Computed tomography (CT) has been widely used in modern medical diagnosis and treatment. However, ionizing radiation of CT for a large population of patients becomes a concern. Low-dose CT is actively pursued to reduce harmful radiation, but faces challenges of elevated noise in images. To address this problem and improve low-dose CT image quality, we develop an image-domain denoising method based on cycle-consistent adversarial networks (CycleGAN). Different from previous deep learning based denoising methods, CycleGAN can learn data distribution of organ structures from unpaired full-dose and low-dose images, i.e. there is no one-to-one correspondence between full-dose and low-dose images. This is an important development of learning-based methods for low-dose CT since it enables the model growth using previously acquired full-dose images and later acquired low-dose images from different patients. As a proof-of-concept study, we used the NIH-AAPM-Mayo Clinic Low Dose CT Grand Challenge data to test our CycleGAN denoising method. The results show that the proposed method not only achieves better peak signal-to-noise ratio (PSNR) for quarter-dose images than non-local mean and dictionary learning denoising methods, but also preserves more details reflected by images and structural similarity index (SSIM). Our investigation also reveals that a larger sample size leads to a better denoising performance for CycleGAN.