Learning-Based Metal Artifacts Removal in Head CT

摘要

Due to the presence of metal fillers, metal artifacts have always affected the effectiveness of computed tomography (CT) inspection. Moreover, metal artifact reduction (MAR) is still one of the major problems in clinical head CT. In order to reduce the metal artifacts in the dental region of CT images, we develop an artifact removal algorithm based on a deep convolutional neural network (CNN). The proposed approach consists of two-step. Firstly, we build a database consisting with and without artifact head CT image. In this step, a deformable image registration (DIR) method is implemented to preprocess data before CNN training. Therefore, pairs of with and without artifacts data are acquired from our dataset. Secondly, in the CNN training step, we build a simple 17-layer CNN architecture to learning the metal artifacts. Experimental results show the greater MAR capability of the proposed method. The computed tomography values, PSNR, and SSIM of ROIs also show the evident improvement.