Artifact Removal in Sparse-Angle CT Based on Feature Fusion Residual Network

Abstract

To reduce the radiation dose from computed tomography (CT) scans and obtain high-quality images, various methods based on deep learning have been proposed for artifact removal in sparse CT. In this article, a new widely applicable method for artifact removal from sparse-angle CT images is proposed. We propose a feature fusion residual network (FFRN), which achieves excellent performance in removing artifacts from different anatomical regions of sparse angle CT images. In the FFRN, the residual skip dense block (RSDB) is introduced in the shallow layer to adequately utilize the feature information from the Conv layer of the residual block (RB). An improved RB is used in the FFRN deep layer to reduce the network complexity and training difficulty. The RSDB implements local feature fusion by skipping connections to enhance feature extraction. The use of a $3\times 3$ convolution kernel in the RSDB and improved RB achieved better performance compared with a $1\times 1$ convolution kernel. Weight normalization (WN) was used instead of batch normalization (BN) to improve the accuracy of the deep network. We use the original Hounsfield (HU) values of CT images for learning. The result obtained is comparable to the label image. In addition, it was verified that artifacts in the sparse-angle CT images can be better predicted without changing the image size.