A Study On Some Problems In Image Reconstruction For Low-Dose CT System

The X-ray computed tomography (CT) technology has been developing for more than fourty years since its first commercial application in early 1970s. Nowdays, CT technology is one of the important im-aging approaches for modern medicine. Owing to the high spatial resolution and high sensitivity, X-ray CT technology plays an irreplaceable role in clinical diagnosis. While having made great contributions to the human healthcare, X-ray CT technology is limited by its relatively high radiation doses. Therefore, the re-search works on low-dose CT (Low-dose CT, LDCT) technology become particularly critical for clinical application.Based on the analyses in a few of aspects of LDCT technology, this thesis focuses on the development of the imaging algorithms which directly improve the quality of LDCT imaging. These imaging algorithms include new smoothing algorithms for noised LDCT projection data and those for improving the quality of LDCT reconstructed image based on the traditional filter back projection (FBP) method. The main research works in this thesis are presented as follow:(1) Based on an anisotropic diffusion weighted priori model, a new Maximum A Posterior (MAP) al-gorithm was proposed for smoothing LDCT sinogram. For this work, a new weighted priori model was constructed by adding an anisotropic diffussion factor to the quadratic priori model based on Markov ran-dom field (MRF), which is a classic simple model and can be solved easily. To meet different requirements of suppressing noise for edges and flat regions in the sinogram, the diffusion property was analyzed in de-tail for the new anisotropic diffusion proir model. During the diffusion process, the new prior model can alter the diffusion strength and direction for individual regions by adaptively differentiating between edges and flat regions. After introducing the new prior model into the MAP statistical iterative reconstruction al-gorithm frame, the new MAP algorithm was applied in order to suppress noise for LDCT sinogram. The experimental results show that the new algorithm can improve the quality of the LDCT reconstructed im-age.(2) Based on the dictionary learning and sparse representation theory, a joint regularization term was proposed. Additionaly, a new algorithm was built by introducing the joint regularization term into the the penalized weighted least square optimization algorithm. The proposed joint regularization term contains a dictionary learning regularization term which can properly represent image edges, structural details and other information, as well as a quadratic regularization term which can filter the noise properly. Thus, the new model has the advantages of maintaining the edge detail and smoothing the noise. A new algorithm model was further built by combining the penalized weighted least square algorithm with the proposed joint regularization term. In order to solve the new model easily, another new equivalent algorithm was proposed by using the corresponding surrogate function according to the optimization transfer principle. Then, the iterative formula was given and the solving procedure for the original model was simplied. The improved LDCT image was reconstructed by FBP from smoothed projection data. The new algorithm was tested in the simulation data and the clinical data. Compared with the classic smoothing algorithms for LDCT sino-gram, the new algorithm has the advantages in both visual effects and signal-to-noise ratio value.(3) In the wavelet domain, a new directional non-local means algorithm was proposed for processing LDCT images. The new algorithm processed the degenerated LDCT images resconstructed by FBP from the noised projection data. Thus, the new algorithm is a processing algorithm after the reconstruction. By analyzing the features of the noise and the artifacts, the principle of the non-local means filter algorithm was used to process the LDCT images. To overcome the limitations of the classic non-local means filtering algorithm using the isotropic Gaussian kernel function, the new algorithm utilizes an improved weighted distance kernel, i.e., oval weighted kernel. The new weighted kernel function has anisotropic property, which can achieve a better compromise between maintaining edges and supressing artifacts. Based on this oval weighted kernel function, the directional non-local means filtering algorithm was constructed. Firstly, the new algorithm decomposed the degraded LDCT image into two scales by the two-dimensional station-ary wavelet transform (SWT). Then, the proposed directional non-local means algorithm was applied to the high-frequency components to suppress the artifacts. Moreover, the algorithm can use the different direc-tional oval weighted function and matching window. Finally, the improved LDCT image was reconstructed by inverse stationary wavelet transform (ISWT) based on the low-frequency component and the three pro-cessed high-frequency components. The impact of parameters on algorithm performance was evaluated in a few of experiments. Additionally, some advices were given on how to select the parameters. The perfor-mance of the algorithm was verified in both the simulated image and the real image. The results show that the new algorithm can achieve a superior balance between the suppressing artifacts and protecting edges.(4) In the wavelet domain, a new post-processing algorithm was proposed for LDCT images based on morphological component analysis (MCA). As a modern image processing method, MCA is just at the ear-ly stage. The new algorithm integrated the advantages of MCA with those of wavelet transform, leading to a good performance of artifacts removal for LDCT images. First, the new algorithm carried on two-dimensional SWT for LDCT images. Then, dictionaries were trained based on each of the high-frequency components by the dictionary learning method, respectively. At last, combination diction-aries were obtained. On this basis, the new algorithm utilized the MCA method to remove the artifacts in the high-frequency component obtained by wavelet transform. The improved LDCT images were obtained by the ISWT. Through analyzing the effect of processing the real degenerated LDCT images, the validity and superiority of the new algorithm were verified.