Histology Images from Uniform Tumor Regions in TCGA Whole Slide Images

Daisuke Komura from The University of Tokyo created a collection of 1,608,060 H&E-stained histological image patches from various human cancers. The data was derived from 8,736 diagnostic slides from 7,175 patients in the TCGA dataset, with uniform tumor regions annotated by pathologists. Patches were cropped from these regions at six magnification levels and resized to 256 x 256 pixels.

Use Cases

• Training image classification models for cancer type identification based on histology patches.
• Developing segmentation models to identify uniform tumor regions in whole slide images.
• Conducting multi-scale analysis of tumor morphology using patches at six different magnification levels.
• Benchmarking feature extraction methods for histopathological image analysis.

Strengths

• Contains 1,608,060 image patches, providing a substantial volume for model training.
• Data is derived from 8,736 diagnostic slides across 32 solid cancer types from the TCGA legacy database.
• Regions were annotated by two trained pathologists, focusing on uniform tumor areas and avoiding noncancerous structures.
• Patches are provided at six standardized magnification levels, from 0.5 to 1.0 μm/pixel.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Last update date is unknown; freshness unverified.
• The dataset is focused on uniform tumor regions, which may not represent the full heterogeneity of tumor microenvironments.

Provenance

Source

The Cancer Genome Atlas (TCGA) legacy database via the GDC.

Collection Method

Whole Slide Images were downloaded, annotated by pathologists for uniform tumor regions, and processed into patches using a custom tool (keras-OpenSlideGenerator).

Time Range

Data was downloaded from the GDC legacy database between December 1, 2016 and June 19, 2017.