CellHIST-Bench

Dataset Summary

CellHIST-Bench is a benchmark dataset for histology-based spatial gene expression inference. It provides paired whole-slide histopathology images, spatial transcriptomics labels, cell segmentation results, spot-centered patch metadata, and cell-to-patch correspondence information.

The dataset supports two levels of spatial transcriptomics annotations:

• Spot-resolution spatial transcriptomics, where each spot is associated with a local histology patch and a spot-level gene expression vector.
• Single-cell-resolution spatial transcriptomics, where each cell is associated with a single-cell gene expression vector and can be linked to image-detected cells.

CellHIST-Bench is designed to support benchmarking of weakly supervised learning methods for predicting spatial gene expression from histopathology images.

---

Dataset Organization

The dataset is organized by tissue or cancer type. Each folder contains whole-slide images, cell segmentation files, patch metadata, patch-cell relationship files, and gene expression labels.

A typical directory structure is shown below:

/data/west_bench
├── LUNG
│   ├── wsis
│   ├── cellvit++_seg
│   ├── patches
│   ├── patches_cell
│   ├── st_spot_label
│   └── st_cell_label
├── IDC
├── SKCM
└── ...

---

Spot-Resolution Samples

For spot-resolution spatial transcriptomics samples, the dataset contains the following files:

/data/west_bench
├── LUNG
│   ├── wsis
│   │   ├── TENX62.tif
│   │   └── ...
│   │
│   ├── cellvit++_seg
│   │   ├── TENX62_seg.parquet
│   │   ├── TENX62.h5
│   │   └── ...
│   │
│   ├── patches
│   │   ├── TENX62.h5
│   │   └── ...
│   │
│   ├── patches_cell
│   │   ├── TENX62.json
│   │   └── ...
│   │
│   └── st_spot_label
│       ├── TENX62.h5ad
│       └── ...

wsis/

This folder stores whole-slide images.

Each file is a .tif image:

TENX62.tif

The WSI file contains the original histopathology image. These files are usually very large and may require specialized tools such as OpenSlide or tifffile for reading.

cellvit++_seg/

This folder stores cell segmentation results generated by CellViT++.

Each sample may contain:

TENX62_seg.parquet
TENX62.h5

The *_seg.parquet file is a DataFrame containing cell-level segmentation information. Typical columns include:

['geometry', 'class', 'cell_id']

The corresponding .h5 file stores cell-level information. Typical keys include:

['cell_coords', 'cell_embedding', 'cell_class_id']

where:

• cell_coords stores cell coordinates;
• cell_embedding stores cell-level image embeddings;
• cell_class_id stores predicted cell class IDs.

patches/

This folder stores metadata for spot-centered patches.

Each .h5 file contains:

['barcode', 'coords']

where:

• barcode is the spatial transcriptomics spot barcode;
• coords records the patch coordinates in the WSI.

Note that the raw patch image array is not stored in this file. Users need to crop the patch from the corresponding WSI according to the provided coordinates.

patches_cell/

This folder describes the cells contained in each spot-centered patch.

Each .json file stores a dictionary indexed by spot barcode. A typical entry is:

{
  "barcode": {
    "cell_index": [99],
    "in_spot": [1]
  }
}

where:

• cell_index indicates the indices of cells located in the patch;
• in_spot indicates whether each cell is located inside the spatial spot region.

Each patch corresponds to a local image region of:

224 × 224 pixels

st_spot_label/

This folder stores spot-level gene expression labels.

Each .h5ad file contains a spot-level gene expression matrix:

num_spots × num_genes

The row index corresponds to the spot barcode.

---

Single-Cell-Resolution Samples

For single-cell-resolution spatial transcriptomics samples, the dataset contains all files used in spot-resolution samples, with additional files for single-cell gene expression labels and spot-cell mapping.

A typical structure is shown below:

/data/west_bench
├── LUNG
│   ├── wsis
│   │   ├── TENX141.tif
│   │   └── ...
│   │
│   ├── cellvit++_seg
│   │   ├── TENX141_seg.parquet
│   │   ├── TENX141.h5
│   │   ├── TENX141_st_cell_idx.parquet
│   │   └── ...
│   │
│   ├── patches
│   │   ├── TENX141.h5
│   │   └── ...
│   │
│   ├── patches_cell
│   │   ├── TENX141.json
│   │   └── ...
│   │
│   ├── st_spot_label
│   │   ├── TENX141.h5ad
│   │   └── ...
│   │
│   └── st_cell_label
│       ├── TENX141.h5ad
│       └── ...

*_st_cell_idx.parquet

This file stores the mapping between spatial transcriptomics cell indices and image-detected cell IDs.

Typical columns include:

['st_cell_index', 'cell_id']

where:

• st_cell_index is the cell index in the spatial transcriptomics data;
• cell_id is the cell ID detected from the WSI by CellViT++.

st_cell_label/

This folder stores single-cell-level gene expression labels.

Each .h5ad file contains a single-cell gene expression matrix:

num_cells × num_genes

The row index corresponds to:

st_cell_index

---

Data Fields

Whole-Slide Images

File	Description
*.tif	Whole-slide histopathology image

Cell Segmentation Files

Field	Description
geometry	Cell geometry or polygon information
class	Predicted cell category
cell_id	Unique cell identifier in the WSI
cell_coords	Cell coordinates
cell_embedding	Cell-level image embedding
cell_class_id	Numeric cell class ID

Patch Metadata Files

Field	Description
barcode	Spatial transcriptomics spot barcode
coords	Patch coordinates in the WSI

Patch-Cell Relationship Files

Field	Description
cell_index	Indices of cells located in the patch
in_spot	Binary indicator showing whether the cell is inside the spot region

Spot-Level Gene Expression Labels

Field	Description
.X	Spot-level gene expression matrix
.obs	Spot metadata
.var	Gene metadata
index	Spot barcode

Single-Cell-Level Gene Expression Labels

Field	Description
.X	Single-cell-level gene expression matrix
.obs	Cell metadata
.var	Gene metadata
index	st_cell_index

---

Supported Tasks

CellHIST-Bench can be used for the following tasks:

1. Spot-level spatial gene expression prediction

   Given a spot-centered histology patch, predict the corresponding spot-level gene expression vector.

