Study On Wavelet Denoising For Low-dose Computed Tomography

Computed Tomography (CT) technology has been widely used in clinics. However, higher radiation dose confines its further application with mass screening such as the examination of people under higher risk of lung cancer. The radiation dose is mainly determined by the scanning parameters of the system. To lower the radiation dose, low-dose protocols have been used clinically as an alternative for above situations. However, the noise existed in low-dose CT images not only decreases the image quality and the accuracy of the diagnosis, but also increases the difficulties in image analysis and processing. In order to improve the image quality of low-dose imaging, this paper proposes wavelet-based denoising methods that integrate the noise properties of low-dose CT projections. After sinogram filtering with the proposed methods, the image are reconstructed by classical filtered back projection (FBP) method.In this study, to obtain the statistical properties of the projection data, experiments with physical phantoms were first performed to acquire low-dose projections repeatedly at a fixed angle. Statistical analysis of the data revealed that the noise of the projection data could be regarded as approximately Gaussian distributed with a nonlinear signal-dependent variance, and non-stationary noise model was established based on above findings. The distribution of signal and noise in wavelet domain was analyzed and wavelet-based denoising algorithms with different basic functions have been investigated and compared. Considering the feature and the distribution of the noise in wavelet domain, the paper proposes two methods for the filtering of low-dose CT projections, namely the adaptive wavelet coefficient denoising based on Bayesian estimation and the local adaptive threshold denoising based on stationary wavelet. In the first method, the Bayesian estimation was used to achieve adaptive wavelet coefficients. After wavelet decomposition, the wavelet coefficients for lower and higher frequency were determined respectively, and then the variance for each layer was estimated with maximum-likelihood (ML) algorithm. Finally, the new wavelet coefficients used for reconstruction were estimated adaptively with minimum mean-square error (MMSE) method to separate the image signals from noise effectively. The latter alleviates the Gibbs ringing effect caused by thresholding process with orthogonal wavelet, and image edges, boundary and details could be reserved for better visual effect. The simulation and experimental results show that the two methods could reduce noise efficiently. Since the proposed framework for noise reduction is algorithm based software, it could be used directly in current CT equipment and has a potential to be used in a broad clinical practice.