TY - JOUR T1 - <strong>Whole-body PET Estimation from Ultra-short Scan Durations using 3D Cycle-Consistent Generative Adversarial Networks</strong> JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 247 LP - 247 VL - 60 IS - supplement 1 AU - Xiaofeng Yang AU - Yang Lei AU - Tonghe Wang AU - Xue Dong AU - Kristin Higgins AU - Walter Curran AU - Hui Mao AU - Jonathon Nye Y1 - 2019/05/01 UR - http://jnm.snmjournals.org/content/60/supplement_1/247.abstract N2 - 247Objectives: Oncology PET protocols utilize administered activities ranging between 10-15 mCi and 2-3 min/bed positions to obtain diagnostic quality PET images. Lowering either the administered activity or scan time is desirable as it decreases the patient’s radiation burden or can improve patient comfort (e.g. reduced motion), respectively. But reducing these parameters lowers overall counts and increases noise, adversely impacting image contrast and quantification. To shorten the PET scan duration while maintaining high image quality, we propose a novel 3D cycle-consistent generative adversarial network (GAN) model to estimate diagnostic quality PET images using ultra-short scan durations. Methods: Cycle GANs learn a transformation to synthesize diagnostic PET data that would be indistinguishable from our standard clinical protocol (10mCi, 2.5 min/bed) using ultra-short duration PET data. The algorithm also learns an inverse transformation such that the cycle ultra-short PET data (inverse of the synthetic estimate) generated from synthetic diagnostic PET is close to the real ultra-short PET. In order to optimize the matching between the synthetic and their cycle to their respective real datasets, both transformations are implemented by a generator network and their outputs are judged by a discriminator. Training the generator takes into account the estimation error loss (error between the synthetic and real datasets, and error between cycle and real datasets) as well as the discriminator feedback, which can make the generator output more realistic and improves the discriminator ability to identify the realism. Compared with diagnostic quality PET, an ultra-short PET shares similar anatomical structure but has a higher noise level. Therefore we introduced residual blocks into the generator to catch the differences between PETs at ultra-short durations and the original clinical durations in the training dataset. Finally, the patches from ultra-short scan PET were fed into well-trained transformation model to estimate the synthetic AC PET patches. The diagnostic quality PET was then generated through a patch fusion process. Twenty one subjects with whole-body PET/CT were retrospectively processed to derive diagnostic quality PET dataset from ultra-short PET with 1/8, 1/4 and ½ of the normal scanning time. These data were then compared to the original diagnostic PET study (2-2.5min/bed). This method was assessed with a leave-one-out cross-validation method. Mean absolute percentage error (MAPE) and mean percentage error (MPE) were used to quantify the differences between the synthetic and original PET images. Analysis of covariance and paired-sample t tests were used for statistical comparison between synthetic and original diagnostic PETs. Results: The average MPE and MAPE in whole body were 3.14±4.23% and 15.16±6.54%, 0.95±2.91% and 13.14±5.52%, and 0.69±2.84% and 11.53±5.75% for the 1/8th, 1/4th and 1/2 duration datasets in all 21 patients. There was no significant mean error difference between the synthetic and original diagnostic PETs for all three ultra-short durations. In the selected regions of normal physiologic uptake (brain, lung, heart, kidneys, and liver), the overall MPE was less than 5% and the MAPE was less than 19% averaged over all three ultra-short scan durations. Lesion-based MPE was less than 3% and MAPE were less than 8% average over all three ultra-short scan durations. Conclusion: We developed a novel learning-based approach to accurately estimate diagnostic quality PET datasets from scan durations as short as 19 seconds based on a 3D cycle-consistent GAN with integrated residual blocks. The proposed deep-learning-based approach has great potential to improve ultra-short PET image quality to the level of diagnostic PET used in clinical settings. ER -