A Comparative Study Of CNN And GAN Deep Learning Networks To Improve The Quality And Quantification Parameters Of Pet Short Frame Images

Objective:(1)The deep learning method was used to compare the difference of image quality and quantization parameters between the reconstructed images obtained from short frames and standard images,and to evaluate its clinical applicability.(2)By constructing different deep learning models,the effectiveness of each model in noise reduction and the accuracy of tumor lesions semi-quantitative parameters were compared.Methods:The PET/CT scan data of 311 patients from March 2020 to December 2021 for definite diagnosis were retrospectively analyzed,and were randomly divided into the training set(n=218)and the test set(n=93)according to a ratio of 7:3.The original list mode PET data of 15 seconds/bed(15s)and 30 seconds/bed(30s)were selected to simulate short frame and low count PET images,and 90 seconds/bed(90s)PET data were taken as the standard images reaching clinical diagnosis.Short frame low count images are used as input to predict full count images by P2 P model and UNet model.Two experienced PET/CT physicians used the 5-point method to score the PET image quality of P2P-15 s,P2P-30 s,3D UNet-15 s,3D UNet-30 s,15s PET,30 s PET and S-PET groups.A.K software was used to obtain SD and SNR of PET image background,SNR,CNR of tumor lesions,and semi-quantitative parameters SUVmax and SUVmean.Kappa test,χ2test,Kruskal Wallis H test and Wilcoxon sign-Rank test were used to analyze the data.Results:Image quality scores were highly consistent among all groups(Kappa=0.72,95%CI: 0.70-0.74,P<0.001),89 patients in the P2P-15 s group,237 patients in the P2P-30 s group,236 patients in the 3D Unet-15 s group,and 303 patients in the 3D Unet-30 s group.There were 1,130 and 303 visual image quality scores ≥3 in the 15 s PET group,30 s PET group and S-PET group,respectively,and the proportions of visual image quality scores were significantly different among different groups(χ2=1252.81,P<0.001).Quantitative analysis results showed that the two deep learning models decreased background SD and increased SNR.When 15 s PET images were used as input,P2 P and 3D UNet models had similar enhancement effect on SNR of tumor lesions,but 3D UNet could significantly increase CNR of tumor lesions(P<0.05).When 30 s PET images were used as input,there were no significant differences in SNR,CNR and SUVmax between the 3D UNet group and the S-PET group(all P>0.05).Conclusion:Both P2 P and 3D UNet can suppress image noise and improve image quality.While 3D UNet reduces the noise of tumor lesions,it can increase the CNR of tumor lesions.In addition,the quantitative parameters of the two tumor models are similar to S-PET,which can meet the requirements of clinical diagnosis,efficacy evaluation and omits study.This study provides the possibility to reduce the scanning time,which also predicts the improvement of patient comfort and patient circulation.