Low-Dose CT Image Blind Denoising with Graph Convolutional Networks

Convolutional Neural Networks (CNNs) have been widely applied to the Low-Dose Computed Tomography (LDCT) image denoising problem. While most existing methods aim to explore the local self-similarity of the synthetic noisy CT image by injecting Poisson noise to the clean data, we argue that it may not be optimal as the noise of real-world LDCT image can be quite different compared with synthetic noise (e.g., Poisson noise). To address these issues, instead of manually distorting the clean CT to construct paired training set, we estimate the noise distribution over the real-world LDCT images firstly and then generate noise samples through Generative Adversarial Network (GAN) such that a paired LDCT image dataset can be constructed. To explore the non-local self-similarity of LDCT images, Graph Convolutional Layers (GCLs) is utilized to obtain the non-local patterns of LDCT images. Experiments were performed using real-world LDCT image dataset and the proposed method achieves much better performance compared with other approaches with respect to both quantitative and visual results.