Research On Low-dose CT Image Noise Suppression Based On Regional Content Perception Low-rank Modeling

X-ray computed tomography(CT)is widely used in clinical diagnosis because of the advantages in high spatial resolution,short scanning time,and low cost.However,high ionizing radiation dose from CT scans would rise the risk of leukemia and tumors for patients,especially for the sensitive groups such as women and children.Low-dose CT(LDCT)technique can effectively reduce the ionizing radiation dose to patients,however the corresponding reconstructed image would suffer from severe noise and artifacts,which thus increase the difficulty of image analysis,and affect the accuracy of clinical diagnosis.In recent years,the low-rank model has demonstrated its effectiveness in low level image processing,including image denoising tasks.Among them,the nuclear norm minimization(NNM)method has been widely used because it can transform the traditional nonconvex low rank constraint into a convex norm optimization problem through constraining the singular value of the objective matrix.However,the traditional NNM method conducts the constraint for all singular values of the target matrix with same intensity,thus can not describe the non-stationary noise characteristics in LDCT images.To this end,this study carried out the research of LDCT images noise reduction based on the non-stationary noise characteristics in LDCT images through the regional-content-aware low rank modeling.The innovation of this study can be summarized as two fold:(1)On the basis of traditional NNM method,this study proposed a novel regional-content-aware weighted nuclear norm minimization algorithm(rwNNM)for LDCT image denoising.The proposed algorithm can adaptively set the weight of the NNM according to the local noise level as well as the magnitude of singular values.Therefore it can effectively improve the local information expression ability of nuclear norm,and thus improve the ability of image texture detail preservation.In addition,to estimate the local noise of LDCT image,this study proposed an improved noise estimation method by using the relationship between singular value decomposition of matrix and principal component analysis,which greatly improved the calculation efficiency.Furthermore,for the block matching operation in NNM model,this paper employed a new block matching strategy based on local statistical characteristics,which improves the accuracy of block matching.(2)This study proposed a novel LDCT image noise reduction method through regional-content-aware low-rank model and learned sparsifying transform learning model(reLRSTL).The proposed method combines the low-rank approximation and learned sparsifyingtransform,and thus naturally encodes the correlations in non-local similar patches and the intrinsic local transform sparsity in CT images.This study also developed an algorithm to improve the performance of LDCT image noise reduction using the reLRSTL model.An alternating optimization method was developed for the algorithm that alternates between the reLR approximation,the sparsifiying transform learning and the image update.Each step of the proposed algorithm has closed-form solution.