Attenuation-adjusted deep learning of pore defects in 2D radiographs of additive manufacturing powders