CAMEO-Breast: A Multimodal Benchmark Dataset of Aligned H&E Patches and Gene Expression Profiles in the Breast

Citation

If you use this dataset, please cite it directly and the original breast study:

@dataset{kuijs_cameo_breast_2026,
  author    = {Kuijs, Merel and Richter, Till and Gindra, Rushin H and Traeuble, Korbinian and
               Matek, Christian and Lukn{\'a}rov{\'a}, Rebeka and Peng, Tingying and Theis, Fabian J},
  title     = {CAMEO-Breast: A Multimodal Benchmark Dataset of Aligned H&E Patches and Gene Expression Profiles in the Breast},
  year      = {2026},
  publisher = {Hugging Face},
  doi       = {10.57967/hf/7909},
  url       = {https://huggingface.co/datasets/theislab/CAMEO-Breast}
}

@article{breast,
  title   = {High resolution mapping of the tumor microenvironment using integrated single-cell, spatial and in situ analysis},
  author  = {Janesick, Amanda and Shelansky, Robert and Gottscho, Andrew D and Wagner, Florian and Williams, Stephen R and Rouault, Morgane and Beliakoff, Ghezal and Morrison, Carolyn A and Oliveira, Michelli F and Sicherman, Jordan T and others},
  journal = {Nature Communications},
  volume  = {14},
  number  = {1},
  pages   = {8353},
  year    = {2023}
}

---

Dataset Description

This dataset is part of the CAMEO framework for multimodal spatial transcriptomics learning. It contains paired histology images and gene expression data derived from 7 publicly available 10x Xenium breast cancer samples from 4 patients, originally hosted on the 10x Genomics website.

Each row represents one niche — a 224×224 pixel crop of an H&E-stained histology slide paired with the single-cell gene expression profiles of all cells located within that crop, together with expert pathologist niche annotations, per-cell coordinates, and cell-type composition. In total, the dataset contains 126,770 niches encompassing approximately 2.3 million cells across 7 samples. We constructed these niche-level paired representations by spatially aligning the histological and transcriptomic modalities using SpatialData, tessellating non-overlapping crops across each slide, and applying a quality control filter to exclude niches with less than 50% tissue coverage. Transcripts from partially included cells are treated as whole-cell data within the niche. Niche-level annotations were provided by our collaborating expert pathologist using the Elston-Ellis grading system.

In addition to raw modality data, the dataset includes a set of precomputed embeddings from several unimodal foundation models to facilitate research on multimodal and unimodal representation learning.

• Organization: Theislab
• Source data: 10x Genomics Xenium breast cancer samples (publicly available)
• License: CC BY-NC-SA 4.0

---

Dataset Structure

Splits

The dataset is stored as a single train split containing all 126,770 niches across all 7 samples.

Split	Niches
Full dataset (train)	126,770

---

Column Descriptions

Each row corresponds to one niche (224×224 px patch). The following columns are included:

Identifiers and labels

Column	Type	Description
name	string	Sample (slide) identifier, e.g. "TENX97_vectorized_annotated_allpolys_noCT". Maps to one of the 7 Xenium samples.
annotation	ClassLabel (int64)	Expert pathologist niche annotation, encoded as an integer. See Niche Label Mapping below.
species	ClassLabel (int64)	Species label. Always 0 (human) in this cohort.
cancer	ClassLabel (int64)	Cancer flag (schema inherited from the combined CAMEO dataset).
tissue	ClassLabel (int64)	Tissue label (schema inherited from the combined CAMEO dataset).

Raw modality data

Column	Type	Shape	Description
image	Image	224×224 RGB	H&E-stained histology patch.
gexp	Array2D float32	(200, 280)	Raw gene expression counts per cell. Up to 200 cells per niche (zero-padded); 280 Xenium panel genes. Use mask to identify valid cells.
spot_gexp	Array2D float32	(1, 280)	Niche-level pseudobulk gene expression (sum over valid cells).
mask	Sequence bool	(200,)	Boolean mask indicating valid cells (True = real cell, False = padding).
cell_coords	Array2D int32	(200, 2)	Cell centroid coordinates (x, y) in pixel space within the 224×224 patch. Padded to 200 rows.

Precomputed embeddings

All embeddings are niche-level representations derived from the raw modalities.

Column	Type	Shape	Description
img_embed	Sequence float32	(1024,)	Image embedding from UNI
conch_embedding	Sequence float64	(512,)	Image embedding from CONCH
ctranspath_embedding	Sequence float64	(768,)	Image embedding from CTransPath.
gexp_embed	Sequence float32	(128,)	Gene expression embedding learned by a self-supervised Graph Attention Network.
scvi_pool	Sequence float64	(128,)	scVI embedding, pooled over valid cells in the niche.
scvi_pseudobulk	Sequence float64	(128,)	scVI embedding computed from the pseudobulk niche expression profile.
pca_pool	Sequence float64	(128,)	PCA embedding (128 components), pooled over valid cells in the niche.
pca_pseudobulk	Sequence float64	(128,)	PCA embedding computed from the pseudobulk niche expression profile.
nicheformer_pool	Sequence float64	(512,)	Nicheformer embedding, pooled over valid cells in the niche.
scgpt_pool	Sequence float64	(512,)	scGPT embedding, pooled over valid cells in the niche.

Niche Label Mapping

The annotation column contains integer class labels corresponding to expert-annotated niche types:

Integer	Niche type
0	DCIS
1	Flat epithelial atypia
2	Invasive Adenocarcinoma
3	Lymphocyte Rich Tumor Stroma
4	NOANNOT
5	Necrosis
6	Normal breast lobules
7	adenosis
8	blood_vessels
9	columnar cell change
10	duct
11	fat
12	stroma
13	tumor_epithelium

---

Loading the Dataset

Standard loading

from datasets import load_dataset

dataset = load_dataset("theislab/CAMEO-Breast")
train_ds = dataset["train"]

# Access one example
example = train_ds[0]
print(example.keys())
# dict_keys(['name', 'image', 'img_embed', 'gexp_embed', 'cell_type_ratio',
#            'annotation', 'species', 'cancer', 'tissue', 'gexp', 'mask',
#            'cell_coords', 'spot_gexp', 'conch_embedding', 'ctranspath_embedding',
#            'pca_pool', 'pca_pseudobulk', 'scvi_pool', 'scvi_pseudobulk',
#            'nicheformer_pool', 'scgpt_pool'])

# The image is a PIL Image
print(example["image"].size)       # (224, 224)

# Gene expression: shape (200, 280), use mask to select valid cells
import numpy as np
gexp = np.array(example["gexp"])         # shape (200, 280)
mask = np.array(example["mask"])         # shape (200,) bool
gexp_valid = gexp[mask]                  # shape (n_cells, 280)

# Decode the niche label
label_name = train_ds.features["annotation"].int2str(example["annotation"])
print(label_name)  # e.g. "Invasive Adenocarcinoma"

Streaming (avoids downloading all ~40 GB upfront)

from datasets import load_dataset

dataset = load_dataset("theislab/CAMEO-Breast", streaming=True)
for example in dataset["train"]:
    # process one niche at a time
    break

Filtering by sample

train_samples = ["TENX97_vectorized_annotated_allpolys_noCT", ...]
train_split = dataset["train"].filter(lambda x: x["name"] in train_samples)

---

License

This dataset is distributed under the Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0) license.