Deep neural network architectures based on unrolling optimization algorithms have been widely adopted in deep-learning based image reconstruction applications in recent years. However, these architectures only mimic the iterative schemes of the corresponding algorithms, but lack rigorous convergence guarantee; and the learned network layers are difficult to interpret. These issues have hindered their applications in clinical use. In this paper, we develop an efficient Learned Descent Algorithm with a Line Search strategy (LDA-LS) and apply it to the nonconvex nonsmooth optimization problem of low-dose CT (LDCT) reconstruction. We show that LDA-LS yields a highly interpretable neural network architecture, where the regularization parameterized as multilayer perception is explicitly integrated into the iterative scheme and learned during the training process. We demonstrate that LDA-LS retains convergence guarantee as classical optimization algorithms, while achieving improved efficiency and accuracy in LDCT image reconstruction problems.
|