A Pilot Study of Attenuation Correction for Static and Dynamic ¹³N-Ammonia Cardiac PET Using a Conditional Generative Adversarial Network

Attenuation correction (AC) of myocardial perfusion PET remains challenging in routine clinical practice due to the potential misalignment of PET and CT data, which results in image artifacts and inaccurate quantitative parameters in PET/CT. In this work, a 3D conditional generative adversarial network (cGAN) was used to directly generate attenuation-corrected PET (PET-AC) images from non-attenuation-corrected PET (PET-NAC) images. This work included 60 patients who underwent rest ¹³ N-ammonia PET/CT examinations. For each patient, one static PET and twenty-one dynamic PET frames were acquired with both AC and NAC. The network was firstly trained by 3D paired static ¹³ N-ammonia PET-NAC and CT-based AC PET (PET-CTAC) images in an end-to-end fashion (PET-DLAC). Then the pre-trained network was fine-tuned (FT) by 3D paired dynamic PET-NAC and PET-CTAC frames (PET-DLAC-FT). Five-fold cross-validation was adopted to train and test the network. Qualitative and quantitative analyses were performed by voxel-wise analysis on the standard intake values (SUV), e.g., RMSE, PSNR, SSIM, and joint histogram analysis versus the PET-CTAC images as reference. Furthermore, quantitative kinetic parameters and time activity curves of the dynamic PET were performed to evaluate the PET-DLAC-FT method. Qualitative and quantitative results show the proposed PET-DLAC and PET-DLAC-FT methods reduce attenuation artifacts significantly and achieve comparable performances with clinical CTAC. Transfer learning-based AC is feasible for dynamic cardiac PET.