Geometric Artifacts Correction for Computed Tomography Exploiting A Generative Adversarial Network

Abstract

Geometrical accuracy of an X-ray Computed Tomography (CT) system is crucial to achieve high quality tomographic reconstructions. Methods to correct the resulting geometric artifacts have been comprehensively described in the past few years. Deep convolution neural network is increasingly used in CT imaging, which has a great potential in image feature learning and processing tasks. In this work, a geometric artifact correction method exploiting generative adversarial networks (GAN) is developed. The U-Net structure is employed as the generator of the network to extract the CT image features with geometric artifacts. The convolutional neural network (CNN) based on image block perception acts as a discriminator for the network, further constraining the optimization of the generator. The proposed method has been shown experimentally feasible for geometric artifacts correction in circular cone beam CT by performing more accurate feature extraction. The peak signal to noise ratio (PSNR) of the corrected phantom images increased by 11.812 on average, and the root mean square error (RMSE) decreased by 9.982 on average.