Previously we showed that respiratory patterns affect the effectiveness of respiratory gated SPECT. This study aims to use a convolutional neural network (CNN) to automatically classify respiration with apnea (AR) and regular respiration (RR) for selecting appropriate rebinning schemes in respiratory gated myocardial perfusion SPECT. We reviewed and classified respiratory traces from 1000 patients tracked from Vicon Motion Systems during their routine Tc-MIBI stress SPECT/CT scans. The traces were first pre-processed and a total of 25,000 data points were used for analysis. All traces were first classified into AR and RR by visual inspection. We randomly chose 700 and 300 processed signals for training and testing our proposed CNN, respectively. Selected SPECT data for the classified patients were then rebinned into 7 respiratory gates using phase and amplitude-based gating accordingly. Estimated motion amplitude, FWHM of the image profiles across the left ventricle with Gaussian fit, and normalized standard deviation (NSD) of a uniform region in lungs in different gates were measured from the two rebinning methods for two classified respiratory traces. Our results show that the binary CNN classification accuracy reaches 88%. The mean estimated motion amplitudes are 0.67 cm and 0.94 cm for phase gating and amplitude gating respectively for RR while they are 0.35 cm and 1.66 cm for AR correspondingly. The FWHM of inferior left ventricle are 13.57 mm and 13.07 mm for phase gating and amplitude gating respectively for RR, while they are 15.58 mm and 13.99 mm for AR. The NSD values are about 40% lower for phase gating as compared to amplitude gating respectively for both traces. The use of CNN can classify AR and RR with high accuracy which can be used to guide the subsequent rebinning schemes. Amplitude-based gating is more suitable for AR patients while RR patients are less sensitive to different gating methods, and phase gating could be used to lower noise.