EMISTA-Based Quantitative PET Reconstruction.

Linlin Zhao,Huafeng Liu

DOI: https://doi.org/10.1007/978-3-030-87358-5_56

2021-01-01

Abstract:As an important tool for cancer diagnosis and brain function imaging, PET is widely used in clinical fields where quantitative accuracy takes an important role. Several methods have been suggested to improve the quantification performance of PET by introducing quantitative coefficients, taking advantages of the time-of-flight (TOF) information, or modeling the point spread function (PSF). However, some physical effects such as photon attenuation limit the quantification potential, which should be focused on. In this paper, we proposed a novel method based on the combination of the expectation maximization (EM) and the learning ability of neural networks, achieving the quantification of PET reconstruction from raw sinogram data without attenuation correction (AC). The EM module was utilized to recover less-accurate PET images without the consideration of attenuation. And these rough images were finely adjusted by the network module which has strong nonlinearity to improve the quantitative performance. The proposed method was evaluated on simulated phantom data and compared to several existing reconstruction methods. It turns out that our method has great potential of high-quality PET image reconstruction.