Images are one of the important sources of getting information, and the process of getting images can be affected by various factors. Atmospheric turbulence is one of them. Ghost imaging has a positive effect on suppressing atmospheric turbulence, but its reconstruction results are not stable, and it cannot get high-quality images under extreme conditions. In this paper, we simulate atmospheric turbulence using a phase screen, combine computational ghost imaging to simulate the imaging process, and analyze the factors that affect the imaging. We use an end-to-end neural network to input the bucket signal into the network after processing, which can not only reconstruct the target image directly but also save reconstruction time by removing the process of correlation calculation. Simulations show that good reconstruction results can be obtained at low sampling rates and extreme conditions.