PanNuke is released by its authors under CC BY-NC-SA 4.0 for research purposes only. By requesting access you agree to those terms, to cite the PanNuke papers, and to share any derivative under the same licence.

PanNuke in the nuclei store format

A faithful conversion of the PanNuke dataset (Gamper and others, 2019 and 2020) into one store format shared by several nuclei datasets: images and masks in sharded Zarr, one row per image and one row per nucleus in Parquet. Every pixel, every instance and every class label was checked equal to the original fold files on 31-10-2026. The original data, its licence and its citation requirements are unchanged.

Source and licence

Source: PanNuke, Tissue Image Analytics centre, University of Warwick, https://warwick.ac.uk/fac/cross_fac/tia/data/pannuke, folds 1 to 3 as distributed.
Licence: CC BY-NC-SA 4.0, research purposes only, as stated on the source page. This copy carries the same licence. Commercial use is not allowed.
Cite, as the source asks: Gamper, J., Koohbanani, N. A., Benet, K., Khuram, A., Rajpoot, N. (2019). PanNuke: an open pan-cancer histology dataset for nuclei instance segmentation and classification. European Congress on Digital Pathology, 11 to 19. Gamper, J., Koohbanani, N. A., Graham, S., Jahanifar, M., Khurram, S. A., Azam, A., Hewitt, K., Rajpoot, N. (2020). PanNuke dataset extension, insights and baselines. arXiv:2003.10778.

What changed in the conversion

Pixels: PanNuke stores images as float64; they are rounded and clipped to uint8. No resampling.
Masks: PanNuke stores one instance channel per class (six channels, the last being background). They are merged into one instance map: a pixel in two class channels goes to the later channel, and a nucleus left with no pixel is dropped. The class of each nucleus is its channel. The store holds 189,744 nuclei in 7,901 images.
Tissue type and fold are kept per image. Folds are named fold1, fold2 and fold3 by their image counts (2,656, 2,523 and 2,722), as the organisers numbered them.
Pixel size: PanNuke states 40x and no pixel size; the store records 0.25 um/px with the origin assumed and that evidence, so nothing pretends to know more than the source.

Layout

native/pannuke/card.json        what the dataset is: source, licence, scale and its
                                evidence, label origin, class names, arrays present
native/pannuke/data.zarr        Zarr v3, sharded, Zstd level 3, CRC32C checksums
    images     uint8  (7901, 3, 256, 256)   RGB
    instances  uint32 (7901, 256, 256)      0 background, one id per nucleus per image
native/pannuke/images.parquet   one row per image
native/pannuke/objects.parquet  one row per nucleus (GeoParquet, point geometry)
views/pannuke_split1.parquet    the organisers' split 1: fold1 train, fold2 validation,
                                fold3 test
views/pannuke_split2.parquet    split 2: fold2 train, fold1 validation, fold3 test
views/pannuke_split3.parquet    split 3: fold3 train, fold2 validation, fold1 test

images.parquet columns: image (row number, the index into the arrays), key (foldN/NNNN), height, width, um_per_px, split, patient (not reported by PanNuke), slide (not reported), organ (the tissue type as PanNuke spells it), group, annotated_classes, label_origin, intensity_low and intensity_high (per-channel percentiles), and parent_image, parent_x, parent_y, parent_scale (not applicable in the native tier).

objects.parquet columns: image, object (the nucleus's id in instances), native_class, label_origin, x_min, y_min, x_max, y_max (bounding box in pixels) and geometry (the nucleus centre, WKB).

Class names are PanNuke's own, never regrouped. A class code is the position in card.json's native_classes, counted from 1; 0 means no class:

code	native_class	nuclei
1	neoplastic	77,403
2	inflammatory	32,276
3	connective	50,585
4	dead	2,908
5	epithelial	26,572

Images per organ and fold:

organ	fold1	fold2	fold3
Adrenal_gland	134	148	155
Bile-duct	159	103	158
Bladder	21	61	64
Breast	827	749	775
Cervix	155	52	86
Colon	478	468	494
Esophagus	155	128	141
HeadNeck	70	169	145
Kidney	37	56	41
Liver	65	66	93
Lung	80	53	51
Ovarian	54	40	52
Pancreatic	78	89	28
Prostate	77	37	68
Skin	59	87	41
Stomach	50	48	48
Testis	60	79	57
Thyroid	80	84	62
Uterus	17	6	163

Nuclei per image: minimum 0, median 20, maximum 289; 343 images hold no nucleus, which the PanNuke evaluation protocol skips for binary PQ.

Reading the store

With zarr (3.x) and pandas:

import pandas as pd
import zarr

root = zarr.open_group("native/pannuke/data.zarr", mode="r")
image = root["images"][0]          # (3, 256, 256) uint8
instances = root["instances"][0]   # (256, 256) uint32
images = pd.read_parquet("native/pannuke/images.parquet")
objects = pd.read_parquet("native/pannuke/objects.parquet")
classes = objects[objects["image"] == 0].set_index("object")["native_class"]

With pathomix[nuclei-datasets], the package this format belongs to, a store opens with Store.open, a view loads with load_role, and a PyTorch loader comes from pathomix.nuclei.datasets.loading. The package's patches tier is the native tier for PanNuke, since its images already are 256 px patches at the recorded scale; copy native/pannuke to patches/pannuke when a tool expects that tier.

Limits

Labels are PanNuke's semi-automatic labels checked by pathologists, with their known imperfections; nothing was corrected.
The pixel size is assumed, not measured.
Models trained on this data inherit the non-commercial licence.

Downloads last month: 16

Total file size:

1.26 GB

Paper for saahilw/nuclei-pannuke

PanNuke Dataset Extension, Insights and Baselines

Paper • 2003.10778 • Published Mar 24, 2020 • 1