SPIDER-BREAST Dataset

SPIDER is a collection of supervised pathological datasets covering multiple organs, each with comprehensive class coverage. These datasets are professionally annotated by pathologists.

If you would like to support, sponsor, or obtain a commercial license for the SPIDER data and models, please contact us at models@hist.ai.

For a detailed description of SPIDER, methodology, and benchmark results, refer to our research paper:

📄 SPIDER: A Comprehensive Multi-Organ Supervised Pathology Dataset and Baseline Models
View on arXiv

This repository contains the SPIDER-breast dataset. To explore datasets for other organs, visit the Hugging Face HistAI page or GitHub. SPIDER is regularly updated with new organs and data, so follow us on Hugging Face to stay updated.

---

Overview

SPIDER-breast is a supervised dataset of image-class pairs for the breast organ. Each data point consists of:

• A central 224×224 patch with a class label
• 24 surrounding context patches of the same size, forming a composite 1120×1120 region
• Patches are extracted at 20X magnification

We provide a train-test split for consistent benchmarking. The split is done at the slide level, ensuring that patches from the same whole slide image (WSI) do not appear in both training and test sets. Users can also merge and re-split the data as needed.

How to Use

Downloading the Dataset

Option 1: Using huggingface_hub

from huggingface_hub import snapshot_download

snapshot_download(repo_id="histai/SPIDER-breast", repo_type="dataset", local_dir="/local_path")

Option 2: Using git

# Ensure you have Git LFS installed (https://git-lfs.com)
git lfs install
git clone https://huggingface.co/datasets/histai/SPIDER-breast

Extracting the Dataset

The dataset is provided in multiple tar archives. Unpack them using:

cat spider-breast.tar.* | tar -xvf -

Using the Dataset

Once extracted, you will find:

• An images/ folder
• A metadata.json file

You can process and use the dataset in two ways:

1. Directly in Code (Recommended for PyTorch Training)

Use the dataset class provided in scripts/spider_dataset.py. This class takes:

• Path to the dataset (folder containing metadata.json and images/ folder)
• Context size: 5, 3, or 1
  • 5: Full 1120×1120 patches (default)
  • 3: 672×672 patches
  • 1: Only central patches

The dataset class dynamically returns stitched images, making it suitable for direct use in PyTorch training pipelines.

2. Convert to ImageNet Format

To structure the dataset for easy use with standard tools, convert it using scripts/convert_to_imagenet.py. The script also supports different context sizes.

This will generate:

<output_dir>/<split>/<class>/<slide>/<image>

You can then use it with:

from datasets import load_dataset

dataset = load_dataset("imagefolder", data_dir="/path/to/folder")

or

torchvision.datasets.ImageFolder class

---

Dataset Composition

The SPIDER-breast dataset consists of the following classes:

Class	Total Patches
Adenosis	2899
Benign phyllodes tumor	4526
Ductal carcinoma in situ (high-grade)	5632
Ductal carcinoma in situ (low-grade)	5017
Fat	6286
Fibroadenoma	5243
Fibrocystic changes	5027
Fibrosis	6260
Invasive non-special type carcinoma	6142
Lipogranuloma	4941
Lobular invasive carcinoma	5102
Malignant phyllodes tumor	5271
Necrosis	5396
Normal ducts	4891
Normal lobules	5821
Sclerosing adenosis	3423
Typical ductal hyperplasia	5546
Vessels	5469

Total Counts:

• 92,892 central patches
• 984,924 total patches (including context patches)
• 921 total slides used for annotation

---

License

The dataset is licensed under CC BY-NC 4.0 and is for research use only.

Citation

If you use this dataset in your work, please cite:

@misc{nechaev2025spidercomprehensivemultiorgansupervised,
      title={SPIDER: A Comprehensive Multi-Organ Supervised Pathology Dataset and Baseline Models}, 
      author={Dmitry Nechaev and Alexey Pchelnikov and Ekaterina Ivanova},
      year={2025},
      eprint={2503.02876},
      archivePrefix={arXiv},
      primaryClass={eess.IV},
      url={https://arxiv.org/abs/2503.02876}, 
}

Contacts

• Authors: Dmitry Nechaev, Alexey Pchelnikov, Ekaterina Ivanova
• Email: dmitry@hist.ai, alex@hist.ai, kate@hist.ai