TriDoRNet: Reconstruction of Standard-Dose Pet from Low-Dose Pet in Triple (Projection, Image, and Frequency) Domains

Reconstructing high-quality positron emission tomography (PET) from low-dose PET is an alternative way to reduce radiation hazard in PET imaging. As a PET image can be represented in multiple domains, each of which has different emphasized information, taking multiple domains into account could result in better reconstruction of standard-dose PET (SPET) from low-dose PET (LPET). Thus, different from previous studies on a single domain, in this paper, we fully consider the advantages of multi-domain image representation and propose the triple-domain reconstruction network (TriDoRNet) to reconstruct SPET image from LPET sinogram in the projection, image, and frequency domains. Specifically, a denoising network and a reconstruction network are coupled sequentially, where the former denoises the LPET sinogram in the projection domain, while the latter reconstructs the SPET image transferred from denoised sinogram in the image and frequency domains. The respective loss functions are further designed to supervise the training of TriDoRNet in three domains. Extensive experiments conducted on a real PET dataset demonstrate our proposed approach can reconstruct SPET images with the least noise and also the richest structural details compared to the state-of-the-art methods.