End-to-End Unsupervised Learning for Direct Attenuation and Scatter Correction of Whole-body 18F-FDG PET Images Using Cycle GAN

Introduction: PET is an imaging modality of choice for quantitative assessment of numerous physiological and biological processes, providing valuable biomarkers of disease. Meanwhile, Compton scattering degrades the quality and quantitative accuracy of PET images via loss of events along lines-of-response (‘attenuation’) and misleading measured counts indicating incorrect events (‘scatter’). Attenuation and scattered correction, though routine in PET/CT imaging, remain challenging in PET-only and PET/MRI scanners. The present study aims to develop a fully automated unsupervised cycle generative adversarial network (GAN) model to predict attenuation and scatter corrected (ASC) PET images directly from non- attention and scatter corrected images (NASC).

Methods: Images of 122 patients administered with 391 ± 50.11MBq of 18F-FDG before undergoing PET/CT scan were enrolled in this study. All PET images reconstructed by using the ordinary Poisson ordered subsets-expectation maximization algorithm (OP-OSEM, with 2 iterations and 21 subsets) followed by a 5-mm FWHM Gaussian post-reconstruction smoothing. The voxel intensities of images were converted to standardized uptake values (SUVs) to reduce the range of the intensity of PET images. PETCTASC and Non-ASC image were divided by 9 and 3 (based on all images histogram median), respectively to further reduce the dynamic range of the voxel intensities with range of 0~1. The proposed Cycle GAN benefits from an unsupervised loss function term for handling misregistered images, anatomical variations among patients, and inter-and intra- imaging sessions. The cycle GAN has two cycles, called forward and backward cycles. In each cycle, residual block-based generator maps the NASC image into ASC domain and discriminator calculates the distance between generated and target images. In the forward cycle, images are translated to ASC domain. To ensure accurate translation, the generated images from ASC domain are translated back to NASC domain during the backward cycle. Comprehensive image quality metrics including mean absolute error (MAE), absolute relative error (ARE), relative error (RE), root mean squared error (RMSE), mean squared error (MSE), peak signal to noise ratio (PSNR), structural similarity index (SSIM), and Frechet inception distance (FID) were calculated for 22 patients as a holdout test set.

Results: The cycle GAN achieved MAE of 0.15 ± 0.06 on the test set. ARE (%) and RE (%) were 16.62 ± 4.65 and 5.20 ± 12.17, respectively. The RMSE and MSE for predicted images of cycle GAN were 0.19 ± 0.07 and 0.04 ± 0.03, respectively. The cycle GAN achieved 31.56 ± 1.41 PSNR, 0.96 ± 0.01 SSIM, and 0.31 ± 0.33 FID. The cycle GAN needed roughly 5 seconds to predict the corrected images. We also show outlier cases where the cycle GAN failed to accomplish a good performance, requiring further considerations. Figures 1 and 2 depict a generated image and outlier case, respectively, compared with their non-ASC, CT-ASC, and difference map.

Conclusions: The proposed cycle GAN model was promising for simultaneous end-to-end attenuation and Compton scattering correction of PET images, which is suitable for prospective clinical use due to the fast and accurate estimation of corrected images. The proposed algorithm has the potential to address ASC in PET-only and PET/MRI scanners.

• Download figure
• Open in new tab
• Download powerpoint