Low Dose CT Image Restoration by Incremental Learning and Ant Colony Optimization

Patch priors play a key issue in patch based image restoration tasks. The main idea is to learn a good and adaptive image prior over small image patches. However, expected patch log ikelihood (EPLL) framework achieved a good performance in image restoration applications, it mainly employs the Gaussian Mixture model (GMM) prior over a large number of training patches which gains a huge computational cost. On the other hand, the GMM parameters are learned with the Expectation Maximization (EM) algorithm which usually suffers from local optimum problem. This paper proposes a post filter by extending the EPLL framework to reduce noise in CT images acquired on low radiation dose, in order to get a better quality image with fine details preserved. The basic idea is to incrementally estimate the mixture component parameters to skip over the significant drawback of the (EM) algorithm which greatly depends on the initialization of model parameters. Incremental method with split-merge mixtures learning (SMILE) is utilized for learning mixture models by consecutive split and merge operations. Then, the Ant Colony Optimization (ACO) algorithm is used to select homogeneous patches which tend to cluster together due to the local spatial smoothness in the image. Instead of restoring each patch separately, the maximum a posteriori (MAP) estimation is employed to denoise the homogeneous patches selected by ACO algorithm. The proposed method leads to better clustering results and superior denoised results.