Limited-angle X-ray tomography (LAT) plays an important role in non-destructive inspection in situations where full 360° data acquisition is physically limited, such as specialized medical imaging and industrial applications. However, the incompleteness of the projection data makes the underlying inverse problem highly ill-posed. Traditional analytical methods and iterative reconstruction methods based on standard models suffer from serious “missing wedge” artifacts and low numerical stability. Although stand-alone deep learning approaches have shown promising results in artifact removal, they often lack the reliability of physical information and exhibit unexpected generalization in non-distributional scanning processes.
qizi et al. (2026) studied this question.