Iteratively Reweighted Least-Squares Implementation for Accurate Extraction of Prior Knowledge for Bayesian Image Reconstruction

Extracting prior knowledge from previous high quality normal-dose computed tomography (NdCT) data for Bayesian reconstruction of current low-dose CT (LdCT) images has attracted great research interests recently. While most efforts focused on registering the previous NdCT data to the current LdCT reconstruction, this work investigated an alternative strategy by extracting the local structure-specific spectral information of the NdCT data to preserve the local image textures in the LdCT reconstruction, because many clinical studies have revealed the image textures are clinically desirable. In implementation, we adapted a Gaussian Markov random field (GMRF) model to incorporate the spectral information for knowledge-based penalized weighted least squares (PWLS) LdCT reconstruction. Specifically, we explored a novel idea of extracting the spectral information by statistically estimating the weighting coefficients of the GMRF prior model. We adopted an iteratively-reweighted least squares regression (IRLSR) method to statistically estimate the linear relationship between the central voxel intensity and adjacent neighboring voxel intensities, which overcomes the limitations of ordinary least square regression. This novel idea led to very encouraging results as demonstrated by outcomes of the Bayesian reconstruction of LdCT image data.