NCT-CRC-HE-100K: histological images of colorectal cancer and healthy tissue
Kather, Halama, Marx (NCT Heidelberg)
100,000 stained tissue patches, labeled into nine tissue types, for training cancer classifiers.
A set of 100,000 non-overlapping 224x224 pixel patches from H&E-stained histology slides of human colorectal cancer and normal tissue, color-normalized. Each is labeled with one of nine tissue classes such as tumor epithelium, stroma, lymphocytes, and mucus. It is a common benchmark for digital pathology models.
Access
- Cost
- FreeFree download from Zenodo; no account needed.
- License
- Open licenseCC BY 4.0
- Sign-up
- No sign-up
- How you get it
- Bulk downloadZIP
What's in it
- One row is
- One image patch with a tissue-class label
- Covers
- Patches from patients at NCT Biobank and UMM pathology archive (Germany)
- Years
- 2018 (one-time release)
- Updated
- One-time release
- Size
- 100,000 patches
Quality
Clean, widely used benchmark with clear labels; a static deposit from 2018.
- Patches come from few sites and patients, limiting generalization
- Class labels were applied at the patch level, not by multiple pathologists
Connects to
- Collection of textures in colorectal cancer histology: Smaller sibling set from the same group with eight tissue classes
- The Cancer Genome Atlas: Whole-slide images that models trained here can be tested on
- MedMNIST+: Includes a downsampled pathology task derived from this set
Questions it could answer
- How well do models trained on these patches transfer to whole slides from other hospitals?
- Which tissue classes are most often confused?