CaGAN: a Cycle-consistent Generative Adversarial Network with Attention for Low-Dose CT Imaging

Although lowering X-ray radiation helps reduce the risk of cancer in patients, low-dose computed tomography (LDCT) technology usually leads to poor image quality, such as amplified mottle noise and streak artifacts, which severely impact the diagnostic results. To improve diagnostic performance, we propose an algorithm based on a cycle-consistent generative adversarial network (CycleGAN) to suppress noise and reduce artifacts. In addition, we include attention mechanisms in the proposed network to expand the receptive field and capture richer contextual dependencies. Unlike traditional methods that manually match similar local blocks, our proposed method can autonomously learn the relationship between local features and their global dependencies. Specifically, two different types of attention modules (criss-cross self-attention (CCSA) and channel attention (CA)) are adopted to enhance feature interdependencies in the spatial and channel dimensions separately. Because of the CCSA mechanism, noise and artifacts can be restored using cues from all local feature locations; the CA mechanism adaptively reassigns the weights of each feature map. Furthermore, we also performed a diagnostic quality assessment of the results and ablation studies of the loss functions and the structural modules, which showed the validity of our proposed method. Extensive experiments show that our proposed method achieves better metrics and visual effects than state-of-the-art methods.