Histology images from uniform tumor regions in TCGA Whole Slide Images (TCGA-UT)

<h1><strong>TCGA-UT Dataset Documentation&nbsp;</strong></h1> <h2>Quick Links</h2> <ul> <li><strong>Dataset on Hugging Face</strong>: For users interested in benchmarking foundation models or feature extractors, please visit <a href="https://huggingface.co/datasets/dakomura/tcga-ut">TCGA-UT on Hugging Face</a></li> <li><strong>Original Paper</strong>: <a href="https://doi.org/10.1016/j.celrep.2022.110424">Universal encoding of pan-cancer histology by deep texture representations</a></li> </ul> <p>&nbsp;</p> <h2>Dataset Overview</h2> <p>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 &amp; eosin (H&amp;E) stained histological samples across 32 different types of solid cancers.</p> <h3>Key Features</h3> <ul> <li><strong>Size</strong>: Over 1.6 million image patches</li> <li><strong>Resolution</strong>: All patches are standardized to 256 x 256 pixels</li> <li><strong>Source</strong>: Derived from The Cancer Genome Atlas (TCGA) dataset</li> <li><strong>Quality</strong>: Curated by trained pathologists</li> <li><strong>Coverage</strong>: 32 different cancer types</li> <li><strong>Patient Base</strong>: 7,175 patients from 8,736 diagnostic slides</li> </ul> <h2>Data Collection Process</h2> <ol> <li><strong>Image Source</strong>: Whole Slide Images (WSI) were downloaded from the GDC legacy database between December 2016 and June 2017</li> <li><strong>Expert Annotation</strong>: Two trained pathologists selected at least three representative tumor regions per slide</li> <li><strong>Quality Control</strong>: 926 slides were removed due to various quality issues (poor staining, low resolution, focus problems, etc.)</li> <li><strong>Patch Extraction</strong>: 10 patches were randomly cropped at 6 different magnification levels from each annotated region</li> </ol> <h2>File Structure</h2> <p>Files are organized using the following format:</p> <div> <div>&nbsp;</div> <div> <div><span>Copy</span></div> </div> <div> <div><code>[cancer_type]/[resolution]/[TCGA Barcode]/[region]-[number]-[pixel resolution].jpg</code></div> </div> </div> <h3>Resolution Key</h3> <ul> <li>0: 0.5 &mu;m/pixel</li> <li>1: 0.6 &mu;m/pixel</li> <li>2: 0.7 &mu;m/pixel</li> <li>3: 0.8 &mu;m/pixel</li> <li>4: 0.9 &mu;m/pixel</li> <li>5: 1.0 &mu;m/pixel</li> </ul> <h2>License</h2> <ul> <li><strong>Non-Commercial Use</strong>: CC-BY-NC-SA 4.0</li> <li><strong>Commercial Use</strong>: Please contact <a href="mailto:ishum-prm@m.u-tokyo.ac.jp">ishum-prm@m.u-tokyo.ac.jp</a> for licensing</li> </ul> <h2>Citation</h2> <p>If you use this dataset in your research, please cite:</p> <div> <div>&nbsp;</div> <div> <div><span>Copy</span></div> </div> <div> <div><code>Komura, D., et al. (2022). Universal encoding of pan-cancer histology by deep texture representations. Cell Reports 38, 110424. https://doi.org/10.1016/j.celrep.2022.110424</code></div> </div> </div> <h2>For Model Benchmarking</h2> <p>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 <a href="https://huggingface.co/datasets/dakomura/tcga-ut">dakomura/tcga-ut</a>. The Hugging Face version provides:</p> <ul> <li>Predefined train/validation/test splits (both internal and external facility-based splits)</li> <li>Ready-to-use benchmarking framework for foundation models</li> <li>WebDataset format support for efficient data loading</li> <li>Example implementations for state-of-the-art model evaluation</li> </ul> <p>&nbsp;</p>