Generative adversarial network for denoising in dual gated myocardial perfusion SPECT using a population of phantoms and clinical data

Previously we proposed to use a generative adversarial network (GAN) in denoising dual respiratory-cardiac gating (DG) images for myocardial perfusion SPECT. In this study we further compared the use of various training datasets and demonstrated the GAN denoising effectiveness on clinical DG data. Five 4D Extended Cardiac Torso phantoms with cardiac motion, different anatomies, respiratory characteristics and activity uptakes were used in the simulation, modeling 6 respiratory and 8 cardiac gates, i.e., a total of 48 DGs. One hundred and twenty noisy LEHR projections were generated analytically and then reconstructed by the OS-EM algorithm with 6 subsets and 5 iterations. A clinical dataset for a patient who underwent SPECT/CT 1 hr post injection of 1332 MBq Tc-99m sestamibi was re-binned into 7 respiratory and 8 cardiac gates, and then reconstructed by the ML-EM algorithm with 24 iterations. The GAN was implemented using Torch. Using patients' own data, eighteen DG images were paired with the corresponding cardiac gate for training. We also evaluated the use of other patients' datasets for training by increasing the patient database from 1 to 4. The noise level measured as the normalized standard deviation (NSD) on a 2D uniform region of liver and the FWHM on the image profile drawn across the left ventricle wall were compared. In simulations, the NSD/FWHM (cm) of before and after cardiac GAN training were 0.298/1.373 and 0.083/1.489. They were 0.153/3.132, 0.118/2.646, 0.112/1.652 and 0.105/1.375 for training using one to four patients' datasets respectively. The clinical data showed that the use of GAN can lower the noise (NSD 0.246 vs 0.110) with minimal degradation of resolution (FWHM 1.682 vs 1.912). The use of patient's own data for training provide superior denoising results. More phantom and patient data are warranted to confirm our findings.