Low-Dose Pet Image Reconstruction Using Self-Supervised Deep Kernel Representations Learned from MR

Positron emission tomography is a modality for imaging physiological process in vivo. Image quality depends on the level of ionising radiation exposure, motivating PET image processing methods which recover standard-dose image quality from low-dose signal. This work presents DK-ADMM; a self-supervised deep learning approach for low-dose PET image reconstruction. Our method explicitly models deep latent space features using a kernel function derived from coregistered MR images for improved noise control and anatomical consistency. We compare our results with baseline methods and demonstrate improved performance, and generalisability across patients without subject-specific fine tuning.