Dictionary learning for medical image synthesis

Dictionary-based image synthesis can be viewed as converting the style of a given image to another desired style. These image synthesis methods rely on a database of patches that have been extracted from images of both the original style (source domain) and desired style (target domain). Dictionary learning approaches have two main components. The first consists in learning dictionaries from the patches of the source and target domains. The second consists in finding the sparse coefficients that enable combining the elements of the dictionaries to reconstruct the source image for a given subject, while at the same time using the same coefficients to generate an image of the target domain for the given subject. This chapter aims to give a theoretical introduction to sparse coding and dictionary learning and illustrate the use cases of example-based sparse dictionary matching techniques in medical image synthesis. The advantages of sparse representations and dictionary learning for medical image synthesis are covered, as well as the shortcomings of this technique compared with the current state-of-the-art.