Computed Tomography Artifact Reduction Employing a Convolutional Neural Network Within the Context of Dimensional Metrology

Abstract Utilizing accurate, nondestructive testing methods to improve quality control and reduce manufacturing errors has gained prominence in light of industry development in various fields. Industrial computed tomography (CT) scanning carries considerable weight among all conventional methods because of their unique features, such as providing a three-dimensional specimen model. Due to the prevalence of metals with high linear attenuation coefficients in industrial applications, beam hardening and scatter artifacts are two of the most prevalent artifacts in any reconstructed volume. Other notable artifacts include those with a nonideal focal spot and conical beam radiation. These artifacts may manifest as a distortion of gray value peaks, systematic discrepancies, blurring-like cupping, and streaking in reconstructed images, degrading volume reconstruction quality. In this paper, the effect of these artifacts is illustrated and mitigated by adopting our proposed method, a combination of conventional and contemporary techniques, including the use of a pretrained convolutional neural network (CNN). Five tests are replicated in different geometric parameters to perform a geometric configuration analysis, indicating how effective the proposed approach is at encountering different geometric situations. The results demonstrate that the proposed method has substantially achieved its goal of improving the accuracy of dimensional metrology performed on our phantom.