A semi-supervised learning method of latent features based on convolutional neural networks for CT metal artifact reduction.

Abstract

X-ray computed tomography (CT) has become a convenient and efficient clinical medical technique. However, in the presence of metal implants, CT images may be corrupted by metal artifacts. The metal artifact reduction (MAR) methods based on deep learning are mostly supervised methods trained with labeled synthetic-artifact CT images. However, this causes the neural network to be biased toward learning specific synthetic-artifact patterns and leads to a poor generalization for unlabeled real-artifact CT images. In this study, a semi-supervised learning method of latent features based on convolutional neural networks (SLF-CNN) is developed to remove metal artifacts while ensuring a good generalization ability for real-artifact CT images. The proposed semi-supervised method extracts CT image features in alternate iterations of a synthetic-artifact learning stage and a real-artifact learning stage. In the synthetic-artifact learning stage, SLF-CNN is fed with paired synthetic-artifact CT images and is constrained using mean-squared-error (MSE) loss and perceptual loss in a supervised learning fashion. In the real-artifact learning stage, the network weight is updated by minimizing the error between the pseudo-ground truths and the predicted latent features. The feature level pseudo-ground truths are obtained by modeling latent features using the Gaussian process. The overall framework of SLF-CNN adopts an encoder-decoder structure. The encoder is composed of artifact information collection groups to map the input artifact-affected synthetic-artifact CT images and real-artifact CT images into latent features. The decoder is composed of stacked ResNeXt blocks and is responsible for decoding latent features with high-level semantic information to reconstruct artifact-free CT images. The performance of the proposed method is evaluated through contrast experiments and ablation experiments. The contrast experimental results indicate that the artifact-free CT images obtained by SLF-CNN have good metrics values, which are close to or better than those of typical supervised MAR methods. The metal artifacts in artifact-affected CT images are eliminated and the tissue structure details are preserved using SLF-CNN. The ablation experiment shows that adding real-artifact CT images greatly improves the generalization ability of the network. The proposed semi-supervised learning method of latent features for MAR effectively suppresses metal artifacts and improves the generalization ability of the network.