Teacher-student Network for CT Image Reconstruction via Meta-learning Strategy

Deep neural networks (DNN) have been widely used in computed tomography (CT) imaging, with promising performance. Meanwhile, most of them are supervised learning strategies, and their performances highly depend on the amount of the pre-collected training samples. In the training, the highdose CT images are usually chosen as labels, but this data is sometimes hard to be collected due to the cancer-risk of highdose CT scanning. Instead, the unlabeled low-dose CT images are easy to access, but they fail to incorporate a large amount of latent information contained into network training. To address these two issues, in this work, we present a couple teacherstudent DNN strategy for low-dose CT image reconstruction via meta-learning strategy. Specifically, this strategy mainly consists of two network, i.e., teacher network and student network. In the teacher network training, only a small amount of samples with high-quality labels (low-dose/high-dose CT image pairs) are included. Then, the unlabeled low-dose CT data are enrolled into this trained teacher network for processing to obtain the temporary high-quality ones. Finally, the unlabeled data with their temporary high-quality ones and another a small number of pre-collected samples with high-quality labels are combined into the student network training. For simplicity, the proposed method is terms as "metaCT", which is similar to the metalearning strategy containing teacher network and student network. Moreover, the present metaCT is fully flexible to adopt the existing DNN-based CT image reconstruction model as the teacher/student network, while the recursive ResNet framework was used in the two network in our work. Experiments on the Mayo clinic dataset demonstrate that the present metaCT method is effective in low-dose CT image reconstruction with a small amount of labeled data and a large amount of unlabeled data.