Histopathology images for end-to-end AI, based on TCGA-BRCA

<p>These are histopathological images which are derived from the TCGA-BRCA breast cancer histology dataset at&nbsp;https://portal.gdc.cancer.gov/ (please check this website for the original data license). They can be used for end-to-end artificial intelligence (AI) workflows such as DeepMed (https://github.com/KatherLab/deepmed) which aim to predict high-level features directly from digital images with weakly supervised transfer learning. Here, we use two subsets of these digitized images:</p> <p>1) TCGA-BRCA-A2, these are all images from Walter Reed National Military Medical Center (tissue source site code A2, N=100 images) in the TCGA-BRCA database (tcga-brca-a2-deepmed-tiles.zip)</p> <p>2) TCGA-BRCA-E2, these are all images from&nbsp;Roswell Park Comprehensive Cancer Center (tissue source site code E2, N=90 images) in the TCGA-BRCA database (tcga-brca-e2-deepmed-tiles.zip)</p> <p>see also&nbsp;https://gdc.cancer.gov/resources-tcga-users/tcga-code-tables/tissue-source-site-codes&nbsp;</p> <p>The images were preprocessed according to the Aachen Protocol for Deep Learning Histopathology which is available at&nbsp;https://zenodo.org/record/3694994. Specifically, digital whole slide images (SVS format) of hematoxylin &amp; eosin (H&amp;E) stained slides were tessellated (without manual annotations) into tiles of 256x256 px edge length at 1 &micro;m/px. Then, images were color-normalized using the Macenko method as described before (https://www.nature.com/articles/s43018-020-0087-6) and saved as JPEG files. For the A2 cohort, an additional ZIP archive is provided in which only 100 random image tiles are saved for each patient (tcga-brca-a2-deepmed-tiles_100.zip). In addition, we provide a CLINI and a SLIDE table as defined in the &quot;Aachen Protocol&quot;. The CLINI table contains clinico-pathological data for all included patients and it is derived from clinical information on www.cbioportal.org as well as from Thorsson et al.&nbsp;(https://pubmed.ncbi.nlm.nih.gov/29628290/). We recommend to use the A2 dataset for training and the E2 dataset for testing. Please cite the relevant papers if you re-use this dataset, more information is available on www.kather.ai</p>