Recent advances in civil engineering increasingly leverage artificial intelligence (AI) for image- and volumetric data driven tasks such as microstructural analysis, material characterization, crack detection, and material separation for recycling. Despite growing interest, the adoption of AI in these areas remains
constrained by two key barriers: labor intensive development pipelines dominated by dataset preparation and annotation, and the limited generalizability of trained models when applied to new data distributions. These challenges are particularly evident in imaging data of cementitious materials, where heterogeneous textures, multiscale features, and variations in material composition and imaging conditions complicate both annotation and model transferability. To address these challenges, this paper proposes a domain specific annotation and training framework that reduces dataset preparation effort while supporting adaptable model development. The framework combines automated and user-guided mask annotation across sequential image data with a lightweight training pipeline that enables the rapid creation of inference ready models. Demonstrations on microstructural imaging data from cementitious composites, illustrate how the framework streamlines annotation workflows and improves label consistency under realistic data variability. The proposed methodology provides a scalable foundation for applying deep learning across a range of civil engineering imaging applications, from laboratory scale material studies to larger scale structural monitoring.