Distillation guided deep unfolding network with frequency hierarchical regularization for low-dose CT image denoising

Linlin Yang, Hongying Liu, Yuanyuan Liu, Fanhua Shang, Liang Wan, Licheng Jiao

Abstract: Computed tomography (CT) is an important means of medical diagnosis. Low-dose CT (LDCT) avoids the problem of high incidence of disease by reducing the radiation dose, but reduces the quality of images. Deep unfolding networks are used to improve LDCT image quality because of its excellent performance and theoretical interpretability. However, the training of the network is usually based on paired LDCT-NDCT images, and the input image is processed iteratively and independently stage-by-stage. There are problems of accumulation of information distortion and lack of feature interaction between different stages. To solve these problems, we propose a distillation guided deep unfolding network with frequency hierarchical regularization (FHRDUN). It uses the pre-trained image denoising network as the teacher network, and adopts the initially clean image attained by distilling teacher network along with the LDCT images to supervise the training of student network. We also design the noise estimation module (NUNet) and frequency hierarchical regularization module (FHRM). NUNet is devised for initial noise removal and to synthesize LDCT images with the denoised images obtained by the regularization module. FHRM achieves the information exchange across stages and the regularization of frequency-division subimage. Extensive experiments on Mayo, Piglet and Kaggle datasets demonstrate superior performance of FHRDUN compared with state-of-the-art methods. Compared with existing methods, the PSNR results for four different doses are increased on average by 0.28 dB on the Piglet dataset.