FD-GAN: A Dual-Domain Approach with Fourier Domain Discriminators for Denoising Low-Dose CT Images

Low-Dose CT (LDCT) denoising is critical for minimizing radiation exposure while maintaining diagnostic image quality. Conventional model-based and transform domain approaches often result in over-smoothed images with diminished structural details. To overcome these limitations, deep learning techniques have emerged as powerful alternatives, leveraging data-driven models to achieve superior noise suppression and structural preservation. In this work, we propose a novel model named FD-GAN that incorporates dual discriminators operating in both the image and Fourier domains. By jointly optimizing spatial and Fourier domain information, FD-GAN addresses common issues such as over-smoothing, loss of high-frequency details, and training instability observed in traditional Generative Adversarial Network (GAN) models. Our framework integrates pixel-wise reconstruction loss with adversarial and Fourier domain losses to ensure global consistency and local detail enhancement. Extensive experiments on the NIHAAPM-Mayo Clinic LDCT dataset demonstrate that FD-GAN outperforms conventional Mean Squared Error (MSE)-based and single-domain GAN models in both quantitative metrics and visual quality. Furthermore, FD-GAN offers improved stability during training and robustness against artifacts, making it a promising solution for clinical LDCT denoising applications.