Understanding the mesostructure of concrete relies heavily on the accurate distinction of its internal phases from X-ray computed tomography (X-CT) images. However, conventional intensity-based segmentation techniques often misidentify heterogeneous aggregates and cement matrix due to more severe overlapping grayscale ranges, leading to unreliable mesostructure quantification. To address this long-standing limitation, a lightweight deep-learning framework enhanced with a Squeeze-and-Excitation (SE) attention mechanism—referred to as SE U-Net—was employed to achieve fine-scale multiphase separation in concrete containing heterogeneous aggregates. The model reached a mean intersection-over-union (mIoU) of 91.63% while maintaining only a quarter of the parameters of a traditional U-Net, ensuring computational efficiency without compromising accuracy. More importantly, the subsequent three-dimensional visualization based on the segmented outcomes provides an accurate representation of the spatial arrangement of the mesostructure. This approach enables more precise identification of heterogeneous aggregates and fine voids in grayscale-overlapping regions, enabling reliable quantitative evaluation of aggregate area, size, distance, and void content with errors below 6%, whereas conventional thresholding exhibits over 30% deviation. This work demonstrates a practical route for precise and efficient image-based characterization of concrete and establishes a robust foundation for linking mesostructural features with macroscopic performance in sustainable cementitious materials.
Jin et al. (Sun,) studied this question.