Commonly implanted medical devices containing metal parts (i.e., dental fillings, coils, hip replacements) generate streaks in computed tomography (CT) images, thereby impeding diagnosis and interfering with radiation therapy planning. Inventors at RPI created a novel technique to boost the efficacy of neural networks for metal artifact reduction (MAR) in CT images. Currently, deep neural network-based techniques need to be trained on synthetic, paired images. Unfortunately, these images may not accurately reflect clinical reality and technical factors. Neural networks that disentangle these artifacts with unpaired, clinical data show promising results. However, artifact disentanglement networks (ADN) are not able to recover the structural details of CT images based on generative adversarial losses only, requiring human insight and intervention. To overcome this limitation, the inventors designed a learning algorithm to train the disentanglement network. It simultaneously optimizes the network’s loss functions and constrains the recovered images to have a low-dimensional manifold representation enabling a superior approach that effectively extract features of real clinical CT images.