Assessment of the Deep Learning-based Attenuation Correction in the Image Domain for 68Ga-PSMA PET Imaging using Artifact-free Dataset

One of the image quality degradation factors in PET imaging is photon attenuation, wherein the presence of an attenuation map is critical for the task of Attenuation Correction (AC). Different approaches have been proposed to address the lack of standard transmission or CT scans in PET imaging, particularly deep learning-based algorithms to correct attenuation and scatter directly in the image domain without the need for anatomical images. However, artifacts in the PET CT-AC images like the respiration motion, halo, or truncation artifacts would decrease the accuracy of the model training. In this work a novel approach was followed to detect artifact corrupted PET-CT images (where PET-CT images are regarded as reference for model training for the task of AC). Then, the PET-CT images with artifact were removed from the training dataset to create a clean dataset. In the next iteration, the clean dataset was used to train a deep learning model to perform AC in the image domain without requiring anatomical images. The proposed framework was evaluated for AC in whole-body 68Ga-PSMA PET imaging. The artifact-free trained network exhibited lower voxel-wise values of mean error (-0.009±0.43 SUV) and mean absolute error (0.09±0.41 SUV) in comparison with the trained network with unclean dataset. The evaluation on external validation dataset before and after cleaning the training dataset demonstrated the improved AC performance due to the application of the proposed data purification framework. The training of the model before and after cleaning the training dataset demonstrated dramatic improvement in AC of whole-body 68Ga-PSMA PET images in the image domain.