README.md

metadata

dataset_info:
  features:
    - name: image
      dtype:
        image:
          mode: RGB
    - name: instances
      sequence:
        image:
          mode: '1'
    - name: categories
      sequence:
        class_label:
          names:
            '0': Neoplastic
            '1': Inflammatory
            '2': Connective
            '3': Dead
            '4': Epithelial
    - name: tissue
      dtype:
        class_label:
          names:
            '0': Adrenal Gland
            '1': Bile Duct
            '2': Bladder
            '3': Breast
            '4': Cervix
            '5': Colon
            '6': Esophagus
            '7': Head & Neck
            '8': Kidney
            '9': Liver
            '10': Lung
            '11': Ovarian
            '12': Pancreatic
            '13': Prostate
            '14': Skin
            '15': Stomach
            '16': Testis
            '17': Thyroid
            '18': Uterus
  splits:
    - name: fold1
      num_bytes: 283673837.64
      num_examples: 2656
    - name: fold2
      num_bytes: 267595457.439
      num_examples: 2523
    - name: fold3
      num_bytes: 293079722.82
      num_examples: 2722
  download_size: 1665092597
  dataset_size: 844349017.8989999
configs:
  - config_name: default
    data_files:
      - split: fold1
        path: data/fold1-*
      - split: fold2
        path: data/fold2-*
      - split: fold3
        path: data/fold3-*
license: cc-by-nc-sa-4.0
task_categories:
  - image-segmentation
task_ids:
  - instance-segmentation
language:
  - en
tags:
  - medical
  - cell nuclei
  - H&E
pretty_name: PanNuke
size_categories:
  - 1K<n<10K
paperswithcode_id: pannuke

PanNuke

Dataset Description

• Homepage: PanNuke Dataset for Nuclei Instance Segmentation and Classification
• Leaderboard: Panoptic Segmentation

Description

PanNuke is a semi-automatically generated dataset for nuclei instance segmentation and classification, providing comprehensive nuclei annotations across 19 tissue types and 5 distinct cell categories. The dataset includes a total of 189,744 labeled nuclei, each accompanied by an instance segmentation mask, and contains 7,901 images, each sized 256×256 pixels. The images were captured at x40 magnification with a resolution of 0.25 µm/pixel. The dataset is highly imbalanced, with the "Dead" nuclei category being particularly underrepresented.

Please note that the dataset was created by extracting patches from whole-slide images (WSIs). As a result, some nuclei located at the edges of patches may be cropped, with fewer than 10 visible pixels in certain cases.

Dataset Structure

The dataset is organized into three folds: fold1, fold2, and fold3, consistent with the original dataset structure. Each fold contains data in a tabular format with the following four columns:

• image: The RGB tile of the sample.
• instances: A list of nuclei instances. Each instance represents exactly one nucleus and is in binary format (1 - nucleus, 0 - background)
• categories: An integer class label for each nucleus, corresponding to one of the following categories:
  0. Neoplastic
  1. Inflammatory
  2. Connective
  3. Dead
  4. Epithelial
• tissue: The integer tissue type from which the sample originates, belonging to one of these categories:
  0. Adrenal Gland
  1. Bile Duct
  2. Bladder
  3. Breast
  4. Cervix
  5. Colon
  6. Esophagus
  7. Head & Neck
  8. Kidney
  9. Liver
  10. Lung
  11. Ovarian
  12. Pancreatic
  13. Prostate
  14. Skin
  15. Stomach
  16. Testis
  17. Thyroid
  18. Uterus

Citation

@inproceedings{gamper2019pannuke,
  title={PanNuke: an open pan-cancer histology dataset for nuclei instance segmentation and classification},
  author={Gamper, Jevgenij and Koohbanani, Navid Alemi and Benes, Ksenija and Khuram, Ali and Rajpoot, Nasir},
  booktitle={European Congress on Digital Pathology},
  pages={11--19},
  year={2019},
  organization={Springer}
}

@article{gamper2020pannuke,
  title={PanNuke Dataset Extension, Insights and Baselines},
  author={Gamper, Jevgenij and Koohbanani, Navid Alemi and Graham, Simon and Jahanifar, Mostafa and Khurram, Syed Ali and Azam, Ayesha and Hewitt, Katherine and Rajpoot, Nasir},
  journal={arXiv preprint arXiv:2003.10778},
  year={2020}
}