Investigation of Noise Reduction in Low-dose SPECT Myocardial Perfusion Images with a Generative Adversarial Network

The purpose of this work was to investigate the possibility of administrated dose reduction while preserving crucial information and clinical value in SPECT-MPI images. In this study, we collected list-mode data from 330 consecutive patients with a dedicated cardiac SPECT scanner using a two-day Tc-99m sestamibi rest/stress protocol at the clinical standard-dose level. A supervised deep learning approach was adopted to predict standard-dose images from 50%, 25%, and 12.5% of the standard-dose level in the projection space. In order to evaluate the proposed deep learning-based framework quantitatively, the peak signal-to-noise ratio (PSNR), root mean squared error (RMSE), and structural similarity index (SSIM) were measured at different dose reduction levels. In addition to the PSNR, SSIM, and RMSE quantitative metrics, we performed a Pearson correlation coefficient analysis on the derived parameters from the QPS package by Cedars-Sinai software. According to the quantitative analysis, reconstructed images at the half-dose level produced the highest PSNR (42.49 ± 2.37) and SSIM (0.99 ± 0.01), as well as the lowest RMSE (1.99 ± 0.63). Analysis of the predicted standard-dose images at the half-dose and quarter-dose revealed that the implemented deep learning model could improve image quality effectively. Although the underlying information in low-dose images beyond the quarter of the standard dose level is not recoverable due to the extremely high noise level.