Unsupervised learning-based deep sparsifying transform network for joint CT metal artifact reduction and super-resolution reconstruction

Abstract

Due to the presence of metallic implants, metal artifacts degrade the quality of computed tomography (CT) images. Existing deep learning-based metal artifact reduction (DL-based MAR) methods rely on paired datasets, i.e., the ground truth CT image and its metal artifact corrupted version, for training. However, it is difficult to obtain paired datasets in clinical scenarios. Unsupervised learning-based MAR algorithms can use unpaired datasets to address the difficulty of obtaining paired data, but still face the following limitations: (i) most MAR network architectures lack interpretability and exhibit redundant learnable parameters, due to their empirical design nature; (ii) the representation ability of existing encoder-decoder-based network architectures is limited, and they often ignore the image resolution. To overcome these limitations, we introduce an unsupervised learning-based deep sparsifying transform network, dubbed UnDeepST, which is designed for the reconstruction of CT images with both metal artifact reduction (MAR) and super-resolution (SR) capabilities. UnDeepST is model-interpretable and has a smaller number of learnable parameters due to less recycling use of encoder and decoders, compared to previous unsupervised learning-based MAR methods. Furthermore, we design a task fusion module to assist MAR with the help of SR to reconstruct high-quality and high-resolution CT images. To the best of our knowledge, we are the first to merge the MAR and SR tasks to achieve mutual learning of information across different tasks. By designing various loss functions, UnDeepST can be trained on unpaired datasets in an end-to-end training manner. Experimental results demonstrate that UnDeepST can achieve competitive recovery quality and resolution compared to benchmark algorithms.