CytoNuke

Overview

CytoNuke contains 83 H&E images of head and neck squamous cell carcinoma from 13 patients, with hand-drawn nucleus and whole-cell outlines. Each image is 256 × 256 pixels at a reported 0.5 µm per pixel.

Some cell boundaries could not be identified, so whole-cell masks are available for only part of the labelled nuclei. The images come from CPTAC-HNSCC; the dataset creators adjusted their stain colours.

License: CC BY 4.0, with credit to the dataset creators. The original CPTAC/TCIA images also carry CC BY 3.0; keep both notices.

Files and splits

Split	File	Images	Patients	Source nuclei	Source whole cells
train	train.parquet	43	9	2,344	1,547
val	val.parquet	11	2	360	259
test	test.parquet	29	2	1,287	801

All images from a patient stay in one split: nine patients in training, two in validation and two in testing. These groups were made for this version using seed 42; they differ from the paper's evaluation groups. Classes were not balanced between groups, and the small patient count limits comparisons. Exact patient lists are in manifest.json.

CytoNuke Dataset.zip keeps all original images and outlines. Source descriptions, licenses and attribution notices are included alongside the Parquet files.

Fields

Each row contains one image and its masks, stored as PNG bytes. Images are 256 × 256 pixels with three colour channels; masks are 256 × 256 arrays.

Field	Contents
id	Original image filename without .png
image	RGB colour image; uint8 values from 0 to 255
inst	Numbered nucleus mask; uint16
inst_cell	Numbered whole-cell mask, including the nucleus; uint16
type	Simple nucleus mask: 1 for a marked nucleus, 0 elsewhere; uint8
pairs	Suggested nucleus-to-cell matches, listed as nucleus_id and cell_id
patient	Patient ID taken from the original CPTAC filename
slide_name	Original slide name, including .svs
file_name, source_image_id	Original image filename and annotation-file image ID
mpp_x, mpp_y	Reported pixel size: 0.5 µm along both axes

Positive mask numbers are the original annotation IDs; zero means no marked object. A nucleus and its cell have different numbers. Keep the two masks separate: a whole cell includes its nucleus, so they overlap.

A zero in the cell mask can mean that a cell was not annotated. It is not proof that no cell is present. There is no tissue-region mask.

Labels

The source labels identify nuclei and cells, not different cell types:

Original class number	Name	Stored in
0	NUCLEUS	inst; type uses 1 for a marked nucleus
1	CELL	inst_cell

pairs lists 2,312 suggested nucleus-to-cell matches. A match is included only when a nucleus lies completely inside one cell and neither object overlaps another object in the other mask. The matches were calculated from the masks, not checked by a person. Partial, missing or ambiguous matches are left out; the objects themselves are kept.

Use these matches rather than assuming equal numbers identify a nucleus and its cell. The original archive keeps all 3,991 nucleus outlines and 2,607 cell outlines. There are no separate cytoplasm-only masks.

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/cytonuke", 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"])  # RGB uint8, HWC
nuclei = decode_png(row["inst"]).astype(np.uint16)  # HW instance labels
types = decode_png(row["type"])  # HW nucleus classes: 0 or 1
cells = decode_png(row["inst_cell"]).astype(np.uint16)  # HW whole-cell labels
pairs = row["pairs"]  # Inferred links between independent source IDs

Use val.parquet or test.parquet to read another split.

Plot examples

Continue from the loading example.

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

example_ids = [
    "C3L-00997-21.svsx35y38",
    "C3L-02651-21.svsx127y34",
    "C3L-04844-21.svsx178y33",
]
samples = (
    pl.scan_parquet(path)
    .filter(pl.col("id").is_in(example_ids))
    .select("id", "image", "inst", "inst_cell")
    .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"])
    nucleus_mask = decode_png(sample["inst"])
    cell_mask = decode_png(sample["inst_cell"])
    row_axes[0].imshow(rgb)
    for ax, mask, color in zip(
        row_axes[1:], [nucleus_mask, cell_mask], [(45, 140, 240), (245, 155, 45)]
    ):
        overlay = rgb.astype(np.float32).copy()
        overlay[mask > 0] = 0.55 * overlay[mask > 0] + 0.45 * np.array(color)
        overlay[find_boundaries(mask, mode="inner")] = color
        ax.imshow(overlay.astype(np.uint8))
    row_axes[0].text(0, 1.02, sample["id"], transform=row_axes[0].transAxes, fontsize=8)
    for ax in row_axes:
        ax.axis("off")
for ax, title in zip(
    axes[0], ["H&E image", "Nucleus instances", "Whole-cell instances"]
):
    ax.set_title(title, pad=20)
fig.subplots_adjust(
    left=0.02, right=0.98, bottom=0.02, top=0.94, hspace=0.18, wspace=0.03
)
plt.show()

Blue marks nuclei; orange marks whole cells.

