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

摘要

Positron emission tomography (PET) is a widely used imaging modality in clinical environment and research studies. In clinical, long-scanned dynamic PET is usually adopted for an acceptable image quality with increasing radioactive risk. Therefore, low-dose PET (LdPET) has attracted much attention, recently model-based image reconstruction (MBIR) and convolutional neural network (CNN) have been demonstrated the efficiency of noise reduction. In this study, we employed a deep learning based Poisson negative log-likelihood problem to model the physical PET imaging. Following that, model was solved and split into three sub-problems by Alternating Direction Method of Multipliers (ADMM) algorithm with convergence guarantee. Then multiple ADMM iterations were mapped in a deep network architecture for optimizing each sub-problem, i.e., reconstruction, nonlinear and multiplier modules, dubbed as ADMM-net. For data acquisition, human brain data was acquired. In training phase, data pairs of Poisson down sampled sinogram and full-dose MLEM reconstructed image were used as network input and ground truth. The learnable parameters were optimized by minimizing the loss function of normalized mean square error (NMSE) using Adam. In order to demonstrate the generalization ability of proposed network, we trained network with 2D data pairs and tested it with sinograms. Moreover, different down sampling ratio were applied in training and testing. In results, our proposed ADMM-net outperformed the traditional EM reconstruction and existing algorithms for LdPET, such as nonlocal mean (NLM) and TV regularized methods in terms of MSE, as well as improved the computational speed.