Deep Learning for Low-Dose CT Noise Removal Using Dilated Convolution and Perceptual Loss

Low-dose computed tomography has been recommended to reduce the radiation risks of CT scans for patients. However, the reconstructed CT image will be considerably degraded because of photon starvation. Both traditional noise removal techniques and neural networks have been used to enhance the quality of low-dose CT images. In this study, a deep neural network is proposed to mitigate this problem. The network employs dilated convolution, batch normalization, and residual learning. Moreover, a nontrainable edge detection layer is proposed helping to produce sharper edges in the output image without introducing additional complexity. This network is optimized by a combination of mean-square error and perceptual loss to preserve textural details in the CT image that are critical for diagnosis. This objective function solves the over-smoothing problem and grid-like artifacts caused by per-pixel loss and perceptual loss, respectively. The experiments demonstrate the effects of each modification to the network and confirm that the proposed network achieves better performance relative to the state of the art methods.