Aiming at the exposure to radiation dosage in CT imaging possessing cancer-inducing potential, and the use of a single data resulting in the residual noise and excessive smoothness of the structure in reconstruction results, a multi-stage joint denoising model for low-dose CT images based on two different data domains is proposed. In the first stage of the generator, the residual U-net model was used to recover the low-dose projection data. and the multi-scale information was added to the up-sampling by embedding skip connections in the coding and decoding process to accelerate the training convergence speed. After noise removal, the projection data was switched from frequency space to spatial domain by means of filtering back projection. In the second stage, multi-scale convolution was used to denoise CT reconstructed images again to enrich the convolution diversity and improve the reconstruction accuracy. VGG network was introduced to capture the perceptual difference between CT images of different doses, and the ability of network representation was enhanced. Experimental results show that the proposed method has a higher PSNR, and compared with the single domain transformation, it can effectively use the complementary effect of the projection data and image data to suppress noise and artifact, and improve the reconstruction effect.