Study Of Image Quality Improvement Algorithm For Low-Dose CT

Recently, CT(Computed Tomography) has made a great contribution to clinical diagnosis. Low-dose CT image, which is attracted more and more attention, have been gradually used in clinics to lower the radiation. However, downscaling the radiation intensities over the entire scan results in decreased image quality of CT image, leading to the reduction of clinical accuracy consequently. Therefore, it is meaningful to study on the study of improving image quality of low-dose CT image.Downscaling the radiation intensities over the entire scan results in increased quantum noise which is the main reason why the image quality of low-dose CT image decreases. Quantum noise, which depends on the pixel value of image, can be approximately modeled by a Poisson distribution. Firstly, this thesis goes on a detailed study of discrete shearlet transform and the traditional denoising algorithm based on shearlet for Poisson noise. Then, this thesis proposes a new de-noising method for low-dose CT image, according to the detailed study of discrete shearlet transform and the characteristic of noise in low-dose CT. Besides, the proposed method is proven to improve the image quality of low-dose mammography. The following is the main work of this thesis:a. A study on the reason of the degradation of the image quality of low dose CT. Downscaling the radiation intensities over the entire scan results in increased quantum noise which is the main reason why the image quality of low-dose CT image decreased. Therefore, Poisson is adopted to simulate quantum noise in low-dose CT.b. Aiming at the limitation of traditional wavelet transform for the representation of high dimensional function, this thesis goes on a detailed study on shearlet which has a rapid development recently. Shearlet is proven to have good characteristics: good localization, the excellent characteristics of multi-scale, multi-direction and the optimal sparse representation. This thesis also studies thestructure of shearlet. From the study above, Nonsubsampled Shearlet Transform(NSST) is proven to be suitable for medical image denoising. Further, this thesis proposes a variance stabilizing transform by combining discrete shearlet transform with Anscombe transform and has a detailed analysis of effectiveness of the transform for the noise removal of Poisson noise.c. A comparative study of the traditional algorithm for improving the image quality of low-dose CT based on shearlet: Threshold de-nosing method based on shearlet, shearlet denoising method based on total variation regularization, shearlet denoising method based on maximum a posteriori estimation. Based on this study, this thesis proposes a new shearlet-based adaptive de-noising method to remove Poisson noise in this thesis. We first transform Poisson noise into a nearly Gaussian noise by a shearlet-based multiply variance stabilizing transform(SMVST) which combines NSST with Anscombe transform in shearlet domain. Second, the positions of non-noise shearlet coefficients are found by an adaptive de-noising method based on correlation of adjacent shearlet coefficients. The shearlet coefficients of an image are supposed follow Laplace prior distribution, and the position of non-noise coefficients are obtained by MAP(maximum a posteriori estimation) estimation. Finally, the de-noised image is estimated from the position of non-noise coefficients using the iterative scheme based on Hybrid Gradient Descent(HGD). The experiment result has demonstrated that the proposed method is effective both on visual effect and quantitative evaluation.d. The proposed method is applied to improve the image quality of real low-dose CT and low-dose mammography, which further verify the effectiveness for the noise removal of X ray quantum noise.