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

Citation

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

@dataset{kuijs_cameo_thymus_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-Thymus: A Multimodal Benchmark Dataset of Aligned H\&E Patches and Visium Gene Expression Profiles in the Thymus},
  year      = {2026},
  publisher = {Hugging Face},
  doi       = {10.57967/hf/7908},
  url       = {https://huggingface.co/datasets/theislab/CAMEO-Thymus}
}

@article{thymus,
  title   = {A spatial human thymus cell atlas mapped to a continuous tissue axis},
  author  = {Yayon, Nadav and Kedlian, Veronika R and Boehme, Lena and Suo, Chenqu and Wachter, Brianna T and Beuschel, Rebecca T and Amsalem, Oren and Polanski, Krzysztof and Koplev, Simon and Tuck, Elizabeth and others},
  journal = {Nature},
  volume  = {635},
  number  = {8039},
  pages   = {708--718},
  year    = {2024}
}

---

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 a 10x Visium thymus cohort comprising 19 samples from 11 donors, spanning fetal (post-conception weeks 11–21) and pediatric (neonate to 3 years old) tissue. The dataset was originally created to map T cell development during pre- and early postnatal stages, and regions are annotated using the Cortico-Medullary Axis (CMA), a common coordinate framework developed by the original study's authors.

Unlike the Xenium-based CAMEO cohorts, this dataset is spot-based: each row represents one niche — a 224×224 pixel crop of an H&E-stained histology slide paired with the transcriptome of the single Visium spot whose centroid falls within that crop. In total, the dataset contains 45,096 niches across 19 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.

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: Yayon et al., Nature 2024
• License: CC BY-NC-SA 4.0

---

Dataset Structure

Splits

The dataset is stored as a single train split containing all 45,096 niches across all 19 samples.

Split	Niches
Full dataset (train)	45,096

---

Column Descriptions

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

Note: This is a spot-based Visium dataset. Each niche contains exactly one spot (no cell-level decomposition, no padding mask). This differs from the Xenium-based CAMEO-Lung and CAMEO-Breast datasets.

Identifiers and labels

Column	Type	Description
name	string	Sample (slide) identifier, e.g. "WSSS_F_IMMsp11765870". Maps to one of the 19 Visium samples.
annotation	ClassLabel (int64)	CMA region annotation, encoded as an integer. See Niche Label Mapping below.
tissue	ClassLabel (int64)	Tissue label. Always 0 (thymus) in this cohort.
species	ClassLabel (int64)	Species label. Always 0 (Homo sapiens) in this cohort.
sample_source	string	Biobank or repository from which the sample was obtained, e.g. "Human Developmental Biology Resource".
assay	string	Spatial transcriptomics assay used. Always "Visium Spatial Gene Expression V1" in this cohort.
stain	string	Histological stain. Always "HnE" in this cohort.
tissue_section_thickness	string	Thickness of the tissue section, e.g. "15 μm".

Raw modality data

Column	Type	Shape	Description
image	Image	224×224 RGB	H&E-stained histology patch.
gexp	Array2D float32	(1, 2000)	Visium spot-level gene expression counts. Shape (1, 2000): 1 spot × 2000 panel genes.
cell_coords	Array2D int32	(1, 2)	Spot centroid coordinates (x, y) in pixel space within the 224×224 patch.

Precomputed embeddings

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

Column	Type	Shape	Description
img_embed	Sequence float64	(1024,)	Image embedding from UNI
conch_embedding	Sequence float64	(512,)	Image embedding from CONCH
ctranspath_embedding	Sequence float64	(768,)	Image embedding from CTransPath.
scvi_pool	Sequence float64	(128,)	scVI embedding of the spot transcriptome.
scvi_pseudobulk	Sequence float64	(128,)	scVI embedding computed from the spot transcriptome (equivalent to scvi_pool for spot-based data).
pca_pool	Sequence float64	(128,)	PCA embedding (128 components) of the spot transcriptome.
pca_pseudobulk	Sequence float64	(128,)	PCA embedding of the spot transcriptome (equivalent to pca_pool for spot-based data).
nicheformer_pool	Sequence float64	(512,)	Nicheformer embedding of the spot transcriptome.
scgpt_pool	Sequence float64	(512,)	scGPT embedding of the spot transcriptome.

Niche Label Mapping

The annotation column contains integer class labels corresponding to CMA regions along the cortico-medullary axis:

Integer	Region
0	Capsular
1	Cortical CMJ
2	Cortical level 1
3	Cortical level 2
4	Cortical level 3
5	Medullar CMJ
6	Medullar level 1
7	Medullar level 2
8	Medullar level 3
9	Sub-Capsular
10	unassigned

---

Loading the Dataset

Standard loading

from datasets import load_dataset

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

# Access one example
example = train_ds[0]
print(example.keys())
# dict_keys(['name', 'image', 'gexp', 'cell_coords', 'tissue_section_thickness',
#            'annotation', 'sample_source', 'tissue', 'assay', 'stain', 'species',
#            'conch_embedding', 'ctranspath_embedding', 'img_embed',
#            '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 (1, 2000) — extract the spot vector
import numpy as np
gexp = np.array(example["gexp"])[0]     # shape (2000,)

# Decode the CMA region label
label_name = train_ds.features["annotation"].int2str(example["annotation"])
print(label_name)  # e.g. "Sub-Capsular"

Streaming (avoids downloading all ~4 GB upfront)

from datasets import load_dataset

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

Filtering by sample

train_samples = ["WSSS_F_IMMsp11765870", ...]
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.