The Research On Metal Artifacts Reduction In Computed Tomography

In computed tomography, metallic implants such as hip prostheses, dental fillings, and spine fixation in the scanning field will create the so-called metal artifacts in the reconstructed images. These artifacts obscure anatomical structures near the metal objects, making it difficult for radiologists to correctly interpret the images. When a single small metal object appears, it often produces streak artifacts around the implant. However, if large or multiple metal objects are present in the measurement, they are usually accompanied by streak artifacts, also the cupping artifacts and dark shadows originating from beam hardening. Hence, the research on metal artifact reduction(MAR), especially for large or multiple metallic implants cases, is always one of the difficult and hot problems in CT applications.MAR algorithms can be roughly fallen into two classes: projection completion and iterative image reconstruction algorithms. There are residual artifacts and loss of anatomical structures adjacent to metallic implants by traditional MAR methods. Therefore, some improved or novel MAR methods are proposed in this work. The research covers image segmentation, projection interpolation, parallel acceleration and iterative reconstruction, which aims to improve the effectiveness of MAR. The main contributions are summarized as follows:(1) Through the analysis of clinical CT images with metal artifacts, we aim to generate a prior image with complete anatomical information via edge-preserving filter and recovery of adjacent tissue structures. In this work, a novel MAR method based on prior image is proposed. And the proposed algorithm can perform well in secondary artifacts suppression and bone structures preservation. Furthermore, a regional smoothing projection completion approach based on the prior is also developed. The proposed interpolation technique can effectively solve projection inconsistencies or multiple metal objects problems in the process of MAR.(2) Forward projection of original uncorrected image, metal image and prior image and filtered back projection procedures contribute to about 90% of the total computational time during the process of proposed MAR. Based on the characteristic of high computation costs and obvious parallelism, we use a ray-driven forward projection, a pixel-driven back projection and texture memory interpolation to realize the parallel acceleration under the CUDA framework. In comparison to CPU, the GPU brings a significant gain with an acceleration that the computation time is shortened by about 7 times.(3) To simulate the beam hardening, scatter, noise, etc… that produce metal artifacts, we use the Poisson model to generate the noisy projection data. In projection domain, the raw projection data can be segmented into metal projections and non-metal projections. Using the metal projections, we propose a tangent back projection method based on the relationship between metal projections and metal implant in image space to determine the location and shape of metallic objects in the reconstructed image. Using the non-metal projections, we develop an iterative metal artifact suppression algorithm based on the minimization of proposed constrained optimization model. The results demonstrate that the proposed algorithm can successfully suppress metal artifacts, reduce noise and improve anatomical structure visibility. Furthermore, through the analysis of the connection of l0 norm penalty function and l1 norm penalty function, a user-defined weight function is used as penalty weight to generate the weighted total variation(TV) model. The reweighted TV reconstruction algorithm can be implemented through the alternation of solving the weighted TV minimization problem and updating the weights procedure. And the experimental results indicate that the proposed MAR algorithm not only suppresses artifacts and noise but also restores clear edge structures.