README.md

metadata

license: cc-by-nc-sa-4.0
task_categories:
  - image-segmentation
  - object-detection
language:
  - en
tags:
  - medical
  - pathology
  - histopathology
  - h-and-e
  - pd-l1
  - lung
  - nsclc
  - ignite
size_categories:
  - n<1K
configs:
  - config_name: he
    data_files:
      - split: train
        path: he/train-*
      - split: test
        path: he/test-*
  - config_name: pdl1
    data_files:
      - split: train
        path: pdl1/train-*
      - split: validation
        path: pdl1/validation-*
      - split: test
        path: pdl1/test-*
  - config_name: nuclei
    data_files:
      - split: train
        path: nuclei/train-*
      - split: validation
        path: nuclei/validation-*
      - split: test
        path: nuclei/test-*

IGNITE Data Toolkit (mirror)

Mirror of the IGNITE Data Toolkit by Spronck et al. (Radboud UMC), originally distributed on Zenodo (10.5281/zenodo.15674785) and accompanied by DIAGNijmegen/ignite-data-toolkit. The dataset accompanies "A tissue and cell-level annotated H&E and PD-L1 histopathology image dataset in non-small cell lung cancer" (arXiv:2507.16855).

License: CC BY-NC-SA 4.0 - non-commercial, share-alike. Attribution to the original authors is required.

Contents

155 unique patients, 887 fully annotated regions of interest from a multi-stain, multi-centric, multi-scanner cohort (Radboud UMC, Sacro Cuore Don Calabria, TCGA-LUAD/LUSC). The release splits into three task-defined subsets, exposed here as named configs:

Config	Task	ROIs
he	H&E tissue compartment segmentation	408
pdl1	PD-L1+ tumor cell detection	344
nuclei	PD-L1 IHC nuclei detection	135

H&E tissue segmentation (he config)

Splits follow data_overview.csv:

Split	ROIs	Notes
train	269	Train pool — paper uses 5-fold CV via the validation_fold col.
test	139	Held-out evaluation set (62 TCGA + 77 Radboud ROIs).

Each row is one ROI with paired image/mask in two field-of-view variants:

Column	Type	Description
image	Image	Base ROI (inner annotated region only)
mask	Image	16-class pixel mask aligned to image
image_with_context	Image	Same ROI extended to a 1792x1792 view (annotated context)
mask_with_context	Image	16-class pixel mask aligned to image_with_context
validation_fold	string	5-fold CV assignment (fold0..fold4); empty for test rows
patient_id	int32	Patient identifier
roi_id	int32	ROI index within patient
name	string	patient<id>_he_roi<idx> (matches the original release)
source	string	rumc, scdc, or tcga
specimen_type	string	resection, biopsy, or tissue_microarray
organ	string	Anatomical site (lung, liver, bone, brain, ...)
histological_subtype	string	adenocarcinoma, squamous_cell_carcinoma, ...
stain	string	Always H&E for this config
scanner	string	WSI scanner model
shape	string	Original (height, width) tuple as a string
area_mm2	float32	Annotated tissue area in mm^2
original_tcga_id	string	TCGA case ID for TCGA-sourced ROIs (empty otherwise)

Labels (also shipped as he_label_map.json):

ID	Class	ID	Class
0	Unannotated	9	Erythrocytes
1	Background	10	Bronchial epithelium
2	Tumor epithelium	11	Mucus/Plasma/Fluids
3	Reactive epithelium	12	Cartilage/Bone
4	Stroma	13	Macrophages
5	Inflammation	14	Muscle
6	Alveolar tissue	15	Liver
7	Fatty tissue	16	Keratinization
8	Necrotic tissue

The paper's evaluation pipeline treats class 0 ("Unannotated", i.e. surrounding context in _with_context masks) as an ignore label during Dice/IoU computation. Downstream loaders should mirror that to reproduce paper-comparable scores.

Mirror-specific note: In the original Zenodo release, base ROI masks (the inner-crop view) store class label L as the byte value (256 - L) mod 256 (e.g. label 4 -> byte 252). The _with_context masks already store labels directly. In this HuggingFace mirror both mask and mask_with_context are written with the canonical 0..16 labels - base masks were pre-decoded during upload, so downstream code does not need to handle the encoding quirk.

The paper recommends training-time 5-fold CV via the validation_fold column on the train split, and reports final numbers on the held-out test split.

PD-L1 / nuclei detection (pdl1, nuclei configs)

These configs hold images plus per-image metadata only (same columns as he except no mask/_with_context fields and no validation_fold). The detection ground truth is in MS-COCO JSON format and is shipped as raw sidecar files because COCO-style nested annotations are a poor fit for columnar parquet:

Path	Subset	Notes
coco/pdl1_annotations.json	pdl1	Main annotations
coco/pdl1_test_set_all_readers.json	pdl1	Multi-reader test set
coco/nuclei_annotations.json	nuclei	Main annotations
coco/nuclei_test_set_all_readers.json	nuclei	Multi-reader test set

Use the row's name field (== image_id in COCO images[*].file_name = "<name>.png") to look up bounding-box / point annotations.

Splits follow data_overview.csv directly (no fold column for the detection tasks).

Loading

from datasets import load_dataset

# H&E tissue segmentation
he_train = load_dataset("Angelou0516/IGNITE", "he", split="train")  # 269 ROIs
he_test  = load_dataset("Angelou0516/IGNITE", "he", split="test")   # 139 ROIs
print(he_test[0]["mask_with_context"])  # PIL Image L-mode, labels 0..16

# PD-L1+ tumor cell detection
pdl1 = load_dataset("Angelou0516/IGNITE", "pdl1")

# PD-L1 IHC nuclei detection
nuc = load_dataset("Angelou0516/IGNITE", "nuclei")

For detection COCO annotations, download the JSON sidecars with huggingface_hub.hf_hub_download.

Sidecar files (raw)

• he_label_map.json — class id -> name
• data_overview.csv — per-ROI metadata (887 rows x 17 cols), authoritative for splits / folds
• coco/*.json — detection annotations (4 files, see table above)

Citation

@article{Spronck2025ignite,
  title   = {A tissue and cell-level annotated H\&E and PD-L1 histopathology image dataset in non-small cell lung cancer},
  author  = {Spronck, Joey and van Eekelen, Leander and van Midden, Dominique and others},
  journal = {arXiv preprint arXiv:2507.16855},
  year    = {2025},
  doi     = {10.48550/arXiv.2507.16855}
}

Mirror maintained by Angelou0516. For the official authoritative release see the Zenodo record and GitHub toolkit.