Probability of detection based evaluation of XCT resolution limits and voxel sizing errors in defect detection

This study investigates sizing error and the impact of scan resolution on defect detectability and porosity analysis in aluminum alloy samples, using a probability of detection framework. X-ray computed tomography (XCT) scans were generated at resolutions of 10 µm, 20 µm, 35 µm, and 70 µm. The highest-resolution XCT scan (10 µm) was established as the ground truth, serving as a reference to quantify detection performance, sizing errors, and segmentation accuracy (via dice coefficient and Jaccard index) across lower-resolution scans. Defects were segmented, matched across resolutions using centroid-based registration with a 60 µm tolerance radius, and evaluated via POD curves generated from probit regression on hit/miss data. Results show that detectability declines markedly with lower voxel size: the defect size achieving 90% POD at 95% confidence for equivalent diameter rises from 100 µm at 20 µm voxel size to 235 µm at 35 µm voxel size. Lower resolutions systematically overestimate defect size due to partial volume effects, inflate sphericity (from ≈ 0,80 to ≈ 0.87), and reduce segmentation overlap (Dice coefficient drops from > 0.99 
to ≈ 0.97). Our study shows that optimizing scanning resolution is essential, as critical defects can be missed or mis-sized at lower resolutions, directly impacting downstream analyses such as finite element or fatigue simulations. These sizing errors propagate into such analyses, potentially leading to over- or underestimation of defect size and safety margins. The results and generated POD curves offer practical guidance for selecting scan parameters that effectively balance imaging fidelity and sizing error against processing time in nondestructive evaluation.