Previously we proposed to use a conditional generative adversarial network (cGAN) in denoising low dose static myocardial perfusion SPECT reconstructed images (cGAN-recon). In this study we further compared its performance with projection-based denoising (cGAN-prj). A population of 25 XCAT phantoms was used to model normal and abnormal patients with different anatomies and 99mTc-sestamibi distributions. An analytical projector of a LEHR collimator with attenuation and detector-collimator response modeling was used to simulate 64 projections for 180° with noise level based on a standard 987 MBq injection and 16 min scan time. A series of low dose projections were obtained by scaling the full dose count rate to be 1/20 to 1/2 of the standard. Projections were then reconstructed by OS-EM method with 48 updates. The cGAN was implemented using Tensorflow with the Adam optimizer running up to 400 training epochs. Standard and low dose SPECT projections and reconstructions pairs of 20 phantoms were selected for training the cGAN separately for each noise level, while and projections or reconstructed images from another two phantoms were applied for validation, three other phantoms were used for testing. The denoised projections were then reconstructed for further analysis in normalized mean square error (NMSE), normalized standard deviation (NSD) and structural similarity index (SSIM) along with cGAN-recon, Gaussian and Butterworth filtered images. Results show that artifactual defects are observed for cGAN-recon for lower count levels (≤1/8) while not observed at cGAN-prj. The NMSE, SSIM and NSD values are generally better for cGAN-prj as compared to others. The cGAN denoising improves image quality as compared to conventional post-reconstruction filtering. Denoising based on cGAN-prj is superior to cGAN-recon approach.