A Deep Recurrent Neural Network With FISTA Optimization for CT Metal Artifact Reduction

Metal Artifact Reduction (MAR) is one of the most challenging problems in Computed Tomography (CT) imaging. In CT imaging, metal implants in patients' bodies cause artifacts due to several factors, such as beam hardening effects, statistical property changes of the X-ray beams, and the shapes of metal implants. Although some promising results have been achieved by previously proposed model-based iterative reconstruction (IR) techniques, there is still much room for improvement. One of the problems is that the image prior models used in the IR techniques are too simple to capture the truly complex nature of the CT images. Recent advances in neural network deep learning (DL) can help address this problem and potentially improve MAR results significantly. In this work, we describe a novel DL-based technique for CT MAR. In this technique, we introduce a novel deep neural network based on an IR formulation and a convex optimization technique known as the FISTA (Fast Iterative Shrinkage-Thresholding Algorithm). The neural network, called the RNN-MAR, is an RNN (Recurrent Neural Network) composed of a set of proposed RFUs (Recurrent FISTA Units). While the structure of the RFU has some connections to the GRU (Gated Recurrent Unit), it is specifically designed for CT MAR. The RNN-MAR conducts dual-domain learning (image and sinogram) but can do this using only one objective function. Furthermore, unlike previous CT MAR techniques, the RNN-MAR does not use a binary metal trace. Instead, we used a novel real-valued sinogram domain confidence map, leading to smoother edges. Results from extensive experiments indicate that our RNN-MAR outperforms state-of-the-art DL MAR techniques in terms of the Root Mean Squared Error (RMSE), Peak Signal-to-Noise Ratio (PSNR), and Structure Similarity (SSIM) as well as in terms of visual appearance.