Synthetic, automatically labelled training data for machine learning based X-ray CT image segmentation: Application to 3D-textile carbon fibre reinforced composites

Artikel i vetenskaplig tidskrift, 2025

Composite parts with 3D-textile reinforcement show promise in high-performance applications. For widespread use, accurate material characterisations are required. Characterisation of the textile architecture in the as-manufactured state may be performed with X-ray CT. Due to the similarity between the chemical composition of carbon fibres and epoxy based matrices, the contrast of X-ray CT scans is poor. Therefore, segmentation with classical methods is difficult or even impossible. Alternatively, machine learning based segmentation approaches may be used. One drawback of machine learning-based algorithms is the need for large datasets whose ground truth labellings require extensive manual labour. This can be circumvented by utilising automatically labelled synthetic X-ray CT data. In this work, a novel pipeline that generates synthetic CT image datasets, with automatically labelled ground truths, is developed. The pipeline is entirely based on free and/or open source software. It is demonstrated that segmentation model, trained on only such data, is able to accurately segment a real X-ray CT scan of a 3D-reinforced carbon fibre composite sample. A pixel-wise agreement of 88% is reached when compared to a manual segmentation. This implies potentially large time savings in segmentation tasks, which could accelerate characterisation of textile composites in their as-manufactured state.