PET-QA-NET Training

Non-ASC PET images were input to the DL model to generate direct ASC PET images (using CT-ASC PET with PSF + TOF as reference). Additional information on image preprocessing and the network is provided in the supplemental section and Supplemental Figure 1 (http://links.lww.com/CNM/A441). The primary training process was performed as training/validation (669/100 patients) on the HUG data set. Two tests from HUG were used for further evaluation: a clean test set (100 patients) and a test with artifacts (1218 patients). Because of the high variability across the different centers, we used transfer learning with fine-tuning in the 8 different centers (20% for fine-tuning and 80% for the test set).