A 3D attention residual encoder–decoder least-square GAN for low-count PET denoising

Abstract

In this paper, to reduce patient scan times and maintain high image quality, we propose a 3D attention least-square (LS) generative adversarial network (GAN) to estimate positron emission tomography (PET) images with long scan times from short-scan-time images; this network is called 3D a-LSGAN. To explore the structural information between slices, a 3D network implementation is used. We take a low-count 3D PET image scanned for 75 s as the input and generate a high-count (HC) 3D PET image corresponding to an estimated scan time of 150 s. Specifically, a U-Net-like deep learning network is combined with a residual network and self-attention strategy to transfer the important information from the encoder part to the corresponding decoder part of the network. In addition, the mean square error (MSE) loss is added to the adversarial loss to form a new loss function that removes artifacts and yields high-quality PET images. The qualitative and quantitative experimental results show that the proposed 3D a-LSGAN method for low-count PET image noise reduction performs better than the state-of-the-art methods considered.