Deep Convolutional Neural Network for Noise Reduction in Low-Dose CT

To reduce the noise and artifacts in low-dose CT images, a deep convolution network with perceptual loss is proposed. The proposed model obtains a good performance on objective indicators, and makes the denoised images more acceptable to radiologists. The method learns the mapping between low-dose CT images and noise images, and combines batch normalization to accelerate the convergence speed and improve the noise reduction performance. Introducing the perceptual loss extracted by the pre-trained VGG network to effectively solve the problem of aggressive denoising. In addition, the combination of perceptual loss and L1 loss solves the problem of insufficient noise reduction. The experimental result shows that compared with the current advanced algorithms, the proposed network can better balance the noise reduction and the integrity of the organizational structure. After processing, the noise is reduced significantly and the small structure fidelity is perfectly maintained. And the texture of the soft tissue window is more complete which is the difficulty in the field of low-dose CT denoising.