Initial Investigation of Using a Generative Adversarial Network for Denoising in Dual Gating Myocardial Perfusion SPECT

Abstract

Dual respiratory-cardiac gating (DG) freezes myocardial perfusion SPECT by reducing both cardiac and respiratory motion. However, detected counts are divided into more bins in DG as compared to sole cardiac or respiratory gating, leading to further noise amplification. This study aims to reduce the noise level in each DG bin using a generative adversarial network (GAN). We used the XCAT phantom to generate a Tc-99m-MIBI DG dataset with cardiac and respiratory motion. A 5-s respiratory and a 1-s cardiac cycle were divided into 24 and 48 frames respectively, and then grouped to 6 respiratory and 8 cardiac gates, i.e., a total of 48 gates. We used an analytical projector to simulate a LEHR collimator with 120 realistic noisy projections over 180°, which were reconstructed by OS-EM using average CT for attenuation correction. Eighteen DG bins, i.e., a total of 2052 images (18×114 axial slices), were paired with the corresponding single cardiac gate for training. Six other DGs of the same cardiac gate were tested by the GAN, which consisted of a 16-layer generator and a 5-layer discriminator implemented on a NVIDIA GTX 1080 GPU. The DG images w/ and w/o denoising were then registered to the end-expiration gate and summed to a cardiac image. The noise level, measured as the normalized standard deviation (NSD) on a 2D uniform region on the liver, was compared for a single cardiac gate (C8) and the registered cardiac image w/o (C8-REG) and w/ (C8-GAN) denoising. The NSD was 0.219, 0.229 and 0.093 for C8, C8-REG and C8-GAN respectively. The noise level substantially reduced for the DG images when GAN was applied, with minimal increase of motion blurring. The GAN has potential to improve the DG performance thus clinical feasibility, and further evaluation on patient data is warranted to validate its actual clinical performance.