Abstract
3218
Introduction: Attenuation and scatter correction (ASC) are two main crucial steps of PET imaging toward quantitative imaging. Artifacts in PET imaging befall in different areas including PET images (halo artifact), inference between PET images and anatomical (CT and MRI) images (mismatch, misregistration and motion), and CT/MR artifact which propagate to PET images (metal, truncation). In the current study, we propose an emission-based ASC algorithm which would be immune to the above mentioned artifacts. The main aim is detection and correction of artifacts in PET/CT and PET/MRI scanners.
Methods: In this study, we enrolled 313 artifact-free clinical PET/CT images, where ground truth ASC images were generated using CT-based ASC. The datasets were randomly split into 80% training/validation and 20% test sets. All PET images were converted to SUV units and normalized by an empirical value of 9. We implemented a modified attention residual U-Net architecture to map non-ASC images used as input to CT-based ASC as target. For model evaluation, voxel-wise mean error (ME), mean absolute error (MAE), relative error (RE%), absolute relative error (ARE%), and structural similarity index (SSIM) were calculated. In addition to testing the network on clean data set, we evaluated the network performance in the presence of artifacts. To this end, we included patients with metal artifacts, halo artifacts, motion artifacts and truncation artifacts and inspected visually the efficiency of the proposed algorithms in dealing with different artifacts.
Results: We achieved ME of 0.02±0.05, MAE of 0.20±0.07, RE of -1.32±2.5%, ARE of 10.0±4.5% and SSIM of 0. 98±0.01 in the hold out test set. We reported different artifacts in the original CT-ASC images and corrections made using the proposed algorithms. The Proposed method could effectively disentangle different PET image artifacts including halo, metal, truncation, motion, and mismatch.
Conclusions:
We developed a deep learning-based algorithm to correct halo, motion, mismatch, metal and truncation artifacts in PET/CT and PET/MRI. The proposed algorithm could be used as effective and fast quality assurance tool to routinely detect and correct PET image artifacts in clinical setting.