Accuracy of deep learning-based attenuation correction in 99mTc-GSA SPECT/CT hepatic imaging

Abstract

Objective The aim of this study was to generate pseudo CT images for attenuation correction (AC) from non-AC SPECT images and evaluate the accuracy of deep learning-based AC in ^99mTc-labeled galactosyl human serum albumin (^99mTc-GSA) SPECT/CT hepatic imaging. Methods A cycle-consistent generative network (CycleGAN) was used to generate pseudo CT images of 40 patients with normal liver function. The test cohort consisted of one patient with normal liver function and one patient with abnormal liver function. SPECT images were reconstructed without AC (SPECT_NC), with conventional CTAC (SPECT_CTAC), and with deep learning-based AC using pseudo CT images (SPECT_GAN). The accuracy of each AC method was evaluated using the total liver count and the structural similarity index (SSIM) of SPECT_CTAC and SPECT_GAN. The coefficient of variation (%CV) was used to assess uniformity. Results The total liver counts in SPECT_GAN were significantly improved over those in SPECT_NC and differed from those of SPECT_CTAC by approximately 7% in both patients. The %CV values in SPECT_CTAC and SPECT_GAN were significantly lower than those in SPECT_NC. The mean SSIM values in SPECT_CTAC and SPECT_GAN for patients with normal and abnormal liver functions were 0.985 ± 0.00189 and 0.977 ± 0.00191, respectively. Conclusions The accuracy of AC with a deep learning-based method was similarly performed as the conventional CTAC method. Our proposed method used only non-AC SPECT images for AC, which has great potential to reduce patient exposure by eliminating real CT examination.