Deep Learning-based Low-dose Cardiac Gated SPECT: Implementation in Projection Space vs. Image Space

Abstract

The reduction of radiation exposure in SPECT-MPI is an important research topic. However, lowering the injected activity degrades image quality, thus impacting the diagnostic accuracy of this modality. In this study, we enrolled a total of 335 clinical gated SPECT-MPI images from a dedicated cardiac SPECT scanner acquired in list-mode format. All patients underwent a two-day rest/stress protocol and the obtained gated images were retrospectively used to convert low-dose to standard-dose images in both projection and image spaces. A deep generative adversarial network was employed to predict standard-dose images from 50% low-dose images. The proposed network was evaluated using quantitative metrics, such as the peak signal-to-noise ratio (PSNR), root mean squared error (RMSE), and structural similarity index measure (SSIM). Moreover, a Pearson correlation coefficient analysis was performed on the half-dose and predicted standard-dose images with respect to the reference standard-dose images. The results demonstrated that the highest PSNR (46.30 ± 2.23) and SSIM (0.98 ± 0.01), and the lowest RMSE (1.32 ± 0.54) were obtained from the image space implementation. Pearson analysis showed that the predicted standard-dose images yielded ρ = 0.960 ± 0.011 and ρ = 0.947 ± 0.027 in the image and projection spaces, respectively. Overall, considering the quantitative metrics, the noise was effectively suppressed in the predicted standard-dose images for both implementations. Yet, standard-dose image estimation in the image space resulted in superior quantitative accuracy and image quality.