SCGF-GAN: Spatial-Channel Guided Fusion and Dual-Stream Frequency-Spatial Network for Low-Dose CT Denoising
TL;DR
Experiments demonstrate that SCGF-GAN effectively suppresses complex artifacts while preserving fine structural details, outperforming existing methods in both quantitative metrics and perceptual image quality.
Abstract
Low-dose computed tomography (LDCT) reduces radiation exposure but introduces complex quantum noise and non-stationary streak artifacts, which may compromise the visualization of subtle lesions and diagnostic accuracy. This study aims to develop an effective deep learning approach to suppress artifacts while preserving anatomical details.We propose a spatial–channel guided fusion generative adversarial network (SCGF-GAN). A hybrid spatial–frequency dual-stream block is incorporated into the generator to suppress global streak artifacts in the frequency domain while preserving anatomical edges via a spatial gating branch. A spatial–channel guided fusion (SCGF) module is further designed to constrain quantum noise propagation through channel selection and contextual spatial masking. In addition, a multi-task discriminator and focal frequency loss are employed to jointly enhance local texture fidelity and global distribution consistency.Experiments on the 2016 NIH-AAPM-Mayo Clinic Low-Dose CT Grand Challenge dataset demonstrate that SCGF-GAN effectively suppresses complex artifacts while preserving fine structural details, outperforming existing methods in both quantitative metrics and perceptual image quality.SCGF-GAN provides a promising solution for LDCT Denoising by balancing noise suppression and structural preservation, with potential to improve clinical diagnostic reliability.