An Iterative Algorithm for Unrolled Unsupervised PET Image Reconstruction

In recent years, the potential of convolutional neural networks (CNNs) for PET image reconstruction has been widely acknowledged. However, the effectiveness of CNNs relies on the availability of a substantial data set containing paired low-quality and high-quality reference PET images. Acquiring such a data set is often not feasible in clinical practice. The deep image prior (DIP) based on unsupervised learning can effectively address this problem by utilizing only prior information and measurement data of patients. In this paper, we propose an accelerated restart algorithm that integrates the DIP framework into the maximum-a-posteriori (MAP) PET image reconstruction. Our framework combines a CNN with the expectation maximization (EM) update step to ensure data consistency. The experimental results on both computer-simulated and clinical PET data sets demonstrate the effectiveness of our proposed method.