2. Single-cell-level gene expression prediction

   Given image-derived cell information and local histological context, predict the gene expression profile of individual cells.

3. Weakly supervised spatial gene expression inference

   Evaluate whether weakly supervised models can infer molecular profiles from histopathology images.

4. Cell-aware representation learning

   Use cell segmentation, cell embeddings, and patch-cell relationships to learn cell-aware histology representations.

5. Cross-resolution spatial transcriptomics analysis

   Compare spot-level and single-cell-level spatial gene expression prediction under a unified data organization.

---

Loading Examples

Load Spot-Level Gene Expression Labels

import scanpy as sc

adata = sc.read_h5ad("LUNG/st_spot_label/TENX62.h5ad")

print(adata)
print(adata.X.shape)  # num_spots × num_genes
print(adata.obs.head())
print(adata.var.head())

Load Single-Cell-Level Gene Expression Labels

import scanpy as sc

adata = sc.read_h5ad("LUNG/st_cell_label/TENX141.h5ad")

print(adata)
print(adata.X.shape)  # num_cells × num_genes
print(adata.obs.head())
print(adata.var.head())

Load Patch Metadata

import h5py

with h5py.File("LUNG/patches/TENX62.h5", "r") as f:
    print(list(f.keys()))
    barcodes = f["barcode"][:]
    coords = f["coords"][:]

print(barcodes.shape)
print(coords.shape)

Load Cell Segmentation Results

import h5py
import pandas as pd

seg_df = pd.read_parquet("LUNG/cellvit++_seg/TENX62_seg.parquet")
print(seg_df.head())

with h5py.File("LUNG/cellvit++_seg/TENX62.h5", "r") as f:
    print(list(f.keys()))
    cell_coords = f["cell_coords"][:]
    cell_embedding = f["cell_embedding"][:]
    cell_class_id = f["cell_class_id"][:]

print(cell_coords.shape)
print(cell_embedding.shape)
print(cell_class_id.shape)

Load Patch-Cell Relationship Files

import json

with open("LUNG/patches_cell/TENX62.json", "r") as f:
    patch_cells = json.load(f)

first_barcode = list(patch_cells.keys())[0]

print(first_barcode)
print(patch_cells[first_barcode])

Load Spot-Cell Mapping for Single-Cell-Resolution Samples

import pandas as pd

mapping = pd.read_parquet("LUNG/cellvit++_seg/TENX141_st_cell_idx.parquet")

print(mapping.head())
print(mapping.columns)

---

Recommended Evaluation Settings

CellHIST-Bench supports evaluation under different task settings.

Spot-Level Prediction

The model takes a spot-centered histology patch as input and predicts the corresponding spot-level gene expression vector.

Input:

spot-centered histology patch

Output:

spot-level gene expression vector

Label file:

st_spot_label/*.h5ad

Single-Cell-Level Prediction

The model takes cell-level visual information and local histological context as input and predicts single-cell gene expression.

Input:

cell-level image features and local patch context

Output:

single-cell gene expression vector

Label file:

st_cell_label/*.h5ad

Mapping file:

cellvit++_seg/*_st_cell_idx.parquet

---

Dataset Splits

The dataset can be split by tissue type, sample ID, or task setting depending on the evaluation protocol.

A typical split may include:

train
validation
test

When evaluating generalization, users should ensure that slides from the same biological sample are not shared across training and test sets.

If official split files are provided, users should follow the official split protocol.

---

Intended Use

This dataset is intended for research on:

• histology-based spatial gene expression prediction;
• weakly supervised learning for spatial transcriptomics;
• computational pathology;
• cell-aware histology representation learning;
• integration of histology images and spatial omics data;
• single-cell and spatial transcriptomics analysis.

This dataset is intended for research use only.

---

Out-of-Scope Use

This dataset should not be used for:

• clinical diagnosis;
• treatment recommendation;
• patient identification;
• direct medical decision-making;
• commercial use without proper permission.

---

Limitations

Large WSI Files

Whole-slide images are large and may require specialized tools and sufficient storage capacity.

Patch Images Are Not Directly Stored

The patches/*.h5 files provide patch coordinates and barcodes, but do not directly store raw patch image arrays. Users need to crop patches from the corresponding WSI.

Cell Segmentation Depends on CellViT++

Cell-level information is derived from CellViT++ segmentation. Downstream analysis may therefore be affected by the accuracy and bias of the segmentation model.

Resolution Differences

Spot-resolution and single-cell-resolution samples have different label granularities. Results across these two settings should be interpreted carefully.

Gene Expression Sparsity

Spatial transcriptomics and single-cell transcriptomics data are sparse and noisy. Prediction performance should be interpreted with appropriate biological and technical considerations.

---

Ethical Considerations

The dataset contains histopathology images and molecular profiles derived from biomedical samples. Users are responsible for ensuring that their use of the dataset complies with applicable ethical, institutional, and legal requirements.

The dataset should only be used for research purposes and should not be used for patient identification, clinical diagnosis, or treatment recommendation.

Contact

For questions about the dataset, please open an issue in this Hugging Face dataset repository.