Research On Sinogram Noise Reduction Method And Post-processing Approaches For Low-dose CT

Since the British engineer G. N. Hounsfield successfully developed the world first CTmachines in the early1970s, the x-ray computed tomography (CT) imaging technology playsan irreplaceable role in the field of radiation, has been widely applied in clinical diagnosis.However, the ray radiation problem during x-ray CT examinations has caused wide-spreadconcern. The low-dose CT scanning technology can not only alleviate the harm caused bycumulated radiations to patients but also can further expand the scope of application of CTtechnology. Among all the methods proposed so far to obtain LDCT images, the mostpractical and widely used method is to lower the x-ray tube current, however reduce tubecurrents make the projection data corrupted by quantum noise, leading to degradedreconstructed images with noise and streak artifacts. Therefore, how to improve the quality ofreconstruction image in the lower radiation dose CT has been hot topic in CT research field,and has the important scientific research value and the clinical use value.This article mainly aims at the noise problems in the projection data and in reconstructedimages under the low-dose environment, the main contributions are as follows:1. For the insufficiency of Gaussian MRF prior model in the GS-PRWLS algorithm, basedon anisotropic diffusion model in partial differential equation, we designed a new GaussianMRF prior model with anisotropic smoothing effect, and according to the prior model andnoise statistics characteristics of the projection data, using the maximum a posteriori (MAP)method to get an adaptive sinogram restoration algorithm. The algorithm can adaptivelyadjust the degree of smoothness according to the noise level and the region feature in thesinogram, avoiding the drawbacks of GS-PRWLS algorithm using a fixed parameter. Visualeffect together with quantitative analysis of the experimental result shows the developedapproach has the excellent performance in protection of the edge and removal of streak artifacts in the reconstructed image.2. Aiming at the noise problems in the projection data, a new sinogram smoothingapproach based on a total variation prior is proposed. Based on the MAP statistical iterativealgorithm, the new algorithm integrates the statistical properties of the projection data and thetotal variation prior information into sinogram recovery to achieve the purpose of suppressingnoise and preserving edge. The new total variation prior model is based on the original totalvariation model and introduces an adaptive punishment. The punishing power can make thealgorithm automatically distinguish between the edge and the flat areas and adaptively adjustthe smoothing degree according to the structural characteristics and the level of noise of thesmoothed pixel. Experimental results indicate that the proposed algorithm has the excellentperformance in suppressing noise and preserving edge and the image quality is superior toother methods.3. Obvious streak artifacts and noise often appear in low-dose CT (LDCT) images,seriously degrading CT image quality. Aiming at improving the quality of LDCT images, anartifact removal method based on dictionary learning and morphological component analysis(MCA) is proposed. The method formulates streak artifacts removal as an imagedecomposition. First, the LDCT image is first decomposed into the low-frequency (LF) andhigh-frequency (HF) parts by a bilateral filter, then the image decomposition method based onMCA is applied to the high-frequency part in order to separate out the streakartifacts. Different from the traditional decomposition method based on MCA, This methoddoes not need to select fixed dictionary by experience, nor need to collect sample images inadvance, but seperates the dictionary learned from HF part into artifact subdictionary andtissue subdictionary according to the directional characteristics and frequency characteristicsof artifacts. Based on sparse representations theory, the HF part is decomposed into artifactcomponent and tissue component by performing sparse coding, so as to achieve the purposeof removal of artifacts. At last, for further suppressing residual artifacts and noise, a dictionarylearning (DL) method is applied. The results of numerical simulation and clinical dataexperiments indicate the proposed algorithm can not only effectively remove artifacts and noise but also maintain structural details in the low-dose CT image.