Deep Monochromatic Metal Artifact Reduction for Computed Tomography

Abstract

Computed tomography (CT) is a three-dimensional medical imaging modality that uses X-ray beams to generate cross-sectional images of the human anatomy. Although CT is widely used in medical diagnosis, the presence of metal implants often severely impair the diagnostic value of CT images. The presence of metal implants causes errors in the image that are called “metal artifacts”. Existing metal artifact reduction (MAR) algorithms are either ineffective or require a large training dataset which is difficult to attain due to the inaccessibility of clinical data. Thus, this study proposes a novel end-to-end convolutional neural network with autoencoder embeddings for MAR that overcomes the shortcomings of existing methods. Unlike existing methods that simulate training data using artificially synthesized metal implant shapes, our research proposes a new data synthesis technique that uses randomly generated polygons to automate the data simulation process. Experimental results prove that this method drastically improves the efficiency of the data generation process. Our proposed network also significantly outperforms state-of-the-art MAR techniques, achieving an MSE < 7 × 10− 6, an SSIM index > 0.994, and a PSNR > 58 dB on a simulated training dataset of 130 samples.