Noise reduction in low-dose CT with stacked sparse denoising autoencoders

Many approaches have been proposed to improve the quality of low-dose CT images from noisy projections. These approaches can be categorized into three groups: Sinogram filtering approaches, iterative reconstruction approaches and post-reconstruction restoration approaches. Sinogram filtering approaches directly smooth the raw data before filtered back-projection (FBP) is applied. Iterative reconstruction approaches optimize a prior-regularized objective function by iterative ways. Despite the successes achieved by these two kinds of approaches, researchers on these approaches are often limited in practice due to the difficulty of gaining well-formatted projection data from the commercial CT scanner. The post-reconstruction restoration approaches, which don't rely on the projection data, can be directly applied on low-dose CT and easily integrated into the current CT systems [1]. In this summary, we will focus on the third group. Inspired by the superior performance achieved by non-linear deep neural networks in the field of image processing, a stacked sparse denoising autoencoders (SSDA) based noise reduction method for low-dose CT imaging is presented in this summary. The experiments demonstrate the feasibility and effectiveness of our method.