Lizard 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 Lizard paper, and to share any derivative under the same licence.

Lizard in the nuclei store format

A faithful conversion of the Lizard dataset (Graham and others, 2021) into one store format shared by several nuclei datasets: images and masks in sharded Zarr, one row per region and one row per nucleus in Parquet. Every pixel, every instance and every class label was checked equal to the original PNG and .mat files on 31-10-2026. The original data, its licence and its citation requirements are unchanged.

Source and licence

Source: Lizard, Tissue Image Analytics centre, University of Warwick, https://warwick.ac.uk/lizard-dataset, the released regions: 238 colon regions at 20x from five public sources, every nucleus segmented and classed into six classes. The authors hold back the TCGA part of the dataset for a challenge, so it is not here.
Licence: CC BY-NC-SA 4.0, research purposes only, as the release's README states. This copy carries the same licence. Commercial use is not allowed.
Cite, as the source asks: Graham, S., Jahanifar, M., Azam, A., Nimir, M., Tsang, Y. W., Dodd, K., Hero, E., Sahota, H., Tank, A., Benes, K., Wahab, N., Minhas, F., Raza, S. E. A., El Daly, H., Gopalakrishnan, K., Snead, D., Rajpoot, N. (2021). Lizard: a large-scale dataset for colonic nuclear instance segmentation and classification. Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, 684 to 693.

What changed in the conversion

Pixels: the PNG regions are stored unchanged.
Labels: the .mat instance map is stored as it is, and each nucleus's class comes from the file's id and class arrays. Every nucleus id in the files owns pixels, so none was dropped: 431,913 nuclei in all. Each nucleus's centre is kept as its geometry.
Regions keep their source names (crag_12, pannuke_3, and so on) in group, and their fold from the release's info.csv in split. The 28 pannuke_* regions are PanNuke tiles, so any experiment that holds PanNuke apart must keep them out.
Pixel size: 0.5 um/px, as the paper states for the whole dataset.

Layout

native/lizard/card.json         what the dataset is: source, licence, scale and its
                                evidence, label origin, class names, arrays present
native/lizard/data.zarr         Zarr v3, sharded, Zstd level 3, CRC32C checksums
    images     uint8  (238, 3, H, W)   RGB, padded to the largest region;
                                                 true sizes are in images.parquet
    instances  uint32 (238, H, W)      0 background, one id per nucleus
native/lizard/images.parquet    one row per region
native/lizard/objects.parquet   one row per nucleus (GeoParquet, point geometry)
patches/lizard/                 the same, doubled to 0.25 um/px and cut into
                                18,299 patches of 256 px; a nucleus crossing a
                                patch edge appears in both patches
views/lizard_folds.parquet      the release's three folds as roles: fold1, fold2, fold3

images.parquet columns: image (row number, the index into the arrays), key (the region name), height, width (the true size inside the padded array), um_per_px, split (the fold), patient (not reported), slide (not reported), organ (colon), group (the region name, whose prefix is the source), annotated_classes, label_origin, intensity_low and intensity_high (per-channel percentiles), and parent_image, parent_x, parent_y, parent_scale (filled in the patches 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 Lizard's own, never regrouped. A class code is the position in card.json's native_classes, counted from 1, the same order as the .mat codes; 0 means no class:

code	native_class	nuclei
1	neutrophil	4,116
2	epithelial	210,372
3	lymphocyte	92,238
4	plasma	24,861
5	eosinophil	2,979
6	connective	97,347

Regions per source and fold (the release's info.csv):

source	fold1	fold2	fold3	total
consep	0	0	16	16
crag	21	25	18	64
dpath	20	25	24	69
glas	29	20	12	61
pannuke	0	0	28	28

Nuclei per fold and class:

fold	neutrophil	epithelial	lymphocyte	plasma	eosinophil	connective	total
fold1	1,514	74,529	30,053	8,664	842	33,769	149,371
fold2	1,839	74,941	34,732	8,218	1,138	33,391	154,259
fold3	763	60,902	27,453	7,979	999	30,187	128,283

Region sizes run from 337 to 2175 rows and 500 to 2145 columns. Nuclei per region: minimum 17, median 1368, maximum 9,354.

Reading the store

With zarr (3.x) and pandas:

import pandas as pd
import zarr

root = zarr.open_group("native/lizard/data.zarr", mode="r")
images = pd.read_parquet("native/lizard/images.parquet")
row = images.iloc[0]
image = root["images"][0, :, : row.height, : row.width]      # (3, h, w) uint8
instances = root["instances"][0, : row.height, : row.width]  # (h, w) uint32
objects = pd.read_parquet("native/lizard/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; both read the patches tier.

Limits

Labels are model output refined by pathologists, as the paper describes; nothing was corrected.
The pannuke_* regions duplicate PanNuke tiles at half resolution; the consep_* and glas_* regions duplicate parts of CoNSeP and GlaS. Hold-outs across datasets must account for that.
The TCGA regions of Lizard are not part of the release.
Models trained on this data inherit the non-commercial licence.

Downloads last month: 16

Total file size:

3.35 GB