Artifact Removal In Sparse-Angle CT Based On Feature Fusion Residual Network

CT technology has played a decisive role in promoting medical diagnosis and treatment,and has achieved remarkable rapid development.Especially in the past ten years,it has been changing with each passing day,constantly introducing new ones,and now it has become an indispensable diagnostic method in the field of medical diagnosis.Since the radiation sources used in CT imaging are all X-rays,there are potential hazards to the human body.In order to reduce the radiation dose from computed tomography(CT)scans and obtain high-quality images,various deep learning-based methods have been proposed to removal artifacts in sparse-angle CT.First,this paper implements several different neural network structures,and compares the performance of these networks in sparse angle CT image artifact removal while keeping other parameters unchanged.Experimental results show that there are two network structures that have achieved good results in artifact removal,and good results have been achieved in terms of parameter amount,training difficulty,and artifact removal quality.In-depth analysis of the characteristics of the two network structures,pave the way for the subsequent proposal of better performance networks.Finally,this paper proposes a new method for removing artifacts from sparse angle CT images.This paper proposes a feature fusion residual network(FFRN),which achieves excellent performance in removing artifacts from different anatomical regions of sparse angle CT images.In FFRN,the residual skip dense block(RSDB)is introduced in the shallow layer to make full use of the intermediate feature information from different residual blocks.The RSDB implements local feature fusion by skipping connections to enhance feature extraction.The use of the 3×3convolution kernel in the RSDB achieved better performance compared with the 1 × 1 convolution kernel.Weight normalization(WN)was used instead of batch normalization(BN)to improve the accuracy of the deep network.We use the CT value of the CT image for learning,and the result obtained is equivalent to the label image.Experimental results show that without changing the image size in the network,the artifacts in sparse-angle CT images can be better eliminated.