TCGA-UT Dataset Documentation 

Quick Links

• Dataset on Hugging Face: For users interested in benchmarking foundation models or feature extractors, please visit TCGA-UT on Hugging Face
• Original Paper: Universal encoding of pan-cancer histology by deep texture representations

Dataset Overview

The TCGA-UT dataset is a large-scale collection of histopathological image patches from human cancer tissues. It contains 1,608,060 image patches extracted from hematoxylin & eosin (H&E) stained histological samples across 32 different types of solid cancers.

Key Features

• Size: Over 1.6 million image patches
• Resolution: All patches are standardized to 256 x 256 pixels
• Source: Derived from The Cancer Genome Atlas (TCGA) dataset
• Quality: Curated by trained pathologists
• Coverage: 32 different cancer types
• Patient Base: 7,175 patients from 8,736 diagnostic slides

Data Collection Process

1. Image Source: Whole Slide Images (WSI) were downloaded from the GDC legacy database between December 2016 and June 2017
2. Expert Annotation: Two trained pathologists selected at least three representative tumor regions per slide
3. Quality Control: 926 slides were removed due to various quality issues (poor staining, low resolution, focus problems, etc.)
4. Patch Extraction: 10 patches were randomly cropped at 6 different magnification levels from each annotated region

File Structure

Files are organized using the following format:

Resolution Key

• 0: 0.5 μm/pixel
• 1: 0.6 μm/pixel
• 2: 0.7 μm/pixel
• 3: 0.8 μm/pixel
• 4: 0.9 μm/pixel
• 5: 1.0 μm/pixel

License

• Non-Commercial Use: CC-BY-NC-SA 4.0
• Commercial Use: Please contact ishum-prm@m.u-tokyo.ac.jp for licensing

Citation

If you use this dataset in your research, please cite:

For Model Benchmarking

If you're interested in using this dataset for benchmarking foundation models or feature extractors, we recommend accessing the dataset through the Hugging Face Hub at dakomura/tcga-ut. The Hugging Face version provides:

• Predefined train/validation/test splits (both internal and external facility-based splits)
• Ready-to-use benchmarking framework for foundation models
• WebDataset format support for efficient data loading
• Example implementations for state-of-the-art model evaluation