Isolating the Unwanted: Deep-Learning-based Artifact Segmentation in Industrial CT Scans—Insights from a Case Study on Metal Artifacts

Computed tomography (CT) imaging is governed by the laws of physics. Nonetheless, it is necessary to make simplifications and assumptions that are only partly correct to meet extrinsic constraints, like cycle times or costs. This leads to artifacts in the resulting image. While much of the existing literature which deals with deep learning techniques focuses on direct elimination, there is limited focus on thoroughly understanding these artifacts to develop mitigation strategies that minimize typical risks such as hallucinated features. In our case study, we present an analysis of metal artifacts in industrial CT scans. When performing CT scans of multi-material parts with widely different absorption coefficients, like connector pins, metal artifacts affect the measurement accuracy in a number of ways. They lead to surface distortion, where the transition between the metal and other materials becomes unclear, and they obscure internal features like cracks, voids, or pores in the material. A challenging aspect when dealing with metal artifacts is to correctly detect the boundaries of the metal components. Utilizing deep learning, we investigate the feasibility of accurately identifying metal regions in new, unseen CT scans based on patterns learned from training data. In our experiments, we address the typical questions: “How accurately does the training data need to be labelled?”, “How do spatial and contrast resolution influence the robustness of the segmentation performance?”, and “How to maintain data integrity and avoid infringing potential third party intellectual property?”. This differentiation between metal and non-metal components in a CT scan is an important first step to specifically addressing the metal artifacts. Our approach enables a more robust analysis and improves the reliability of industrial CT imaging workflows.