Kather Colorectal Histology

Overview

Kather Colorectal Histology is an eight-class tissue-texture classification dataset with 5,000 H&E-stained colorectal histology tiles. Each tile belongs to one tissue category.

• Train: 5,000 images, with 625 per class.

• Images: 150×150 RGB TIFF tiles, stored by the builder as RGB PNGs without resizing.

• Official splits: none; the builder exposes the entire tile release as train.

• Download: the 246 MiB tile archive, verified against Zenodo’s published MD5 checksum; the separate larger-image archive is not downloaded.

• License: CC BY 4.0; attribute the dataset and cite the original paper.

Data Structure

Accessing ds[i] returns a dictionary with these fields:

Key	Type	Description
image	PIL.Image.Image	150×150 RGB histology tile, decoded from PNG bytes
label	int	Tissue class index (0-7)
filename	str	Original tile filename, retaining its .tif extension
source_image_id	str	Source name embedded between the tile token and .tif_Row_ in the filename
slide_id	str	Numeric CRC-Prim-HE-NN filename prefix used to group related source images

For example, 10647_CRC-Prim-HE-10c.tif_Row_1_Col_1.tif yields source_image_id="CRC-Prim-HE-10c" and slide_id="CRC-Prim-HE-10". Source suffixes such as _003b and _copy are preserved in source_image_id but grouped under the numeric prefix in slide_id. These fields are derived from filenames, not independently verified slide or patient identifiers. No patient_id is provided.

Raw access with ds.with_format("raw") returns PNG image bytes rather than PIL images. Original TIFF filenames remain unchanged as metadata, and examples are ordered deterministically by archive path.

Classes

ds.features["label"].names lists class names in this fixed index order:

Label	Class Name	Description	Images
0	tumor	Tumor epithelium	625
1	stroma	Simple stroma	625
2	complex	Complex stroma	625
3	lympho	Lymphocytes	625
4	debris	Debris	625
5	mucosa	Normal mucosal glands	625
6	adipose	Adipose tissue	625
7	empty	Background	625

Usage Example

from stable_datasets.images import KatherColorectalHistology

# Downloads and prepares the tiles on first use; subsequent loads reuse the cache.
ds = KatherColorectalHistology(split="train")
print(len(ds))  # 5000
sample = ds[0]
class_names = ds.features["label"].names
print(class_names[sample["label"]])
print(sample["image"].mode, sample["image"].size)  # RGB (150, 150)
print(sample["source_image_id"], sample["slide_id"])

# Omitting split returns a StableDatasetDict containing only train.
ds_all = KatherColorectalHistology()
print(list(ds_all))  # ['train']

Requesting test or validation raises an error: those splits are not published, and the builder does not invent them.

Leakage-Aware Splitting

A random tile-level split can place related source images in both training and validation, making results optimistic. Grouping by slide_id keeps the numeric filename groups together, including letter-suffixed and copied source names. Grouping only by source_image_id is weaker because multiple source images can share the same numeric prefix.

Warning

slide_id is a filename-derived grouping, not a verified patient ID. Keeping these groups disjoint reduces leakage among related tiles but does not establish patient-disjoint evaluation. Patient-level splitting requires additional validated provenance that this builder does not provide.

For example, hold out 20% of the filename groups using GroupShuffleSplit:

from sklearn.model_selection import GroupShuffleSplit
from stable_datasets.images import KatherColorectalHistology

ds = KatherColorectalHistology(split="train")
# Read metadata without decoding every image.
slide_ids = [row["slide_id"] for row in ds.with_format("raw")]
splitter = GroupShuffleSplit(n_splits=1, test_size=0.2, random_state=42)
train_indices, validation_indices = next(
    splitter.split(range(len(ds)), groups=slide_ids)
)
train_ds = ds.select(train_indices)
validation_ds = ds.select(validation_indices)

train_groups = {slide_ids[index] for index in train_indices}
validation_groups = {slide_ids[index] for index in validation_indices}
assert train_groups.isdisjoint(validation_groups)

The hold-out fraction refers to groups, not tiles, and does not guarantee class balance or complete class coverage in each subset. Inspect label counts after splitting. These subsets are user-created, not official dataset splits. The benchmark harness’s default random 10% hold-out (seed 42) is tile-level; report that protocol explicitly rather than treating it as patient-level evaluation.

References

• Dataset record and license

• Original paper

• Malaria Cell Images: a complementary binary microscopy benchmark

Citation

@article{kather2016multi,
  title={Multi-class texture analysis in colorectal cancer histology},
  author={Kather, Jakob Nikolas and Weis, Cleo-Aron and Bianconi, Francesco
          and Melchers, Susanne M and Schad, Lothar R and Gaiser, Timo
          and Marx, Alexander and Z{\"o}llner, Frank Gerrit},
  journal={Scientific Reports},
  volume={6},
  pages={27988},
  year={2016},
  doi={10.1038/srep27988}
}