3D-Progressive Growing Generative Adversarial Networks for PET attenuation map generation

Abstract

The Positron Emission Tomography images visualize molecular activity in living tissue. To correctly recover the activity distribution, the patient anatomical information is required to compensate the attenuation effects. This attenuation map is normally obtained through an external imaging technique, such as the Computed Tomography. Recently it has been proposed to recover the attenuation map using image-based methods, such as convolutional neural networks. In this work the capacity of progressive growing Generative Adversarial Networks (GANs) is explored. Their accuracy is compared to traditional GANs from previous works. Despite of being successful in high quality 2D image generation and posses more stable training, their accuracy falls behind prior works, achieving a Mean Average Error of 132+-22 HU against 103+-18 HU from previous results.