---
license: mit
tags:
- pathology
- histology
- cell-phenotyping
- foundation-models
- benchmark
pretty_name: "PathoCell Benchmark Datasets"
---

# PathoCell: A Benchmark for Foundational Models in Computational Pathology

This repository contains the datasets for **PathoCellBench** (formerly PhenoBench), a comprehensive benchmark designed to evaluate the cell phenotyping capabilities of pathology Foundation Models.

The benchmark suite includes four key datasets, all processed into the efficient LMDB (Lightning Memory-Mapped Database) format to facilitate large-scale experimentation.

## Datasets Included

This collection consolidates the following datasets:

*   **[PathoCell](https://huggingface.co/datasets/FabianReith/phenobench/tree/main/data/phenocell)** (formerly PhenoCell): Our new, challenging benchmark dataset for H&E cell phenotyping with 14 granular cell types.
*   **[PanNuke](https://huggingface.co/datasets/FabianReith/phenobench/tree/main/data/pannuke)**: A pan-cancer nucleus segmentation and classification dataset. **Note: This dataset requires a one-time setup step after downloading (see instructions below).**
*   **[Lizard](https://huggingface.co/datasets/FabianReith/phenobench/tree/main/data/lizard)**: A large-scale nucleus segmentation and classification dataset in colon histology.
*   **[ARCTIQUE](https://huggingface.co/datasets/FabianReith/phenobench/tree/main/data/arctique)**: A colorectal cancer histology dataset for tile-based classification.

---

## **IMPORTANT**: Setup Instructions for the PanNuke Dataset

Due to repository file size limits, the large database file for the **PanNuke dataset** (`data.mdb`) was split into smaller parts. Before you can use this dataset, you must reassemble these parts into a single file.

After downloading the repository, navigate into the PanNuke `lmdb` directory and run the `cat` command as shown below.

    # 1. Navigate to the correct directory
    cd data/pannuke/pannuke_lmdb/

    # 2. Run the cat command to merge the parts into a single file
    cat data.mdb.part_* > data.mdb

    # 3. Verify the new file 'data.mdb' has been created.
    #    You can now optionally delete the .part files to save space.
    #    rm data.mdb.part_*

The other datasets (`PathoCell`, `Lizard`, and `ARCTIQUE`) are ready to use immediately after download and do not require this step.

---

## How to Use

You can download the entire dataset collection using the Hugging Face `datasets` library.

    from datasets import load_dataset

    # Download all the dataset files
    # Note: This does not load the LMDB files into memory, it only downloads the raw files.
    # Remember to run the reassembly command for PanNuke after downloading.
    dataset_path = load_dataset("FabianReith/phenobench", cache_dir="./huggingface_data")

    print(f"Dataset downloaded to: {dataset_path.cache_files}")

After reassembly, you can use your own custom data loader to read from the resulting `.mdb` files.

## Dataset Structure

The data is organized into folders, one for each of the four datasets.

````
/
├── data/
│   ├── arctique/
│   │   ├── arctique_lmdb/
│   │   │   ├── data.mdb
│   │   │   └── lock.mdb
│   │   └── ... (other metadata files)
│   │
│   ├── lizard/
│   │   ├── lizard_lmdb/
│   │   │   ├── data.mdb
│   │   │   └── lock.mdb
│   │   └── ...
│   │
│   ├── pannuke/
│   │   ├── pannuke_lmdb/    (Requires reassembly)
│   │   │   ├── data.mdb.part_aa
│   │   │   ├── data.mdb.part_ab
│   │   │   └── ...
│   │   └── ...
│   │
│   └── phenocell/    (PathoCell legacy name)
│       ├── phenocell/
│       │   ├── data.mdb
│       │   └── lock.mdb
│       └── ...
│
├── pathocell_hdf/
│   └── ... (contains the PathoCell dataset in HDF5 format)
│
├── .gitattributes
└── README.md
````

## Citation Information

If you use our benchmark or the **PathoCell** dataset in your research, please cite our work.

*TODO*


Please also ensure you cite the original papers for the **PanNuke**, **Lizard**, and **ARCTIQUE** datasets if you use them.

## License

This dataset collection is licensed under the **MIT License**.