Multi-Stage Progressive Generative Adversarial Network for Low-Dose CT Denoising

To address the balance between spatial details and high-level contextual information in low-dose CT image denoising tasks, a multi-stage progressive generative adversarial network is introduced. This network learns features from CT images in three phases, transitioning from coarse to fine. Wavelet transforms are employed in the initial two phases to expand the perceptual domain of the image. Subsequently, a dense residual network enhances the expression of image details in the third phase. This integration of information across stages facilitates the capture of multi-scale information and abstract features. Additionally, a discriminator equipped with multi-scale re-sidual modules is designed to improve noise artifact discrimination. Experimental results demonstrate significant outperformance of our proposed network over existing algorithms in denoising performance, both quantitatively and qualitatively.