Abstract
Multi voltage threshold (MVT) positron emission tomography digitizer uses some discriminators and corresponding time-to-digital converters to record the timing information of fast scintillation pulses and to determinate the event time in a low-power and cost-effective way compared to a traditional ADC-based PET scanner. Linear fitting (LF) algorithm, and quadratic programming (QP) method is proposed methods to Calculate TOF information. We found that both LF and QP algorithms are used to fit linear models, which are not sufficient to describe non-linear random physical phenomena leads to underfitting and accuracy loss. In this paper, we proposed artificial neural networks system with heuristic search strategy to get lower CTR (173 ps), When the two channels take 4 thresholds, the CTR obtained by QP / MVT is 182ps, and the CTR obtained by LF / MVT is 192ps. An intelligent threshold selection method is proposed to determine the thresholds for the MVT digitizer when the hardware resource is limited. We design a novel neural network structure called sparse attention neural network to select optimum threshold values combination. SAnet/MVT has higher accuracy and faster convergence speed than ANN (CTR of 165 ps), which means that lower circuits overhead and higher accuracy.