X-ray simulation and deep learning: Application to Automatic segmentation of defects in CT images corrupted by artefacts

The use of X-ray computed tomography (XCT) in precision engineering is becoming commonplace to assess the accuracy of a manufacturing process. The ISO 10360-11 standard has recently been issued to define metrological characteristics and methods making use of XCT. Two of the main challenges lies in extracting the 3D surfaces from XCT data and detecting manufacturing defects in images that are prone to artefacts such as beam-hardening, phase contrast, scatter radiation and partial volume effect.

We have developed gVirtualXRay, a fast programming library to simulate X-ray images on graphics processing units (GPUs) [1]. Being able to generate many X-ray images is useful in machine learning (ML). We deployed gVirtualXRay on Supercomputing Wales to train an optimisation algorithm reproduce real XCT images that where highly corrupted by beam-hardening and phase contrast [2]. Haiderbhai et al. Used gVirtualXRay to build a large training dataset of simulated images. This dataset is used to train a generative adversarial network (GAN), a ML approach to create synthetic images.

The segmentation of real images by ML requires a large amount of manual input to build training datasets. In this research, the PhD candidate will segment defects (e.g. cavities and cracks) from XCT scans and produce parametrised models to generate them in CAD models. Corresponding XCT scans will be simulated on Supercomputing Wales, with and without defects, with and without scanning artefacts, to build datasets in a controlled environment (the location and type of defects will be known, as well as the imaging artefacts). These datasets will be used to train and evaluate a convolutional neural network (CNN).

This research will benefit a large part of the manufacturing sector, where NDT by XCT or radiography is used. It will help material scientist detect manufacturing defects in XCT images corrupted by imaging artefacts. It will also enable the automation of defect detection in an industrial context.