Quantitative characterization of cement hydration microstructures from X-ray micro-computed tomography (μCT) images is fundamental to understanding hydration mechanisms and developing data-driven cement materials. However, existing deep learning approaches rely heavily on manually annotated datasets, limiting scalability and reproducibility. This study presents a physics-guided artificial intelligence framework for automated characterization of cement hydration microstructures using μCT imaging. A multi-Otsu thresholding strategy, guided by the physical relationship between X-ray attenuation and material density, was developed to automatically generate pixel-level pseudo-labels representing pore space, hydration products, and unhydrated cement. The proposed workflow produced a curated dataset comprising 7638 labeled μCT images extracted from 28 volumetric regions of interest spanning multiple hydration ages, Blaine fineness levels, and water-to-cement ratios. The pseudo-labeled dataset was used to train and evaluate three semantic segmentation architectures: U-Net, Attention U-Net, and U-Net++. All models achieved excellent segmentation performance, with mean Dice coefficients exceeding 0.94, while the baseline U-Net achieved the highest overall accuracy (Dice = 0.9454, IoU = 0.8982). Quantitative analysis of the segmented microstructures successfully captured the expected temporal evolution of pore space, hydration products, and unhydrated cement. ROI-level statistical analysis showed that hydration age, Blaine fineness, and water-to-cement ratio influenced morphological descriptors across the three microstructural phases, revealing systematic differences in microstructural evolution. The proposed framework reduces reliance on manual pixel-level annotation while enabling reproducible and scalable characterization of cement hydration microstructures. These findings provide a foundation for automated μCT-based cement characterization and future materials informatics applications.