Background: Deep-learning-based denoising improves image quality and quantification accuracy for low count (LC) positron emission tomography (PET). Conventional deep-learning-based denoising methods only require single LC PET image input. This study aims to propose a deep-learning-based LC PET denoising method incorporating computed tomography (CT) priors to further reduce the dose level.

Methods: Fifty patients who underwent their routine whole-body 2-deoxy-2-[¹⁸F]fluoro-D-glucose (¹⁸F-FDG) PET/CT scans in March 2022 were retrospectively and non-consecutively recruited. For full count (FC) PET, patients were injected with 3.7 MBq/kg FDG and scanned for 5 bed positions with 2 min/bed. LC PET of 1/10 (LC-10) and 1/20 (LC-20) count levels of FC PET were obtained by randomly down-sampling the FC list mode data, which were then paired with FC PET for training U-Net (U-Net-1) and cGAN (cGAN-1). Networks incorporated CT images as input (U-Net-2 and cGAN-2) were also implemented. Quantitative analysis of physical and clinical indices was performed and statistically assessed with Wilcoxon sign-rank test with Bonferroni correction. 

Results: Mean square error and structural similarity index were the best for cGAN-2, followed by U-Net-2, cGAN-1 and U-Net-1. The errors of mean standardized uptake value (SUV) and maximum SUV were lowest for cGAN-2, followed by cGAN-1, U-Net-2 and U-Net-1. For cGAN-2, image quality and lesion detectability scores were 3.71±0.94 and 4.25±0.83 for LC-10, 3.57±0.79 and 3.61±1.21 for LC-20, while they were 3.49±0.92 and 4.42±0.08 for FC. Notably, some small lesions were “masked out” on cGAN/U-Net-1 but can be retrieved on cGAN/U-Net-2 denoised PET for LC-20.

Conclusions: Deep-learning-based LC PET denoising incorporating CT priors is more effective than conventional deep-learning-based denoising with single LC PET input, especially at lower dose levels.