Deep-PVC: A Deep Learning Model for Synthesizing Full-Dose Partial Volume Corrected PET Images from Low-Dose Images

Abstract

Moderating the injected activity and/or reducing acquisition time to minimize potential radiation hazards and increase patient comfort are important trends in PET. This work aims to assess the performance of regular full-dose partial volume corrected (FD-PVC) image synthesis from fast/low-dose (LD) brain PET images using deep learning techniques for 18F-FDG, 18F-Flortaucipir, and 18F-Flutemetamol radiotracers. Clinical brain PET/CT studies of 100 patients for each radiotracer were employed. The 5% of the events were randomly selected from the FD list-mode PET data to simulate a realistic LD acquisition. A modified cycle-consistent generative adversarial network (CycleGAN) model was implemented to predict FD-PVC PET images. Quantitative analysis using established metrics, including the peak signal-to-noise ratio (PSNR), structural similarity index metric (SSIM), root mean square error (RMSE), and standardized uptake value (SUV) bias, were performed. The R2 values were 0.95, 0.90, and 0.93 for 18F-FDG, 18F-Flortaucipir, and 18F-Flutemetamol, respectively. PSNR and SSIM values of 33.62±2.35 and 31.81±3.7, 34.44±4.11 and 0.96±0.02, 0.96±0.03, and 0.95±0.04 were obtained for 18F-FDG, 18F-Flortaucipir, and 18F-Flutemetamol, respectively. The CycleGAN is able to generate PVC images similar to references from under sampled LD images.