Lizard-Mitosis

Overview

Lizard-Mitosis contains 4,981 colour images of colon tissue stained with haematoxylin and eosin (H&E). Each image is 256 × 256 pixels. The masks outline individual nuclei and label six cell types plus dividing cells (mitosis).

The HoVer-NeXt researchers created this version by adding mitosis labels to the Lizard dataset. It can be used to train and evaluate nucleus segmentation and classification models.

License: CC BY-NC-SA 4.0. You may use and share the data for noncommercial work. Credit the creators, describe your changes, and use the same license when sharing adapted data.

Files and splits

Split	File	Images	Marked objects
train	train.parquet	3,423	366,483
val	val.parquet	856	91,163
test	test.parquet	702	112,341

These are the original training, validation and test sets. The source calls the validation set valid; it is named val here. Object counts include repeats where images overlap.

Some training and validation images overlap. Evaluation on the validation set may therefore include tissue seen during training. Patient and slide IDs are unavailable.

The data come from lizard_mitosis.zip in Zenodo version 2.

Fields

Each row contains one image and its masks.

Field	Contents
id	Unique image name based on its source split and row number
image	PNG bytes; HWC RGB uint8, shape (256, 256, 3), values 0–255
inst	Numbered instance mask as uint16 PNG bytes; HW (256, 256)
type	Class mask as uint8 PNG bytes; HW (256, 256), values 0–7
source_split	train, valid or test, as named in the source arrays
source_index	Original row number within that split, starting at 0
source_image_file, source_label_file	Original NumPy array paths inside the archive

Image channels are red, green and blue. Zero means background in both masks. Each marked object has its own number within an image. Original object numbers and class labels are kept.

Labels

Number	Class
0	background
1	neutrophil
2	epithelial-cell
3	lymphocyte
4	plasma-cell
5	eosinophil
6	connective-tissue-cell
7	mitosis

The class numbers follow the HoVer-NeXt release. Most outlines mark nuclei; eosinophil outlines can also include some surrounding cytoplasm.

Mitosis means a cell is dividing. Some dividing cells are missing from the labels or have the wrong label.

In 2 images, different parts of the same object have different class labels. These original labels are kept; the affected object numbers are listed in manifest.json.

Download and read a sample

from io import BytesIO

import numpy as np
import polars as pl
from huggingface_hub import hf_hub_download
from PIL import Image

path = hf_hub_download(
    repo_id="csmp-hub/lizard-mitosis", filename="train.parquet", repo_type="dataset"
)
row = pl.scan_parquet(path).head(1).collect().to_dicts()[0]


def decode_png(data):
    with Image.open(BytesIO(data)) as image:
        return np.array(image)


image = decode_png(row["image"])  # HWC RGB uint8, shape (256, 256, 3)
instances = decode_png(row["inst"])  # HW uint16, original instance numbers
classes = decode_png(row["type"])  # HW uint8, original class numbers 0–7

Plot examples

Continue from the loading example to show three training patches, including one with a marked mitosis.

import matplotlib.pyplot as plt
from matplotlib.patches import Patch
from skimage.segmentation import find_boundaries

class_names = [
    "Background",
    "Neutrophil",
    "Epithelial",
    "Lymphocyte",
    "Plasma cell",
    "Eosinophil",
    "Connective tissue cell",
    "Mitosis",
]
colors = np.array(
    [
        [0, 0, 0],
        [255, 120, 20],
        [245, 80, 65],
        [60, 170, 255],
        [170, 100, 255],
        [255, 225, 40],
        [40, 220, 140],
        [255, 60, 210],
    ],
    dtype=np.uint8,
)
indices = [2, 0, 1]
samples = (
    pl.scan_parquet(path)
    .filter(pl.col("source_index").is_in(indices))
    .sort("source_index")
    .collect()
    .to_dicts()
)
fig, axes = plt.subplots(3, 3, figsize=(10, 10))
for row_axes, sample in zip(axes, samples):
    rgb = decode_png(sample["image"])
    inst = decode_png(sample["inst"])
    types = decode_png(sample["type"])
    overlay = rgb.astype(np.float32)
    foreground = inst > 0
    overlay[foreground] = 0.55 * overlay[foreground] + 0.45 * colors[types[foreground]]
    boundaries = find_boundaries(inst, mode="inner")
    overlay[boundaries] = colors[types[boundaries]]
    for ax, pixels in zip(row_axes, [rgb, overlay.astype(np.uint8), colors[types]]):
        ax.imshow(pixels, interpolation="nearest")
        ax.axis("off")
    row_axes[0].text(0, 1.02, sample["id"], transform=row_axes[0].transAxes, fontsize=8)
for ax, title in zip(axes[0], ["H&E image", "Instance outlines by class", "Classes"]):
    ax.set_title(title, pad=24)
legend = [Patch(color=colors[i] / 255, label=class_names[i]) for i in range(1, 8)]
fig.legend(handles=legend, loc="lower center", ncol=4, fontsize=8)
fig.subplots_adjust(
    left=0.01, right=0.99, bottom=0.08, top=0.93, hspace=0.12, wspace=0.03
)
plt.show()

