Chapter 9 - Synthesis: novel radiography view synthesis

Abstract

Medical data synthesis refers to the process of generating unseen medical data as if it was clinically captured or produced. This chapter covers two forms of medical data syntheses, unconditional and conditional, and introduces design principles for these two forms. For unconditional synthesis, we focus on solutions with generative adversarial networks which have been widely adopted to address medical data synthesis problems. Conditional synthesis can be further categorized into homogeneous and heterogeneous domain synthesis. For homogeneous domain synthesis, we highlight its connection with image-to-image translation and show how to train deep image-to-image networks for the translation between two homogeneous domains. For heterogeneous domain synthesis, we note the heterogeneous nature of the source and target domains and focus on solutions that close such domain gaps. To exemplify the design principles, we provide a case study on novel radiography view synthesis. We propose a model called XraySyn that has specialized modules for the heterogeneous domain synthesis between X-ray and CT images and thus facilitates the generation of novel radiography views. Moreover, we introduce prior knowledge of CT imaging and X-ray projection to close the domain gap between CT and radiography. We show it not only improves the quality of radiography image synthesis but also can be leveraged to suppress bones from radiography.