Low-dose computed tomography (LDCT) denoising is an important topic in medical imaging field. In practice, we aim to preserve the LDCT quality and diagnostic performance while lowering the CT radiation dose. Deep learning-based algorithms such as the convolutional neural network (CNN) and the generative adversarial network (GAN) have recently shown promising results in LDCT denoising. However, limited domain knowledge makes it hard to improve the denoising performance. Instead of manually extracting domain knowledge, we offer a new LDCT denoising scheme and a novel GAN architecture that uses segmentation information as domain knowledge. We demonstrate that domain knowledge from mask LDCT images may be transferred to our denoising GAN for improved LDCT denoising performance. We show that the more precise our semantic segmentation model is, the better our GAN denoising performance is. Furthermore, we posit that the domain knowledge provided by segmentation can come from different datasets, which we refer to as coarse-grained domain knowledge sharing.