Two examples without the image underneath. Black marks background; bright instance colours distinguish objects, not cell types. Colours do not indicate nucleus-to-cell matches.

from skimage.color import label2rgb

instance_colors = [
    (1, 0.85, 0.2),
    (0.2, 1, 1),
    (1, 0.3, 0.8),
    (0.5, 1, 0.2),
    (0.65, 0.55, 1),
]
mask_samples = samples
mask_samples = [mask_samples[0], mask_samples[-1]]
mask_layers = [("inst", "Nucleus instances"), ("inst_cell", "Whole-cell instances")]
mask_fig, mask_axes = plt.subplots(2, 2, figsize=(6.4, 8))
for row_axes, sample in zip(mask_axes, mask_samples):
    for column, (ax, (key, title)) in enumerate(zip(row_axes, mask_layers)):
        labels = decode_png(sample[key])
        pixels = label2rgb(
            labels, colors=instance_colors, bg_label=0, bg_color=(0, 0, 0)
        )
        ax.imshow(pixels, interpolation="nearest")
        ax.set_axis_off()
    label = sample["id"]
    row_axes[0].text(0, 1.02, label, transform=row_axes[0].transAxes, fontsize=8)
for ax, (_, title) in zip(mask_axes[0], mask_layers):
    ax.set_title(title, fontsize=10, pad=18)
mask_fig.subplots_adjust(
    left=0.01, right=0.99, bottom=0.02, top=0.94, hspace=0.16, wspace=0.03
)
plt.show()

Preparation and validation

Prepared from the CytoNuke release on Zenodo. The original outlines were turned into pixel masks, keeping the source annotation numbers. Image pixels, class names and image size are unchanged.

The added patient splits and suggested nucleus-to-cell matches are described above. Further conversion details and exact assignments are in manifest.json.

Limitations

Five tiny outlines—one nucleus and four cells—cover no pixels when converted to masks. The masks therefore contain 3,990 nuclei and 2,603 whole cells. The missing outlines remain in the original archive and are listed in manifest.json.

Only tumour nuclei and cells with clear boundaries were annotated. Unmarked areas may still contain cells, and some nuclei have no matching cell mask. There are no tissue-region masks.

The dataset is small, and these splits do not reproduce the paper's evaluation groups. The reported pixel size comes from the paper, rather than measurements stored with each image.

License and citation

The dataset uses CC BY 4.0. You may use, change and share it, including commercially, with credit, a license link and a description of your changes. LICENSE and source-README.txt keep the original terms and description.

Credit the CytoNuke creators: Johannes Raufeisen, Kunpeng Xie, Fabian Hörst, Till Braunschweig, Jianning Li, Jens Kleesiek, Rainer Röhrig, Jan Egger, Bastian Leibe, Frank Hölzle, Alexander Hermans and Behrus Puladi. Cite the dataset and paper:

• Raufeisen et al. CytoNuke Dataset: Towards reliable whole-cell segmentation in bright-field histological images. Dataset release, DOI: 10.5281/zenodo.15310323.
• Raufeisen et al. (2024). Cyto R-CNN and CytoNuke Dataset: Towards reliable whole-cell segmentation in bright-field histological images. Computer Methods and Programs in Biomedicine, 252, 108215. DOI: 10.1016/j.cmpb.2024.108215.

The images originate from CPTAC-HNSCC, provided by the National Cancer Institute Clinical Proteomic Tumor Analysis Consortium through The Cancer Imaging Archive (TCIA). The TCIA collection page lists its tissue-slide images under CC BY 3.0; retain this upstream attribution and license notice too:

• National Cancer Institute Clinical Proteomic Tumor Analysis Consortium (CPTAC). (2018). The Clinical Proteomic Tumor Analysis Consortium Head and Neck Squamous Cell Carcinoma Collection (CPTAC-HNSCC) (Version 15) [Data set]. The Cancer Imaging Archive. DOI: 10.7937/K9/TCIA.2018.UW45NH81.
• Clark et al. (2013). The Cancer Imaging Archive (TCIA): Maintaining and Operating a Public Information Repository. Journal of Digital Imaging, 26, 1045–1057. DOI: 10.1007/s10278-013-9622-7.

This version saves the outlines as pixel masks in Parquet files and adds patient splits and suggested nucleus-to-cell matches. The original archive and notices are included.