There is a risk of radiation exposure in the signal acquisition process of PET images. Under the premise of ensuring the image quality, it is necessary to reduce the tracer dosage as much as possible. But this will cause artifacts in the PET image and lower signal-to-noise ratio. We propose a conditional generative adversarial network (cGAN) based on multi-scale contextual information fusion, which is used to optimize PET image reconstruction to improve the image quality of PET images. By redesigning the skip connection of encoder-decoder network as generator, we proposes a skip connection based on multi-scale context information fusion, which enables the decoder to obtain richer semantic features from the encoder during the decoding process. We adopt the strategy of making the generator network learn the noise distribution of low-dose PET images, which reduces the learning difficulty of the network. The proposed network is evaluated in the low-dose PET dataset. The peak signal-to-noise ratio was 29.948 ± 4.062, the structural similarity index was 0.926 ± 0.030, and the normalized root mean square error was 0.395 ± 0.211. Compared with the traditional denoising algorithm Non-Local Mean (NLM) and block-matching 3D (BM3D) and two deep learning methods RED-CNN and cGAN (U-Net is the generator), the proposed network is superior to the above four methods.