Time Residual Estimation using Artificial Intelligence

A deep-learning based method was implemented to estimate time-residuals (TR) that is performed as part of daily quality control for Biograph Vision 600 high time of flight (TOF) resolution PET/CT scanner. The purpose of this work is to investigate the feasibility of utilizing deep learning to estimate TRs since it is computationally expensive operation due to requirement of non-TOF reconstruction and TOF forward projection and iterative algorithm to estimate TRs. Cylinder scans at 5 different locations were performed to cover the 1 cm radial tolerance from center permitted during positioning of the cylinder. A convolutional neural network with residual connections is utilized. Time-residual ranging from -10 time bins to +10 time bins were simulated to create training, validation, and test data. Time residual error is zero for 38.84% of test crystals, within +/1 time-bins for 87.70 % of test crystals, within +/-2 time-bins for 98.76% of test crystals, within +/-3 time-bins for 99.89, and all the remaining crystals (50 out of 121600) are within +/-7 which is satisfactory for daily QC purposes.