Research On Post-processing Method For Improving The Quality Of Low-dose X-CT Image

At present,X-ray computed tomography(X-CT)imaging technology has been a medical examination and diagnosis method commonly used in clinical practice.Meanwhile,the X-ray radiation exposure and the corresponding hazard have gained considerable attention in regular.The technology of low-dose X-CT(LDCT)is an effective approach to reduce radiation hazards.Of varied methods,the reduction of X-ray tube current is the most commonly used in clinical practice due to its simpleness and feasibility.However,due to the decrease of photon number and beam hardening,lots of steak artifacts and mottle noise appear in the image which is reconstructed by the traditional filter back projection(FBP)algorithm.The degraded image may affects the credibility of clinical diagnosis directly.Therefore,how to obtain high quality reconstructed image under the condition of LDCT scanning is the key of whether it can really serve for the society.In order to improve the quality of LDCT image,this paper studied in three aspects,including sparse representation theory,morphological component analysis and low-rank matrix approximation,and proposed three denoising algorithms for improving the quality of LDCT image.The main works are as follows:(1)An effective sparse representation denoising method based on the grouped dictionaries with adaptive atom sizes(GDwAAS)was proposed.In GDwAAS,a new feature detector is designed by making use of both photometrical and geometrical similarities in images.The former is captured by pixel values of image patches and the latter by steering kernel regression(SKR)coefficients.Then,according to the proposed new feature detector,image patches are clustered to several groups,and each group is used as a training sample set to train a dictionary.With such a procedure,image patches with similar characteristics in eachgroup can be sparsely represented with the corresponding learning dictionary.In order to further improve the adaptivity of dictionary,the patch groups are classified into three different genres,including flat group,edge group and texture group,then the atom size of each dictionary is selected according to the genre the group belongs to and the noise level.Experimental results show that the GDwAAS method achieves better denoising performance than related denoising algorithms,especially in fine structure preservation.(2)An two-step denoising method for LDCT image,which exploits the morphological component analysis(MCA)and non-local means(NLM)was proposed.In the first step,the MCA-based image separation is performed with the proposed dictionary to remove the steak artifacts.The key to successful artifacts removal is the dictionary.The dictionary is firstly established from the learning procedure from the preprocessed images,and then modified by using gradient activity measure.In the second step,the non-local means(NLM)algorithm is used to remove the mottle noise in the residual image.Experimental results from both phantom simulation and real clinical data demonstrate that the proposed method shows better performance in both noise/artifacts removal and structure preservation.(3)An effective image denoising algorithm,which is based on discriminative weighted nuclear norm minimization(D-WNNM),was proposed to improve LDCT image.In the D-WNNM method,local entropy of image is exploited,to discriminate streak artifacts from tissue structure,and to tune WNNM weight coefficients adaptively.Through this process,the steak artifacts and the tissue structure are processed separately,which achieves effective artifact suppression and good protection of structure and details.Additionally,a preprocessed image is used to improve the accuracy of block matching,and the total-variation(TV)algorithm is applied to further reduce the residual artifacts in the recovered image.We evaluate the D-WNNM method on the simulated pelvis phantom,the actual thoracic phantom,and the clinical thoracic data,and compare it to several other competitive methods.Experimental results show that the D-WNNM method has better performance in both artifacts suppression and structure preservation.Particularly,the number of iterations required in the proposed algorithm is substantially reduced(only twice),when compared with that required inthe WNNM method(at least eight iterations).