Positron emission tomography (PET) with various tracers plays a critical role in Parkinson's disease (PD) studies. However, the clinical application of PET with novel or less-available tracers is limited by factors such as high cost and short half-life. To address this limitation, we developed a cross-tracer PET synthesis model based on diffusion model to directly synthesize cross-tracer PET images from widely used18F-FDG PET images for PD studies. The network was initially trained on18F-FDG and11C-CFT datasets, and subsequently fine-tuned on18F-FDG and18F-DOPA datasets. After anonymization, synthetic images were mixed with real ones and underwent visual assessment by radiologists. Quantitative analyses were further performed through voxel-wise comparisons and regional error analysis between synthetic and real images. Visual assessment indicated that the synthetic11C-CFT and18F-DOPA PET images were comparable to real images in terms of quality, noise, and striatal conspicuity, with no significant differences observed in the error maps. Synthetic11C-CFT images achieved an average PSNR of 36. 22 and an SSIM of 0. 96, whereas18F-DOPA images yielded values of 29. 79 and 0. 95, respectively. Bland-Altman analysis indicated high consistency between synthetic and real images. The average standardized uptake value (SUV) bias was 0. 0130. 032 SUV for11C-CFT and -0. 1750. 149 SUV for18F-DOPA. The proposed cross-tracer PET synthesis model achieves comparable performance with11C-CFT or18F-DOPA PET imaging for PD studies.
Yi et al. (Thu,) studied this question.