Dictionary Learning Based Low- Dose CT Image Processing Methods

X-ray Computerized Tomography is widely recognized not only as a reliable and accurate tool for the diagnosis of diseases but also as an efficient method for planning and guiding surgical interventions. Although the high image quality is generally acknowledged, radiation doses associated with CT have received serious concern. The CT dose is cumulative in lifetime, and repeated CT scanning can significantly increase the risk of cancers. Lowering the X-ray tube current or the tube voltage can reduce the radiation doses delivered to patients, but at the cost of a severe deterioration of the CT image quality due to increased quantum noise and artifacts. In this paper, we explore the application of dictionary learning based sparse representation algorithm in low-dose CT (LDCT) image processing.A growing interest in sparse representation and dictionary learning based signal/image processing (DL) was observed in the past decade. In the DL method, a global over-complete dictionary is often trained at the beginning. After the image is decomposed into overlapping patches, which are then represented as a linear combination of a limited number of dictionary atoms using sparse coding technique. In this step, the anatomical structures can be sparsely represented, but noise and artifacts cannot, thus the purpose of noise/artifact removal is achieved. Successful application of DL method has been reported in abdominal LDCT images. In this paper, we further explore the application of DL method in perfusion CT (CTp) and cardiac CT (CCT) imaging. A 3D DL based processing method is proposed and applied to improve low-dose CTp and CCT images. By taking the information along the temporal axis into account, the proposed method is in fact a 3D extension of the 2D DL method. Experiment results show that encouraging improvement of image qualities can be brought by the proposed 3D DL algorithm. It has been validated that the DL method can obtain good processed image quality under 1/5 of the routine X-ray tube current.It is often the case that strong noise and streak artifacts (with strong intensities) can also be sparsely representable and as a result cannot be removed effectively. To overcome this problem, we proposed a novel algorithm called the Artifact Suppressed Dictionary Learning (ASDL) method to process LDCT images with strong noise and artifacts. The ASDL algorithm includes two main steps, and performs noise and artifact reduction at different scales. In the first step, streak artifacts in LDCT images are suppressed by means of a discriminative sparse coding procedure in high frequency bands. Three novel discriminative dictionaries are respectively designed to characterize artifact and normal tissue feature components in different orientations. Then, the second step makes use of the general DL processing to further suppress the noise and residual artifacts. Experiments on both abdominal and mediastinum date validate the effective performance of the proposed ASDL method.