Multi-scale Noise Reduction Algorithm For Low Dose CT Sinograms

Computed tomography imaging is used more widely in medical diagnosis and therapy in the recent years, along with the great improvement in quality and speed of CT facilities. Comparing with other medical imaging technologies like X-ray scan, CT excels in its tomographic nature and high resolution. However, high radiation dose limits its further use in clinics, such as screening of lung cancer. Via limiting the parameters, tube current for example, in CT scans, radiation can be greatly reduced. However the acquired CT sinogram data are contaminated with excessive noise. And the reconstructed images degrade severely by streak artifacts. To improve the quality of low dose CT images, considering the so far excellent performance of multi-scale method in image denoising, this paper explores multi-scale algorithms for noise reduction of low dose CT sinogram data. After statistically-based restoration, sinograms were reconstructed using the classical FBP method for diagnosis as well as further analysis and processing.The noise property in low dose CT sinogram data is characteristic by non-stationary and nonlinear. In this paper, dyadic wavelet transformation was adopted to perform multi-scale decomposition. Firstly, the noise property of low dose sinogram data in wavelet domain was studied, and it turned out to be approximately consistent with what in spatial domain. Therefore, the nonlinear mean-variance relationship obtained in spatial domain could be used directly for the estimation of wavelet coefficients. The consistency of noise properties between spatial and wavelet domain makes statistically-based restoration in wavelet domain feasible. With empirical study, suitable decomposition level of wavelet transform in the restoration process was obtained. After analyzing both the geometric correlation of wavelet coefficients and high local correlation characteristic of sinogram data, similarity-based restoration strategy (framework) was proposed for noise reduction of low dose CT sinograms. Based on this framework, two wavelet domain denoising methods were evaluated in this paper, namely wavelet based bilateral filtering and wavelet based non-local means filtering (NLMF). The former showed some improvement over traditional filtering methods. However, due to the ineluctable limitation of bilateral filter of staircase effect, it couldn't fully satisfy the requirement of low dose denoising. As a counterpart, after statistical modification of traditional non-local means filter, wavelet domain NLMF achieved satisfying results for low dose sinograms. Experiments with computer simulations and phantom data show that the proposed wavelet-based NLMF outperforms the up-to-date wavelet based PWLS smoothing and preserves more details, which are of most importance for medical images.