A synthetic CT dataset for beam‑hardening artifact correction in printed circuit board assembly imaging

An efficient physics-based simulation pipeline is then developed to generate paired artifact-affected and artifact-free CT images. Comparisons with Monte Carlo simulations validate the realism and computational efficiency of the proposed method. For technical validation, we design both numerical simulation and real-data experiments to comprehensively evaluate the dataset. Specifically, we partition the 200 volumetric datasets into 174,798 non-overlapping 512×512 2D CT slices, and then randomly selected 1,596 paired artifact and ground-truth images for model training and 294 artifact images for testing.

Three representative DL-based methods (U‑Net, TransNet, and β‑CycleGAN) are trained and evaluated on this dataset. Both qualitative and quantitative results demonstrate that our dataset consistently enables effective BHC across all three models, and further cross‑dataset experiments on real measured data confirm its reasonable generalization capability.