Synthetic training data for CT image segmentation of microstructures

Abstract

The segmentation of images obtained through techniques such as computed tomography is a key step in generating digital twins of porous microstructures. A common approach to segmentation is the use of supervised machine learning algorithms, such as U-Net. The training data required for such algorithms are usually obtained by manual labeling, which is extremely time consuming and often inaccurate. We present a method for synthesising realistic training data for segmentation algorithms. This method generates the data in a two-step process that iteratively improves the quality of the synthesised training data. Finally, we validate the similarity between synthetic and real data using quantitative and qualitative metrics and further demonstrate the effectiveness of the synthetic data by experimentally validating segmentation results against measured material properties.