Optimizing Attenuation Correction in 68Ga-PSMA PET Imaging Using Deep Learning and Artifact-Free Dataset Refinement

Background/Objectives: Attenuation correction (AC) is essential for achieving quantitatively accurate PET imaging. In 68Ga-PSMA PET, however, artifacts such as respiratory motion, halo effects, and truncation errors in CT-based AC (CT-AC) images compromise image quality and impair model training for deep learning-based AC. This study proposes a novel artifact-refinement framework that filters out corrupted PET-CT images to create a clean dataset for training an image-domain AC model, eliminating the need for anatomical reference scans. Methods: A residual neural network (ResNet) was trained using paired PET non-AC and PET CT-AC images from a dataset of 828 whole-body 68Ga-PSMA PET-CT scans. An initial model was trained using all data and employed to identify artifact-affected samples via voxel-level error metrics. These outliers were excluded, and the refined dataset was used to retrain the model with an L2 loss function. Performance was evaluated using metrics including mean error (ME), mean absolute error (MAE), relative error (RE%), RMSE, and SSIM on both internal and external test datasets. Results: The model trained with the artifact-free dataset demonstrated significantly improved performance: ME = -0.009 +/- 0.43 SUV, MAE = 0.09 +/- 0.41 SUV, and SSIM = 0.96 +/- 0.03. Compared to the model trained on unfiltered data, the purified data model showed enhanced quantitative accuracy and robustness in external validation. Conclusions: The proposed data purification framework significantly enhances the performance of deep learning-based AC for 68Ga-PSMA PET by mitigating artifact-induced errors. This approach facilitates reliable PET imaging in the absence of anatomical references, advancing clinical applicability and image fidelity.