Low-dose computed tomography(CT) has attraced more attention due to its prevalence and advantages in reducing the potential radiation risk, while suffering from increased noise. In this paper, we propose an unsupervised low-dose CT denoising method based on Bayesian Neural Network(BNN) to enhance low-dose CT image quality. Different from supervised deep learning, this work only needs a single image, and not requiring massive label data sets for training. On the other hand, all weights in BNN are random variables represented by certain probability distributions, instead of a fixed value in the ordinary neural network. The results on simulated CT data show that the method captures the statistical characteristics of image structure better than the other methods in the sense of structural similarity.

 

Keywords—low-dose CT, denoising, unsupervised deep learning, Bayesian Neural Network.