100,000 non-overlapping image patches from hematoxylin & eosin stained histological slides of human colorectal cancer and normal tissue, each 224x224 pixels at 0.5 microns per pixel. The dataset includes nine manually annotated tissue classes and was created by Jakob Nikolas Kather from the National Center for Tumor Diseases using samples from the NCT Biobank and UMM pathology archive. A separate validation set of 7,180 image patches from 50 non-overlapping patients is also provided.
Use Cases
- Training image classification models to distinguish nine colorectal tissue classes based on the described categories (ADI, BACK, DEB, LYM, MUC, MUS, NORM, STR, TUM).
- Developing and validating models for automated cancer detection based on histological image patches.
- Benchmarking color normalization techniques using the provided non-normalized image set version.
- Researching computational pathology and feature extraction from standardized 224x224 pixel, 0.5 MPP images.
Strengths
- Contains 100,000 standardized image patches at a consistent resolution of 224x224 pixels and 0.5 microns per pixel.
- Includes nine manually annotated tissue classes relevant to colorectal cancer pathology.
- Provides a dedicated validation set of 7,180 images from 50 non-overlapping patients for robust model evaluation.
- Images are color-normalized using a cited, peer-reviewed method (Macenko's method).
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- Last update date is unknown; freshness unverified.
Provenance
- Source
- National Center for Tumor Diseases (NCT) Biobank, Heidelberg, and University Medical Center Mannheim (UMM) pathology archive.
- Collection Method
- Manually extracted from 86 H&E stained tissue slides from formalin-fixed paraffin-embedded (FFPE) samples.
- Time Range
- null
- Freshness
- Last updated date is unknown.
- Geography
- Heidelberg and Mannheim, Germany.