Whole-body PET Image Synthesis from Low-Dose Images Using Cycle-consistent Generative Adversarial Networks

This work sets out to investigate the performance of full-dose (FD) PET prediction from fast or low-dose (LD) whole-body (WB) PET scans using convolutional neural networks. One hundred patients who underwent WB PET/CT examinations were retrospectively used to develop LD to FD PET conversion models. The patients underwent two separate WB examinations lasting ~27 min (regular scan) and ~3 min (fast or LD scan) acquisition times. The fast (3 min) WB PET examinations are equivalent to 1/8^th of the full-dose PET scan. A residual neural network (ResNet) and a modified cycle-consistent generative adversarial network (CycleGAN) architecture were employed to model the LD to FD PET conversion. The quality of synthetic PET images produced by ResNet and CycleGAN models, referred to as RNET and CGAN, respectively, were evaluated by two nuclear medicine physicians using a five-point scoring. Quantitative metrics, including structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), mean square error (MSE) and standardized uptake value (SUV) bias were calculated within the left/right lung, brain, liver, and one hundred hot spots (malignant lesions) in PET images predicted using CGAN and RNET models. The physicians assigned scores of 3.88 and 4.92 (out of 5) to the CGAN-predicted FD PET images for the neck & trunk and brain regions, respectively. Considering the PSNR and SSIM metrics, the CGAN model exhibited superior performance with PSNR=39.08±3.56 and SSIM=0.97±0.02 compared to the RNET model with PSNR=34.91±1.50 and SSIM=0.93±0.04. Overall, the CGAN model outperformed the RNET model exhibiting lower SUV bias and higher image quality in the predicted FD PET images.