273 Non-invasive imaging biomarkers 3 Figure 2. Handcrafted and deep learning radiomic pipeline. A) In the handcrafted pipeline predefined features are extracted from a manually or (semi-) automatically defined volume of interest (VOI). Feature selection or dimension reduction is performed and these features are consecutively introduced in a statistical or machine learning model. B) Deep learning radiomics does not require VOI delineation, but processes the images in their raw form. The deep learning architecture consists of several hidden layers including convolutional and pooling layers, that extract increasingly complex features and perform feature selection and classification.
RkJQdWJsaXNoZXIy MTk4NDMw