PDDUNet: A primal-dual deep unrolling network for CT metal artifact reduction.

Abstract

In computed tomography (CT) imaging, severe artifacts caused by metallic implants within patients degrade image quality and compromise clinical diagnostic accuracy. Currently, deep unrolling-based methods have demonstrated strong capability in metal artifact reduction (MAR). However, these methods primarily focus only on the unfoldment of data fidelity term, limiting their effectiveness. To overcome this issue, both the data fidelity term and the regularization term were unrolled to construct a Primal-Dual Deep Unrolling Network (PDDUNet) for MAR. First, a unified optimization model formulated for MAR consists of two data fidelity terms and a total variation regularization, where the prior image generated by a Prior-net constitutes one data fidelity term to improve reconstruction performance. Then, the Chambolle-Pock algorithm was employed to solve this model, yielding a single-loop iterative algorithm suitable for neural network unrolling. Finally, each computational step in this iterative algorithm is replaced by a simple neural network, constructing a deep unrolling network for MAR tasks. Extensive experiments were conducted on the synthetic and clinical datasets. Experimental results show that our proposed method achieves superior performance and generalization over previous MAR methods. It reveals that our proposed method offers a new perspective for MAR tasks.