PT - JOURNAL ARTICLE AU - Wu, Yaping AU - Shang, Chong AU - Yuan, Jianmin AU - Fu, Fangfang AU - Meng, Nan AU - Huang, Zhun AU - Feng, Pengyang AU - Xu, Tianyi AU - Feng, Tao AU - Li, Xiaochen AU - Bai, Yan AU - Lin, Yusong AU - Wang, Meiyun TI - Improved short-axies PET image quality via uEXPLORER total-body PET DP - 2021 May 01 TA - Journal of Nuclear Medicine PG - 1527--1527 VI - 62 IP - supplement 1 4099 - http://jnm.snmjournals.org/content/62/supplement_1/1527.short 4100 - http://jnm.snmjournals.org/content/62/supplement_1/1527.full SO - J Nucl Med2021 May 01; 62 AB - 1527Introduction: The uEXPLORER PET scanner has the total-body coverage (1940mm) and ultra-high sensitivity, allows high-quality PET images acquired in a short time. The integrated PET/MR has relative short coverage (320mm), with image quality of short acquisition time difficult to meet the diagnostic requirement. This work attempts to elevate the PET image quality acquired from PET/MR by using machine learning algorithm trained from high quality images collected by uEXPLORER total body scanner. Methods: In total 50 patients were enrolled in this study. Signed informed consent was obtained from each patient before study enrollment. The institutional ethics committee approved this study and all participants gave informed written consent. The PET rawdata collected from a total-body PET/CT scanner (uEXPLORER, UIH, Shanghai, China) in 5 minutes was used for training. PET images reconstructed using the signals from full detector range (1940 mm) were used as high-quality PET images (HQPET), while the images reconstructed by only the signals from 320 mm detector range were considered as low-quality PET images (LQPET). CycleGAN was used to train an image enhancement model to learn the mapping between the paired LQPET and HQPET images through the cyclic loss constraint model. The full convolution neural network (FCN) with residual blocks is used in the generator, and the end-to-end LQPET to HQPET conversion can be realized. In total, the generated paired images from 45 subjects was used as training data, and the images from 5 subjects as test data to evaluate the proposed algorithm. The difference between the proposed method is compared with NLM and pix2pix methods. Results: The normalized root mean squared error (NRMSE), peak signal-to-noise ratio (PSNR), and structural similarity (SSIM) were used to quantify the enhancement effect of the algorithm. The mean values of NRMSE, PSNR, and SSIM of images generated from the test set and original images were (0.211, 37.23dB, 0.969) and (0.284, 36.93dB, 0.955), respectively. Compared with NLM method, the proposed method shows higher image quality, lower noise, and artifacts. The proposed method is superior to pix2pix in preserving edges and details. Conclusions: This work indicates that the CycleGAN model can effectively improve the PET image quality (Figure 1). It has potential to enhance short-axis PET images. Acknowledgments: This paper is supported by the National Key R&D Program of China (2017YFE0103600), National Natural Science Foundation of China (81720108021), Zhongyuan Thousand Talents Plan Project (ZYQR201810117), Zhengzhou Collaborative Innovation Major Project (20XTZX05015).