csmp-hub/schurch-crc

Schürch CRC

H&E images of colorectal cancer tissue, with cell-region instance masks and cell-type labels from paired CODEX imaging. The original study examined cellular neighbourhoods at the colorectal cancer invasive front. Its typed cohort contains 140 tissue cores from 35 patients.

The data is released under CC BY 4.0. Please credit the original study and SPORA, link the license, and describe changes when sharing adapted versions. The original license and citation are included.

Images and annotations

Images retain their native RGB values and resolution of approximately 0.37744 µm per pixel. Each crop is 512 × 512 pixels.

The instance masks contain the original cell-region IDs. These regions were generated from DRAQ5 nuclear staining using the original study's segmentation method. Their extent can be larger than the visible nucleus. The masks are machine-generated annotations and may miss objects or have imperfect boundaries.

Cell types come from the original per-cell table. Labels include epithelial tumour cells, immune populations, stromal populations, and the source's ambiguous and artifact categories. The class file gives the original names and codes. Background is zero; source Unknown uses 255. classes.json lists the codes. Type masks are built by matching each instance ID to its original cell-table label. Instance IDs are unique within each core, so use core_id together with the instance ID to identify a cell across crops.

Instance colours distinguish cell regions. Cell-type colours follow the legend. Black is background.

The two examples show H&E, boundaries over H&E, and masks on their own. Mask colours distinguish instances, not cell types. These examples were judged usable during review; incomplete masks and small shifts can still be present.

Cell-type instance counts

These counts include only cells present in the crop masks. Each source cell is counted once using its slide and original instance ID, even when crops overlap. Partial cells at crop edges are included. Counts include all quality flags; Unknown and source quality labels remain separate from biological classes.

Original class	Train	Test	Unique total	Crop appearances
granulocytes	16,697	5,318	22,015	84,772
vasculature	8,498	3,153	11,651	45,887
CD4+ T cells CD45RO+	12,979	3,564	16,543	63,952
tumor cells	36,176	10,741	46,917	179,293
stroma	14,923	5,100	20,023	78,013
CD68+CD163+ macrophages	28,956	10,421	39,377	153,637
adipocytes	1,395	408	1,803	6,664
plasma cells	3,828	4,628	8,456	33,472
CD8+ T cells	13,659	2,927	16,586	65,307
dirt	6,035	1,171	7,206	28,248
Tregs	2,152	634	2,786	10,906
CD4+ T cells	1,962	317	2,279	8,364
CD11c+ DCs	362	34	396	1,561
B cells	11,335	1,555	12,890	54,018
CD11b+CD68+ macrophages	1,332	163	1,495	5,957
smooth muscle	23,494	4,186	27,680	106,185
undefined	5,171	1,329	6,500	25,171
tumor cells / immune cells	508	1,278	1,786	7,486
immune cells / vasculature	1,185	964	2,149	8,436
immune cells	2,078	1,038	3,116	11,941
NK cells	311	12	323	1,281
nerves	520	136	656	2,430
CD68+ macrophages GzmB+	133	48	181	686
CD68+ macrophages	1,749	353	2,102	8,288
lymphatics	306	22	328	1,298
CD11b+ monocytes	788	23	811	2,892
CD4+ T cells GATA3+	62	5	67	256
CD163+ macrophages	30	8	38	166
CD3+ T cells	184	4	188	744
Unknown	1,085	269	1,354	4,987
Total	197,893	59,809	257,702	1,002,298

instance-counts.csv has one row per crop in train.parquet or test.parquet. Match id to the crop row; source_file names its Parquet file. It records total_instances and one class:<original name> column per class, including zero counts. It also includes the split, slide, patient, crop position, review status and skip_training flag. Background is not counted. Within a crop, each instance ID counts once. Summing rows counts a cell again wherever it appears in overlapping crops; the unique totals above avoid this.

Splits

The original patient splits are retained: 27 training patients and eight test patients. A patient never appears in both splits. Crops overlap; core IDs and crop coordinates allow these relationships to be tracked.

Split	Crops	Tissue cores	Patients
Train	3,780	108	27
Test	1,120	32	8
Total	4,900	140	35

Each 1,440 × 1,920-pixel core contributes 35 crops: five rows and seven columns. Crops start 256 pixels apart, with the final row and column moved to the edge to cover the full core without padding. The crop origins are x = 0, 256, 512, 768, 1,024, 1,280, 1,408 and y = 0, 256, 512, 768, 928. This gives 140 × 35 = 4,900 crops. The masks contain 257,702 distinct source cells; cells in overlapping crops appear more than once.

