Low-dose CT (LDCT) can significantly reduce health risks associated with radiation exposure compared to normal-dose CT (NDCT). However, the lower radiation dose may result in projection data being contaminated by noise, which can hinder the accurate identification of lesion details. Currently, most LDCT image denoising techniques employ supervised learning methods that rely on paired noisy and noise-free datasets for model training. In practical applications, however, obtaining such paired data is often challenging. To address this issue, we propose an unsupervised low-dose CT denoising method called MANet-CycleGAN, which can train a high-quality denoising model via unpaired data. Our design approach is as follows: 1. Eliminate the dependency of the denoising model training on paired data through a cyclic generative adversarial network architecture; 2. Apply the UNet architecture to generator for feature extraction and NDCT image generation, while using a PatchGAN discriminator to enhance the details of the generated images; 3. Introduce channel attention and multi-scale feature extraction capabilities through the Squeeze-and-Excitation (SE) module, Efficient Channel Attention (ECA), and Atrous Spatial Pyramid Pooling (ASPP) to improve image generation quality; 4. Utilize perceptual loss in training process to better preserve the structural features of the image while denoising. We conducted comparative experiments on the Mayo Clinic LDCT Grand Challenge dataset. The results demonstrate that the proposed method outperforms existing methods in both qualitative and quantitative aspects.