Low-Dose CT Denoising Using Octave Convolution with High and Low Frequency Bands

Abstract

Low-dose CT denoising has been studied to reduce radiation exposure to patients. Recently, deep learning-based techniques have improved the CT denoising performance, but it is difficult to reflect the characteristics of signals concerning different frequencies properly. Even though high-frequency components play an essential role in denoising, the deep network with a large number of parameters doesn’t concern it and tends to generate the image still having noise and losing the structure. To address this problem, we propose a novel CT denoising method that decomposes high- and low-frequency features and learns more parameters on important features during training. We introduce a network consisting of Octave convolution layers that take feature maps with two frequencies and extract information directly from both maps with inter- and intra-convolutions. The proposed method effectively reduces the noise while maintaining edge sharpness by reducing the spatial redundancy in the network. For evaluation, the 2016 AAPM Low-Dose CT challenge data set was used. The proposed method achieved better performance than the existing CT denoising methods in quantitative and qualitative evaluations. © 2020, Springer Nature Switzerland AG.