As a medical imaging technique, computed tomography (CT) has been widely used in clinical internal visualization, lesion detection and disease tracking. But excessive radiation will cause adverse effects on patients. Lowering the radiation dose can alleviate this problem, but it will lead to the degradation of CT images. In this paper, we propose a knowledge distillation-based denoising method for low-dose CT images. The teacher network trained with higher dose data can generate soft labels for the student network to avoid information loss. The experiments prove that the model trained with our proposed method outperforms the original model in terms of detail preservation.

deep learning,denoising,knowledge distillation,low-dose CT,neural network