Radiation exposure in positron emission tomography (PET) examination is a major issue for patient safety. PET image quality is severely degraded if low dose of radioactive tracer is administered. With a simultaneous magnetic resonance (MR)-PET scanner, MR anatomical priors can potentially improve PET image reconstruction. In this work, we introduce a framework to synthesize high quality standard dose PET images from the low dose PET and T1 MR images using an atlas guided convolutional neural network (CNN) approach. Compared with the conventional methods, the introduced method demonstrates improved PET image quality for datasets acquired with ten-fold PET dose reduction.