LOW-DOSE PET IMAGE RESTORATION WITH 2D AND 3D NETWORK PRIOR LEARNING

Reducing the dose of positron emission tomography (PET) imaging is a hot research area for avoiding too much radiation exposure. However, low-dose imaging faces the challenges of different degradation factors such as noise and artifacts. To restore high-quality PET images, we propose a mixed 2D and 3D encoder-decoder network to draw the mapping prior between low-dose and normal-dose PET images under the generative adversarial network framework with Wasserstein distance (WGAN). The proposed method has been evaluated on the in vivo dataset, showing encouraging restoration performances when compared to other state-of-the-art methods.