This paper presents a methodology to optimise post-processing parameters in X-ray Computed Tomography (CT) for defect detection in metallic materials. The approach addresses three main goals: minimisation of systematic errors in defect reconstruction, quantification of uncertainty, and reliable defect classification. The proposed methodology aims to remove the systematic error that impacts defect reconstruction, thereby improving the accuracy of defect size and morphology assessment, which is essential for fatigue life prediction, particularly in materials produced through additive manufacturing (AM). An iterative comparison between CT-based defect and fractographic measurements is involved to identify the optimal CT post-processing parameters, such as the grey threshold (GT). The methodology was applied to 11 dog-bone-shaped titanium alloy samples (5.5 mm nominal gauge diameter) produced via electron beam melting. The optimisation procedure resulted in a GT value that was 134% of that obtained using a commercial algorithm, effectively removing the systematic uncertainty associated with Murakami's parameter area. The uncertainty of various defect features, such as equivalent diameter, sphericity and aspect ratio, was calculated by propagating the remaining stochastic uncertainty of area. An unsupervised K-means algorithm categorised unlabelled defects into three major types often encountered in AM: gas pores, keyholes, and lack of fusion. Finally, the labelled defects were processed through a support vector machine to infer the analytical form of the decision boundaries, achieving an accuracy of 99%.
