Research On Low-Dose Noise Suppression And Super-Resolution Reconstruction Algorithm For High Resolution Micro-CT

High Resolution Micro-CT is an imaging modality of new type which uses X-ray imaging principle to achieve high resolution 3D imaging. It is widely used in small animal in-vivo imaging, bone microstructure research, drug development, paleontology research and many other fields due to its advantages of non-destructive, non-intrusive and high resolution. To achieve the high resolution of the system, it makes use of the pipe line from the micro focal spot X-ray source to the X-ray coupled optical detector. Due to both low power energy of X-ray source and low X ray conversion efficiency of the detectors, the photon number may be insufficient, which makes the signal-to-noise ratio (SNR) of the projection data lower and leads to degradation of the reconstructed images by noise. Although the increase of radiation dose can improve the imaging quality, the excessive exposure will bring lesions to biological cells, tissues and organs. Hence, how to effectively suppress the low-dose noise of reconstructed images appears to be a difficult technical problem. And meanwhile, circumscribed by the physical conditions and the actual demands of the applications, the existing X-ray system doesn’t fully meet the requirements of high resolution of the system detection, which makes it very necessary that the digital method is employed to achieve the images with super-resolution quality. To fulfill the application demands of high resolution micro-CT, this project mainly works on these two aspects.For the problem of low-dose noise suppression, this paper mainly focuses on the study of projection data noise characteristics and the denoising algorithm of image reconstruction. Firstly, some statistical methods are employed to analyze the noise model of projection data. Then, by analyzing and improving the classical denoising algorithm:Penalized weighted least-squares (PWLS) algorithm, this paper proposes an edge-preserving method based on the PWLS. This enhanced algorithm makes the result acquire a balance between smoothing and edge retention. The experimental results of actual Micro-CT data show that the proposed approach can get higher SNR and better performance of evaluation compared to the classical PWLS algorithm. In addition, it can get clearer and sharper edges in the images.Projection and back-projection are the most time-consuming steps in PWLS algorithm. It makes low efficiency in 3D cone beam reconstruction especially. This paper develops a parallel optimization algorithm based on the implementation process of PWLS algorithm using the Compute Unified Device Architecture (CUDA) environment introduced by Nvidia. Depending on the different parameters of the main loop, alternative parallel computing methods are deployed, which are based on ray-tracing or voxel-tracing method respectively. Specifically, in the projection-paralleling process, the reconstructed data is bound to the 3D texture and the texture fetch technique is used for accelerating the filtering speed of the reconstructed data; in the back-projection-paralleling process, according to the specified voxel coordinates, the projective coordinates in the detector plane are determined. Besides, the corrected value of projection data is bound to 2D texture to accelerate the interpolation calculation. The experimental results show that the proposed parallel optimization algorithm can obtain almost the same quality of reconstructed images compared with CPU and the execution efficiency has been improved by a factor of 12.For the spatial resolution issue of Micro-CT image, this paper proposes a super-resolution image reconstruction algorithm based on dictionary learning. Firstly, the reconstruction grid is refined and the area weight model is introduced to achieve accurate modeling of the projection process. Then, for the refinement of the reconstruction grid, the number of unknown variables increases which leads to a problem of underdetermined equations, the dictionary learning method is introduced. Namely, with the K-means Singular Value Decomposition (K-SVD) algorithm, a dictionary is learned from the high quality images, and then, on the basis of this dictionary, the sparse constraint is introduced into the reconstruction process to solve the problem. The experimental results show that the proposed approach can effectively improve the spatial resolution of the reconstructed images compared to the conventional bilinear interpolation processing method.