Enhancement and assessment of large vision models for 3D particle reconstruction from X-ray tomography

Three-dimensional (3D) particle reconstruction from X-ray micro-computed tomography (µCT) images is essential for digital twins and understanding the micromechanical behaviors of granular media. Despite large vision models (LVMs) having shown remarkable effectiveness across various domains, their application to accurate 3D particle reconstruction remains underexplored. This study proposes a systematic framework that enhances and leverages LVMs for the reconstruction of arbitrary 3D particles. The proposed framework includes three key steps: (1) enhancing LVMs with higher computational efficiency for two-dimensional (2D) label extraction, (2) mapping stacked 2D labels to 3D, and (3) extracting particle surfaces to generate 3D models. The enhanced approach is applied to reconstruct four distinct samples to validate feasibility. Six conventional and four lightweight LVMs are selected to explore the influence of model size and the number of prompts on reconstruction accuracy. The H-extreme watershed method is chosen as a benchmark for comparison. The results demonstrate that the enhanced framework can accurately reconstruct irregular and complex samples, such as carbonate sands, with a greater than 50% improvement in accuracy compared to the benchmark prediction. Additionally, the framework effectively reduces over- and under-segmentation errors, resulting in accurate reconstruction of microstructural characteristics. This versatile framework presents a promising alternative for investigating complex micromechanical mechanisms of granular media.