DDSF-Net: differentiated degradation and structure-guided fusion for low-dose CT denoising

Low-dose CT (LDCT) images are frequently corrupted by complex and spatially non-uniform noise. This not only obscures subtle textures but also blurs structural boundaries, making it inherently difficult to restore details while preserving structures through denoising. This paper presents a novel model for low-dose CT image denoising, which leverages differentiated degradation and structure-guided fusion. The model utilizes the Fourier transform, combined with Gaussian filtering, to achieve smooth spectral decoupling of LDCT images. Additionally, we design a Differentiated Degradation Strategy (DDS) tailored for high- and low-frequency paths. Furthermore, a Structure-Guided Residual Fusion Module (SG-RFM) is proposed. This module employs a channel attention mechanism to adaptively fuse multi-source features—including denoising results and residual information from different frequency bands—under the guidance of structural priors. This approach effectively suppresses artifacts and structural discontinuities in spectral reconstruction. During the frequency-band fusion training stage, a frequency mixing strategy for neighboring slices is introduced to enhance the model’s robustness against frequency band mismatches and local perturbations. The proposed DDSF framework achieves synergistic optimization of texture detail enhancement and structural consistency. Experimental results demonstrate that the proposed method outperforms existing state-of-the-art denoising approaches in terms of PSNR, SSIM, RMSE, and subjective visual quality.