A Nonlocal Kronecker-Basis-Representation Method for Low-Dose CT Sinogram Recovery

Low-dose computed tomography (LDCT) contains the mixed noise of Poisson and Gaussian, which makes the image reconstruction a challenging task. In order to describe the statistical characteristics of the mixed noise, we adopt the sinogram preprocessing as a standard maximum a posteriori (MAP). Based on the fact that the sinogram of LDCT has nonlocal self-similarity property, it exhibits low-rank characteristics. The conventional way of solving the low-rank problem is implemented in matrix forms, and ignores the correlations among similar patch groups. To avoid this issue, we make use of a nonlocal Kronecker-Basis-Representation (KBR) method to depict the low-rank problem. A new denoising model, which consists of the sinogram preprocessing for data fidelity and the nonlocal KBR term, is developed in this work. The proposed denoising model can better illustrate the generative mechanism of the mixed noise and the prior knowledge of the LDCT. Numerical results show that the proposed denoising model outperforms the state-of-the-art algorithms in terms of peak-signal-to-noise ratio (PSNR), feature similarity (FSIM), and normalized mean square error (NMSE).