ImageFolder(images/)` with 7 class folders; `random_split` 70/15/15, seed 42	0–6 in alphabetical folder order; names in `splits/lunghist700.json`
SPIDER-Breast	224×224 patches at 20× (`metadata.json`); official train / test; val = first 15% of train after `random.seed(42); random.shuffle`	0–17 = the 18 classes sorted alphabetically; names in `splits/spider_breast.json`
MHIST (withheld)	`annotations.csv` Partition; val = train rows `.sample(frac=0.15, random_state=42)`	0 = HP, 1 = SSA
Camelyon17	20,000 patches per hospital, label-stratified 50/50 (`splits/camelyon17/splits.json`, indices into the concatenated `wltjr1007/Camelyon17-WILDS` parquet, verified at 455,954 patches against WILDS v1.0). The stored files follow an initial slide-disjoint partition. The paper's protocol applies the balanced patch-level 70/15/15 re-split in `resplit.json`: concatenate the stored train, val and test files of a hospital, then index	0 = non-tumour, 1 = tumour

The parquet camelyon17 config is already in the paper's re-split. The .npy files are as stored. Because the re-split is at patch level, train and test within a hospital share slides, so in-hospital numbers are an upper bound. Transfer pairs (different hospitals) have no leakage.

Preprocessing

Each FM was embedded once under one pipeline, held fixed across all label fractions and transfer pairs. Preprocessing differs between models, so between-model comparisons at a fixed fraction are confounded by it.

Main datasets (extraction notebooks notebooks/<dataset>/<model>_<dataset>.ipynb, notebooks/Pcam/0X_*.ipynb):

model	weights	resize	crop	normalisation (mean / std)	feature
CLIP	`openai/clip-vit-base-patch32`	224	224	ImageNet	vision `pooler_output`
PLIP	`vinid/plip`	224	224	ImageNet	vision `pooler_output`
CONCH (withheld)	`MahmoodLab/conch`	loader	loader	loader transform	`encode_image(proj_contrast=False, normalize=True)`
UNI (withheld)	`MahmoodLab/UNI`	224	224	ImageNet	CLS
VIRCHOW2 (withheld)	`paige-ai/Virchow2`	224	224	ImageNet	concat(CLS, mean of patch tokens 5:)
GigaPath	`prov-gigapath/prov-gigapath`	256	224	ImageNet	CLS
H-Optimus-0	`bioptimus/H-optimus-0`	224	224	(0.707223, 0.578729, 0.703617) / (0.211883, 0.230117, 0.177517)	CLS
MIDNIGHT	`kaiko-ai/midnight`	224	224	ImageNet (the model card recommends 0.5 / 0.5)	`pooler_output`

ImageNet = (0.485, 0.456, 0.406) / (0.229, 0.224, 0.225). PCam tiles are 96×96 and are resized to 224×224. The CLIP and PLIP PCam notebooks use Resize((224, 224)) without a crop, which gives the same result for square tiles.

Camelyon17 (notebooks/PRISM_camelyon17_extract.ipynb; transforms resolved from each model; splits/camelyon17/preprocessing_table.csv):

model	resize	crop	normalisation	feature	dim
CLIP, PLIP	224	224	CLIP statistics (0.4815, 0.4578, 0.4082) / (0.2686, 0.2613, 0.2758)	vision `pooler_output`	768
CONCH (withheld)	loader	loader	loader	`encode_image(proj_contrast=False, normalize=False)`	512
UNI (withheld)	224	224	ImageNet	CLS	1024
VIRCHOW2 (withheld)	224	224	ImageNet	concat(CLS, mean patch)	2560
GigaPath	224	224	ImageNet	CLS	1536
H-Optimus-0	256	224	model-specific	CLS	1536
MIDNIGHT	224	224	(0.5, 0.5, 0.5) / (0.5, 0.5, 0.5)	`pooler_output`	1536

The CLIP, PLIP and MIDNIGHT Camelyon17 embeddings were re-extracted (notebook cell 9). The repository's first-pass table listed 512 / 512 / 3072 dimensions for these, which no longer applies. The authoritative dimensions are the shapes in checksums.csv.

Verifying and regenerating withheld embeddings

python code/extract_embeddings.py map --model uni --dataset bracs   # which notebook, weights, preprocessing
# run that notebook with your own access to the model and dataset, then
python code/extract_embeddings.py verify --checksums checksums.csv --root /path/to/PRISM --model uni --dataset bracs

Expect EXACT for downloaded copies. For re-extractions expect CONSISTENT: GPU kernels, fp16 autocast and library versions change the low-order bits, so the SHA256 will differ. Labels and shapes must still match exactly, and the fingerprints must agree within tolerance.

Intended use and limitations

Research on the reliability and calibration of frozen pathology representations, label-efficiency studies, and probe design.
Not for clinical use. None of these models or embeddings has regulatory approval.
Tile-level only. The datasets come from a limited number of institutions and scanners.
Commercial use is excluded by CC BY-NC 4.0 and by the BRACS and SPIDER-Breast terms.

Citation

@inproceedings{koyun2026prism,
  title     = {{PRISM}: Pathology Reliability In Scarce-label Medicine. A Tile-level Benchmark of Pathology Foundation Models},
  author    = {Koyun, Alicem},
  booktitle = {NeurIPS 2026 Workshop on Generative AI for Health (GenAI4Health)},
  year      = {2026}
}

Please also cite the source datasets and models. Datasets: PCam (Veeling et al., MICCAI 2018); NCT-CRC-HE-100K (Kather et al., Zenodo 10.5281/zenodo.1214456; PLoS Med 2019); BRACS (Brancati et al., Database 2022, 10.1093/database/baac093); LungHist700 (Diosdado et al., Sci. Data 2024, 10.1038/s41597-024-03944-3); SPIDER (Nechaev et al., arXiv:2503.02876); Camelyon17 (Bándi et al., IEEE TMI 2019) and WILDS (Koh et al., ICML 2021); MHIST (Wei et al., AIME 2021) if you regenerate MHIST. Models: CLIP (Radford et al., 2021); PLIP (Huang et al., Nat. Med. 2023); GigaPath (Xu et al., Nature 2024); H-optimus-0 (Bioptimus, 2024); Midnight (Karasikov et al., 2025); UNI (Chen et al., Nat. Med. 2024), CONCH (Lu et al., Nat. Med. 2024) and Virchow2 (Zimmermann et al., 2024) if you regenerate those cells.

Downloads last month: 5

Total file size:

29.8 GB

Paper for alicemkyn/PRISM-embeddings

SPIDER: A Comprehensive Multi-Organ Supervised Pathology Dataset and Baseline Models

Paper • 2503.02876 • Published Mar 4, 2025 • 4