Content-oriented sparse representation (COSR) denoising in CT images

Denoising has been a challenging research subject in medical imaging in general and in CT imaging in particular, because the suppression of noise conflicts with the preservation of texture and edges. The purpose of this paper is to develop and evaluate a content-oriented sparse representation (COSR) denoising method in CT to effectively address this challenge. A CT image is firstly segmented by thresholding into several content-areas with similar materials, such as the air, soft tissues and bones. After being ex-painted smoothly outside it boundary, each content-area is sparsely coded by an atom from the dictionary that learnt from the image patches extracted from the corresponding content-area. The regenerated content-areas are finally aggregated to form the denoised CT image. The efficiency of image denoising and the ability of preserving texture and edges are demonstrated with a cylinder water phantom generated by simulation. The denoising performance of the proposed method is further tested with images of a pediatric head phantom and an anonymous pediatric patient that scanned by a state-of-the-art CT scanner, which shows that the proposed COSR denoising method can effectively preserve texture and edges while reducing noise. It is believed that this method would find its utility in extensive clinical and pre-clinical applications, such as dedicated and low dose CT, image segmentation and registration, and computer aided diagnosis (CAD) etc.