ABSTRACT
This chapter presents the main principles behind the design and training of a modern deep neural network, starting with an overview of the main architectures of interest for the computer vision community in general, and for the medical domain in particular. Convolutional neural networks are specialized architectures designed to learn good feature representations for multidimensional array signals. Deep neural networks exploit the property that many natural signals are compositional hierarchies, in which higher–level features are obtained by composing lower–level ones. The chapter focuses on fundamental aspects such as hyper-parameter optimization and data augmentation, and provides an overview of topics which are at the forefront of deep learning research, such as transfer learning and domain adaptation. One of the fundamental issues in applying deep learning to medical imaging is solving the limited training size problem. Deep learning relies heavily on high quality annotated data, which is often limited by the cost and complexity of acquisition.
