Unpaired low-dose CT denoising via an improved cycle-consistent adversarial network with attention ensemble

Many deep learning-based approaches have been authenticated well performed for low-dose computed tomography (LDCT) image postprocessing. Unfortunately, most of them highly depend on well-paired datasets, which are difficult to acquire in clinical practice. Therefore, we propose an improved cycle-consistent adversarial networks (CycleGAN) to improve the quality of LDCT images. We employ a UNet-based network with attention gates ensembled as the generator, which could adaptively stress salient features which is useful for the denoising task. By doing so, the proposed network could enable the decoder to acquire available semantic features from the encoder with emphasis, thereby improving its performance. Then, perceptual loss found on the visual geometry group (VGG) is drawn into the cycle consistency loss to elevate the visual effect of denoised images to that of standard-dose computed tomography images as far as possible. Moreover, we raise an ameliorative adversarial loss based on the least square loss. In particular, the Lipschitz constraint is added to the objective function of the discriminator, while total variation is added to that of the generator, to further enhance the denoising capability of the network. The proposed method is trained and tested on a public dataset named 'Lung-PET-CT-Dx' and a real clinical dataset. Results show that the proposed method outperforms the comparative methods and even performs comparably results to that of an approach based on paired datasets in terms of quantitative scores and visual sense.