Fig. 1
From: Deep learning-based metal artifact reduction in CT for total knee arthroplasty
The overall development process of KMAR-Net. A total of 25,000 image pairs were generated through the sinogram handling method to train and validate KMAR-Net. KMAR-Net was trained with the training set of 15,000 image pairs, and the final model was selected with the validation set of 5,000 image pairs. For the simulated test set, the performance of KMAR-Net was evaluated by calculating image quality metrics. Finally, the output images of the KMAR-Net were compared with O-MAR images for the clinical test set.
