Low Ranked Decomposition-based Artifact Suppression Reconstruction Networks For Low Dose CT Images

Computed tomography(CT)scans the human body with X-ray beams to obtain cross-sectional image data inside the human body,which has the advantages of fast scanning time,clear images,and high resolution.It is widely used as an important clinical auxiliary diagnostic tool in disease diagnosis.High-dose ray radiation has potential harm to the human body.Low-dose CT(LDCT)is a direct and effective way to solve this problem by decreasing the radiation dose of CT.However,LDCT will affect the accuracy of clinical diagnosis because low radiation dose can reduce the imaging quality of CT images.Therefore,many scholars are committed to improving the imaging quality of CT images under the condition of the lowest possible radiation dose.In recent years,due to the performance improvement of computer hardware equipment and the rapid development of artificial intelligence,learning based methods have achieved great development in the fields of pattern recognition,natural image processing,and unmanned driving,but also brought new solutions to the problems of LDCT imaging technology.For CT noise reduction,some useful tissue structure and lesion information will be lost while removing noise artifacts.In this paper,we are committed to improving the imaging quality of CT images,using Convolutional Neural Network(CNN)as the basic architecture to suppress artifacts and noise in CT images from the perspective of image decomposition In order to improve the sensitivity of the network to different semantic information contained in different components,this paper uses a low-rank decomposition to separate noise from useful information of the image,and perform corresponding processing in different domains,thereby improving the suppression effect of artifact noise.Two LDCT image denoising networks based on image low-rank decomposition are mainly designed.The specific research contents are as follows:(1)Aiming at the excessive smoothing of CT denoising images,we use a rank-one decomposition reconstruction network to effectively preserve the tissue structure and pathological information that is beneficial to medical diagnosis.The network is able to project the content of CT images into two different domains: low-rank domain and residual domain.And then process the different semantic information in the two domains accordingly.In this way,the artifacts and noises are suppressed as much as possible,and the useful tissue structure information in the CT image is effectively preserved.The experimental results show that the proposed method improves PSNR and SSIM index values by0.9086 d B and 0.0356 respectively compared with RED-CNN,which has the best performance among the four comparison algorithms of BM3 D,Pix2pix,RED-CNN and HFSGAN.(2)Learning a single low-rank component of CT images may lead to insufficient learning of low-rank information.In order to solve this problem,we propose a Low-rank Decomposition Reconstruction Network(LDRN),which improves the extraction effect of low-rank components by introducing various components of SVD decomposition for supervised training.Moreover,the non-low-rank component processing channel is added to reduce the loss of useful information,thereby improving the performance of the noise reduction network.The network can be divided into two parts: content prediction and feature fusion.In the content prediction stage,the complementary strategy of low-rank information and noise learning is adopted,and a multi-low-rank component progressive fusion network is designed to supplement some low-rank information missing from a single low-rank component.At the same time,the noise prediction network is used to learn the non-low rank information.In the feature fusion stage,we integrate features containing different semantic information output from the content prediction network to improve the quality of the reconstructed image.The experimental results show that our method improves PSNR and SSIM index values by 1.0354 d B and 0.0362 respectively compared with RED-CNN,which has the best performance among the four comparison algorithms of BM3 D,Pix2pix,RED-CNN and HFSGAN.