A deep learning-based approach for direct PET attenuation correction using Wasserstein generative adversarial network

Abstract

Positron emission tomography (PET) in some clinical assistant diagnose demands attenuation correction (AC) and scatter correction (SC) to obtain high-quality imaging, leading to gaining more precise metabolic information in tissue or organs of patient. However, there still are some inevitable issues, such as imperceptible mismatching precision between PET and CT imaging, or plenty of ionizing radiation dose exposure in many after-treatment inspections. To cope with the abovementioned issues, we introduced a deep learning-based technique to achieve a direct attenuation correction for PET imaging in this article. Moreover, wasserstein generative adversarial networks and hybrid loss, including adversarial loss, L2 loss and gradient difference loss, were utilized to enforce the deep network model to synthesize PET images with much richer detail information. A comprehensive research was designed and carried out on a total of forty-five sets of PET images of lymphoma patients for the model training stage and test stage. Final performances analysis was totally based on our experimental outcomes, which demonstrated that the proposed algorithm has definitely improved the quality of PET imaging according to qualitative and quantitative study.