Dynamic Threshold Discriminator - Mutual Learning for Scalable Time Stamping Neural Networks

The precision of emission image estimation relies on accurate interaction time estimation, which is becoming increasingly possible with neural network-based time stamping methods. In order to expand the use of these methods to multiple detector systems and facilitate mutual training, we leverage the inherent variance-bias characteristics of scintillator detectors. Our neural network, called the Dynamic Threshold Discriminator (DTD), is designed to estimate appropriate thresholds for given waveforms rather than estimating time stamps directly to exploit the characteristic. We assessed the feasibility of DTD using experimental data and compared it to classical time stamping methods (LED, CFD, LED with energy correction) and a neural network-based time stamping method. The proposed method showed a 12% lower coincidence timing resolution compared to LED with energy correction for a 500 keV energy window even without energy information and demonstrated robust performance regardless of the digitizer's sampling rate.