Evaluating the Effectiveness of Training a Generative Adversarial Network Through Transfer Learning for the Purpose of Completing CT Images

Abstract

This thesis develops an image completion neural network to mitigate computed tomography (CT) artifacts without the use of large clinical datasets or sinogram data. It aims to demonstrate that by using the developed technique, artifacts can be mitigated accurately by simply replacing a given artifact-marred region with context-appropriate anatomy. Such an approach could have implications for how artifacts are dealt within the context of radiotherapy treatment planning. The first chapter contains the preface and thesis outline. Chapters two through four outline the background knowledge that is required to understand the motivation for, and methods of, this work. Chapter five contains a research manuscript intended to address the need for a method which can correct artifacts on CT images used for radiotherapy treatment planning more effectively than current standard of care methods. This study employs a generative adversarial network (GAN)-based image completion network and uses transfer learning techniques in its approach.