For two mask-only examples on a black background:

from skimage.color import label2rgb

mask_fig, mask_axes = plt.subplots(2, 2, figsize=(7, 7))
instance_colors = [(1, 0.3, 0.3), (0.2, 1, 0.5), (0.3, 0.7, 1), (1, 0.8, 0.2)]
for row_axes, sample in zip(mask_axes, [samples[0], samples[-1]]):
    inst = decode_png(sample["inst"])
    types = decode_png(sample["type"])
    row_axes[0].imshow(
        label2rgb(inst, colors=instance_colors, bg_label=0, bg_color=(0, 0, 0)),
        interpolation="nearest",
    )
    row_axes[1].imshow(colors[types], interpolation="nearest")
    row_axes[0].text(0, 1.03, sample["id"], transform=row_axes[0].transAxes, fontsize=8)
    for ax in row_axes:
        ax.axis("off")
for ax, title in zip(mask_axes[0], ["Individual instances", "Classes"]):
    ax.set_title(title, pad=24)
mask_fig.subplots_adjust(
    left=0.01, right=0.99, bottom=0.01, top=0.91, hspace=0.14, wspace=0.03
)
plt.show()

Instance colours distinguish nearby objects; they do not identify cell types. The class colours use the legend above.

Preparation and validation

The images and masks were packed into Parquet files with their pixel values, object numbers and class labels unchanged. Every saved image and mask was compared with the source.

Source details and checks are recorded in source-record.json, class-mapping-source.json, manifest.json and split-audit.json.

Limitations

• Some visible nuclei are unmarked, and some labels are incorrect.
• Training image lizard_mitosis_train_01186 and validation image lizard_mitosis_val_00343 share a 32-pixel-wide strip. Other overlaps may exist. See split-audit.json for details.
• Lizard includes images from PanNuke and other public colon datasets. Check for shared images before training on one collection and evaluating on another.
• The original Lizard paper reports an image scale of about 0.5 µm per pixel at 20×. Scale measurements for individual images are unavailable.

License and citation

The data use CC BY-NC-SA 4.0. For noncommercial reuse, credit the creators, link to the source and license, keep the supplied notices, and describe your changes. Share adapted data under the same license. Full terms are in LICENSE; NOTICE.md describes the changes in this version.

Credit both the HoVer-NeXt dataset creators and the original Lizard creators:

Baumann, E.; Dislich, B.; Rumberger, J. L.; Nagtegaal, I. D.; Martínez Rodríguez, M.; Zlobec, I. (2024). HoVer-NeXt: A Fast Nuclei Segmentation and Classification Pipeline for Next Generation Histopathology — Datasets (Version 2). Zenodo. doi:10.5281/zenodo.11657620.

Baumann et al. (2024). HoVer-NeXt: A Fast Nuclei Segmentation and Classification Pipeline for Next Generation Histopathology. Medical Imaging with Deep Learning. Paper.

Graham, S.; Jahanifar, M.; Azam, A.; et al. (2021). Lizard: A Large-Scale Dataset for Colonic Nuclear Instance Segmentation and Classification. ICCV Workshops, pp. 684–693. Paper.