Providing computing solutions for exascale challenges
A deep learning platform for digital histopathology.
The software is organized into three layers. Layer I imp lements the extraction of patches of dimensions 224x224 pixels from the gigapixel slides of breast lymph nodes tissue. Patches are random sampled from the slide, in which areas of tumour were annotated by a physician. Patches belonging to a tumorous region are assigned a ‘tumor’ label (a Boolean variable equals to True). The extracted data are stored in an intermediate dataset with the corresponding labels. Layer II loads the intermediate dataset of patches and labels and trains a state - of - the - art deep conv olutional network to classify the two patch types. Different models can be chosen by a configuration parameter. Layer III focuses on network robustness and interpretability.