Keep crops from the same patient together when making further splits.

Annotation quality

All crops are retained. A review of 60 selected training crops found 47 usable and 13 unsuitable crops. Another 118 crops overlap rejected areas. This selected sample does not measure the failure rate across the dataset.

skip_training is true for these 131 affected crops. Filtering it out leaves 3,649 training crops. training-index.parquet lists their IDs and coordinates. The original test patients are retained.

quality_status records usable, exclude, or unreviewed. human_training_approved is true, false, or null respectively. The 4,840 unreviewed crops include every test crop. A false skip_training flag does not mean a crop was reviewed. Usable crops may still contain missed cells and small image-to-mask shifts, and review does not verify every cell-type label.

ignored_regions lists rejected areas within each crop as x, y, width, height in pixels, with the top-left corner at (0, 0). The right and bottom edges are excluded from the rectangle. No image or mask pixels have been removed.

Files and columns

train.parquet and test.parquet each contain one row per crop.

Column	Contents
id	Unique crop ID
image	RGB image as PNG bytes; uint8, height × width × 3
inst	Cell-region instance mask as PNG bytes; uint16, height × width; background 0
type	Cell-type mask as PNG bytes; uint8, height × width; background 0, Unknown 255
sem	Null
core_id, patient_id, source_split	Source identity and split
x, y, mpp	Crop origin in the source core and micrometres per pixel
partial_instance_ids	IDs of instances that extend beyond this crop
quality_status, human_training_approved	Review result; approval is null for unreviewed crops
skip_training, ignored_regions	Flag for skipping known affected crops and their rejected areas
source_license, source_mask_kind	Source license and annotation description

cell-annotations.parquet contains the source per-cell labels. crop-instances.parquet lists the instances present in each crop, their pixel counts and whether they extend beyond it. Use these tables to avoid counting overlapping copies as distinct cells.

Loading and viewing examples

from io import BytesIO
import json

import matplotlib.pyplot as plt
from matplotlib.patches import Patch
import numpy as np
import polars as pl
from PIL import Image
from huggingface_hub import hf_hub_download
from skimage.segmentation import find_boundaries


def read_png(data: bytes) -> np.ndarray:
    with Image.open(BytesIO(data)) as image:
        return np.array(image)


repo = "csmp-hub/schurch-crc"
path = hf_hub_download(repo_id=repo, filename="train.parquet", repo_type="dataset")
classes_path = hf_hub_download(repo_id=repo, filename="classes.json", repo_type="dataset")
with open(classes_path) as stream:
    classes = json.load(stream)
classes.setdefault("Unknown", 255)

palette = np.vstack((plt.get_cmap("tab20").colors, plt.get_cmap("Set3").colors))
palette = palette[np.ptp(palette, axis=1) > 0.1]
colours = np.zeros((256, 3))
greys = {"unknown": 0.35, "undefined": 0.55, "dirt": 0.7, "excluded": 0.45, "low": 0.8}
for name, code in classes.items():
    if code:
        colours[code] = greys.get(name.lower(), palette[(code - 1) % len(palette)])

rows = pl.read_parquet(path, n_rows=2)
fig, axes = plt.subplots(2, 4, figsize=(17, 10))
present = set()
for row, panels in zip(rows.iter_rows(named=True), axes):
    image = read_png(row["image"])
    instances = read_png(row["inst"])
    cell_types = read_png(row["type"])
    present.update(np.unique(cell_types).tolist())
    overlay = image.copy()
    overlay[find_boundaries(instances, mode="inner")] = (0, 255, 120)
    masks = plt.get_cmap("turbo")((instances.astype(np.uint32) % 251 * 37 % 251) / 250)[..., :3]
    masks[instances == 0] = 0
    for panel, values, title in zip(
        panels,
        (image, overlay, masks, colours[cell_types]),
        ("H&E", "Boundaries", "Instances", "Cell types"),
    ):
        panel.imshow(values)
        panel.set_title(title)
        panel.axis("off")
handles = [Patch(facecolor=colours[code], label=name)
           for name, code in classes.items() if code and code in present]
fig.legend(handles=handles, loc="lower center", ncol=4, fontsize=8)
fig.subplots_adjust(left=0.01, right=0.99, top=0.97, bottom=0.13, wspace=0.04, hspace=0.1)
plt.show()

This loads a training crop while skipping the 131 affected crops. The filter still includes unreviewed crops. Remove the filter to read all crops.

Sources

• Original Schürch et al. study
• SPORA collection
• CC BY 4.0