Progressive Generative Adversarial Networks: Deep Learning in Head and Neck Cancer CT Images to Synthesized PET Images Generation for Hybrid PET/CT Application

Abstract

We proposed a progressive Generative Adversarial Networks (GAN) to generate synthesized positron emission tomography (PET) images from computed tomography (CT) images for hybrid PET/CT application. It is observed the PET images, generated by using the progressive training strategy, are better in image quality than the PET images generated by using GAN both in mean absolute error (MAE) and peak signal to noise ratio (PNSR) evaluation indicators, indicating that the proposed method is suitable for generating medical images to use in hybrid systems.