With the widespread use of computed tomography (CT) technology, the demand for high-quality CT images is increasing, and how to achieve efficient and high-quality image noise reduction has been a hot research topic in the field of CT. In this paper, we focus on the intelligent noise reduction method of CT images based on attention mechanism. In this paper, U-Net is selected as the base network, and the channel attention module and pixel attention module are chosen for the study. To further improve the performance of pixel attention, we propose to introduce multi-scale convolution and residual modules to optimize it. By adding the channel attention and the optimized pixel attention, a lightweight noise reduction network is proposed in this paper. Compared with CBDNet, this network is lighter and faster, and the peak signal-to-noise ratio of noise reduction results is improved by 0.14 dB on average and the model is more stable under the simulated dataset, and the structural similarity of noise reduction results is improved by 0.036% under the real sample dataset.