Study on improving the quality of low-dose PET images of children based on generative adversarial networks
冯莉娟,马欢,鲁霞,司宇坤,周子昂,阚英,王巍,李楠,张辉,杨吉刚
DOI: https://doi.org/10.3760/cma.j.cn321828-20220705-00212
2022-01-01
Abstract:Objective:To investigate the value of generative adversarial networks-based PET image reconstruction in improving the quality of low-dose 18F-FDG PET images and lesion detection in pediatric patients. Methods:Retrospective analysis of 61 PET images of children (38 males, 23 females, age (4.0±3.5) years) who underwent 18F-FDG total-body PET/CT imaging in Beijing Friendship Hospital, Capital Medical University from August 2021 to December 2021 was performed. The low-dose images (30 s, 20 s, 10 s) of all children extracted by list mode were input into the generative adversarial networks for deep learning (DL) reconstruction to obtain the corresponding simulated standard full-dose images (DL-30 s, DL-20 s, DL-10 s). The semi-quantitative parameters of the liver blood pool and primary lesion of standard full-dose 120 s, 30 s, 20 s, 10 s, DL-30 s, DL-20 s, and DL-10 s images were measured. The target-to-background ratio (TBR), contrast-to-noise ratio (CNR), and CV were calculated. The 5-point Likert scale was used for subjective scoring of image quality, and the detective abilities for positive lesions of each groups were compared. The sensitivities and positive predictive values of positive lesions detection were calculated. Mann-Whitney U test and Kruskal-Wallis rank sum test and χ2 test were used for data analyses. Results:CNR of the 30 s, 20 s, and 10 s groups were lower than those of DL-30 s, DL-20 s, and DL-10 s groups, respectively ( z values: -3.58, -3.20, -3.65, all P<0.05). Score of DL-10 s group was significantly lower than those of 120 s, DL-30 s and DL-20 s groups (4(3, 4), 5(4, 5), 4(4, 5), 4(4, 5); H=97.70, P<0.001). There were no significant differences in TBR, CNR, CV, SUV max and SUV mean of lesions and liver blood pool in 120 s, DL-30 s, DL-20 s, and DL-10 s groups ( H values: 0.00-6.76, all P>0.05). The sensitivities of positive lesion detection in DL-30 s, DL-20 s, and DL-10 s groups were 97.83%(225/230), 96.96%(223/230), 95.65%(220/230), respectively, and the positive predictive values were 96.57%(225/233), 93.70%(223/238), 84.94%(220/259), respectively. The positive predictive value in DL-10 s group was lower than those in DL-30 s and DL-20 s groups ( χ2=23.51, P<0.001). There were more false-positive and false-negative lesions detected by DL-10 s group than those of DL-30 s and DL-20 s groups in different sites. Conclusion:Based on the generative adversarial networks, the image quality of DL-20 s group is high and can meet the clinical diagnostic requirements.