Efficient AI-aided segmentation of concrete CT scans using a predictive annotation interface

Concrete is characterized by a heterogeneous mesostructure consisting of aggregates, mortar, voids, and the interfacial transition zone (ITZ). X-ray Computer Tomography (CT) image processing can capture the 3D mesostructure of concrete, enabling its reconstruction for generating numerical models. A crucial step in this process is the segmentation of CT images. However, common segmentation approaches, such as thresholding, are not feasible for concrete due to the overlapping grayscale values of aggregates and mortar, which make it impossible to define a threshold for their separation. To address this challenge, segmentation software tools based on trained neural networks have proven to be accurate and robust, especially in biomedical CT imaging. In this contribution, we describe a segmentation workflow for concrete CT images using the neural network-based software SuRVoS2. This workflow effectively segments the concrete mesostructure of unaltered samples into three phases: cement, aggregates, and voids. The segmentation procedure begins with a shallow learning step to achieve an initial segmentation of the volume with minimal manual labor. In the second step, a U-Net Convolutional Neural Network (CNN) is trained using subvolumes. The final step involves predicting the entire volume using the trained CNN model and results in accurate segmentation with F1-scores higher than 97 %.