Fast Industrial Computed Tomography for Remanufacturing: Evaluating Measurement Time savings with 3D UNETR Sparse-view Segmentation

Abstract

Industrial computed tomography (CT) provides non-destructive insight into internal structures and full 3D geometry of complex assemblies, but its use in production is limited by long acquisition times. Sparse-view CT can reduce measurement time, yet introduces artefacts and geometric bias that may compromise metrology. We evaluate sparse-view, machine-learning-based component extraction on an assembled angle grinder containing a bevel gear, using five gear variants with wear and damage. A 3D UNETR model is trained for signed distance field (SDF) regression to enable robust surface extraction. Across five leave-one-gear-out runs, we assess segmentation robustness under reduced acquisition time and translate the results into metrology-relevant outcomes via surface deviations and a tolerance-oriented tooth-profile evaluation. Results show that acquisition time can be substantially reduced within the trained regime while maintaining high segmentation quality and largely stable feature-level deviations. For stronger reductions, errors increase non-linearly and become spatially localized, dominating tail statistics and tolerance failures. Beyond demonstrating time savings, the proposed tolerance-centered evaluation links sparse-view segmentation performance to feature-level deviation distributions and pass rates, providing a transferable basis for selecting acquisition-time settings and motivating adaptive inspection strategies that trade measurement time against decision reliability.