Multidimensional noise reduction in C-arm cone-beam CT via 2D-based Landweber iteration and 3D-based deep neural networks

Recently, the necessity of using low-dose CT imaging with reduced noise has come to the forefront due to the risks involved in radiation. In order to acquire a high-resolution image from a low-resolution image which produces a relatively small amount of radiation, various algorithms including deep learning-based methods have been proposed. However, the current techniques have shown limited performance, especially with regard to losing fine details and blurring high-frequency edges. To enhance the previously suggested 2D patch-based denoising model, we have suggested the 3D block-based REDCNN model, employing convolution layers paired with deconvolution layers, shortcuts, and residual mappings. This process allows us to preserve the image structure and diagnostic features of an image, increasing image resolution by smoothing noise. Finally, we applied a bilateral filter in 3D and utilized a 2D-based Landweber iteration method to reduce remaining noise under a certain amplitude and prevent the edges from blurring. As a result, our proposed method effectively reduced Poisson noise level without losing diagnostic features and showed high performance in both qualitative and quantitative evaluation methods compared to ResNet2D, ResNet3D, REDCNN2D, and REDCNN3D.