Generating PET images from low-dose data using a cycle PET reconstruction convolutional neural network

A high-quality PET image usually requires the injection of a sufficient amount of radiotracer, which increases the risk of radiation exposure to the examinee and medical staff. Therefore, performing PET scanning with a reduced injection dose has broad application in clinical practice. However, low-dose PET data can cause problems in the reconstruction process, such as increased noise and degraded image quality. To address these issues, we adopt the idea of a back-projection filter (BPF) based on deep learning methods to reconstruct images. First, the raw data is back-projected (BP) to generate a low-dose BP (LDBP) image, as the PET sinogram is quite large. Second, a cycle PET reconstruction convolutional neural network (CPR-CNN) is introduced to realize a reconstruction process from a LDBP image to a deep learning-based full-dose PET (DL-FDPET) image. CPR-CNN employs two reconstruction models to separately learn bidirectional mappings between pairs of input images to enhance the image reconstruction performance. Our proposed CPR-CNN achieves comparable results in maintaining resolution and denoising of DL-FDPET images.