TCGA-UT: 31 Cancer Type Histology Images from Uniform Tumor Regions

TCGA-UT contains between 100,000 and 1,000,000 histology images extracted from uniform tumor regions within The Cancer Genome Atlas (TCGA) Whole Slide Images. Created by dakomura and updated in April 2025, it provides a benchmarking framework with predefined train, validation, and test splits for cancer classification.

Use Cases

• Classification of 31 cancer types using histology image patches
• Benchmarking pathology foundation models using the provided internal and external test splits
• Feature extraction from uniform tumor regions to analyze morphological differences across malignancies

Strengths

• Includes 31 distinct cancer type labels for multi-class classification
• Scale of 100,000 to 1,000,000 image records
• Provides predefined train/validation/test splits to ensure reproducible benchmarking

Limitations

• Focuses exclusively on uniform tumor regions, which may omit diagnostic features found in the tumor microenvironment or stroma
• Geographic bias inherent to the TCGA source population primarily based in the United States

Provenance

Source

The Cancer Genome Atlas (TCGA) via Zenodo

Collection Method

Extraction of uniform tumor regions from Whole Slide Images (WSI)

Freshness

Last updated April 2025.

Geography

United States