Impact of Segmentation Methods on Predicting Fatigue-Initiating Pores from X-ray Computed Tomography Data

Abstract X-ray Micro Computed Tomography (X-µCT) is increasingly regarded as the gold standard for inspecting additively manufactured components used in fatigue-critical applications. However, segmentation of X-µCT data remains inconsistent across users and applications. Additionally, it is unclear if voxel-wise metrics of segmentation quality, such as the Dice coefficient or Intersection over Union (IoU), are relevant to fatigue performance. In this work, we evaluated global binary thresholding, adaptive thresholding, hysteresis thresholding, and a 2.5D U-Net on X-µCT scans of Powder Bed Fusion – Laser Beam manufactured Ti-6Al-4V rotating bending fatigue specimens from the NIST AMBench 2025 challenge (AMB2025-03-FL) to quantify segmentation-induced measurement bias and assess its impact on predicting the fatigue-initiating pore. To identify the fatigue-initiating pore, the Murakami $$\sqrt{A}$$ parameter was modified using beam theory to account for the stress gradient due to bending. This modification enables localization of the most critical pore, independent of the applied stress. Using the modified $$\sqrt{A}$$ metric, the global binary thresholding, adaptive thresholding, and hysteresis thresholding predicted the fatigue-initiating pore correctly for three out of the four test cases. The 2.5D U-Net was able to predict the fatigue-initiating pore in all four cases despite having a lower Dice coefficient and IoU values when compared to the other segmentation models. In fatigue-critical applications, these limitations are most consequential for small, near-surface pores, where X-ray reflection artifacts cause threshold-based methods to underestimate pore size. Such pores can occur even in high-density PBF-LB parts. Consequently, voxel-wise metrics such as Dice coefficient or IoU do not indicate whether a segmentation approach can identify the true fatigue-initiating pore, as they weight all pixels equally, highlighting the need for fatigue-aware, feature-based methods for segmentation of X-µCT data.