A Deep ADMM-net for iterative low-count PET reconstruction

In clinical, researchers have shown an increasing effort in low-dose PET (LdPET) which reduces the risk of radiotracer while maintaining an acceptable image quality and is challenging in practice. To address this issue, regularized model-based image reconstruction (MBIR) is widely applied and the convolutional neural network (CNN) has been demonstrated the efficiency of noise reduction. In this study, we proposed a deep Alternating Direction Method of Multipliers (ADMM) network with residual CNNs. Human brain data pairs of Poisson sampled sinogram and full-dose MLEM reconstructed image was used as the input and ground truth in training phase respectively.Results showed that ADMM-TV-Net outperformed the traditional EM reconstruction and existing algorithms for LdPET, such as nonlocal mean (NLM) and TV in terms of normalized mean square error (NMSE) and reconstruction speed.