Image Style Transfer–Based High- and Low-Frequency Component Synthesis Neural Architecture for Low-Dose Computed Tomography Denoising

As a solution for addressing the radiation risk caused by computed tomography (CT), low-dose computed tomography (LDCT) has led to extensive research because it directly reduces the CT radiation dose to patients. Recently, through mapping of LDCT images to their corresponding full-dose CT images, many deep learning–based methods have been proposed to mitigate noise and artifacts in LDCT images. For LDCT denoising tasks, networks primarily process in the high-frequency components of CT images. An imbalance in learning the high-frequency and low-frequency components may reduce the network’s ability to preserve fine textures.

In this study, inspired by the image style transfer task, we make the first attempt to apply texture transfer to LDCT and propose a neural architecture for high- and low-frequency component synthesis within one CT image, termed HLFCS. HLFCS simultaneously synthesizes high- and low-frequency components of a CT image and balance the associated network parameter updates within a joint optimization framework. In addition, the structural reparameterization methodology is introduced to further improve the inference speed of the final deployed networks. Experiments demonstrate that HLFCS achieves better results with lower computational overhead compared to several state-of-the-art methods, and it exhibits better stability under different noise levels.