Deep Learning-assisted simultaneous MRI-based Attenuation Correction and Full-Dose Synthesis from Non-Attenuated Low-Dose PET Images

Abstract

The trend to optimize the radiation dose delivered to patients from medical examinations is still relevant and. Dose reduction in PET/CT examinations from both internal and external exposure remains a hot research topic. The goal of this study is to reduce the injected radiotracer and ultimately eliminate the CT component using deep learning techniques. This is achieved by predicting the full-dose attenuation corrected (FD-AC) image from low-dose non-attenuation corrected (LD-NAC) images for three common brain imaging radiotracers, namely 18F-FDG, 18F-Flortaucipir and 18F-Flutemetamol. We included 100 FD-AC brain PET/CT images for each tracer. Only 5% of the events from the FD scan were randomly selected to simulate a realistic LD scan. We used a modified cycle-consistent generative adversarial network (CycleGAN) model to predict the FD-AC (PAC) image from LD-NAC image by considering two approaches (i) using only LD-NAC image as input to the network (PAC1) and (ii) using both LD-NAC and MRI images as input (PAC2). Quantitative metrics, including peak signal-to-noise ratio (PSNR), structural similarity index metric (SSIM), and root mean square error (RMSE) were used for performance assessment. Correlation coefficients (R2) of 0.89, 0.79 and 0.81 for 18F-FDG, 18F-Flortaucipir, and 18F-Flutemetamol were achieved, respectively, for PAC1 approach. PAC2 approach achieved R2 significantly higher than PAC1 (0.96, 0.85, and 0.95 for 18F-FDG, 18F-Flortaucipir, and 18F-Flutemetamol, respectively). Overall, significantly superior performance was achieved by PAC2 approach for all three radiotracers (p<0.05).