Variable preprocessing applied to neural network position estimators for 2-D PET

The goal of this paper is to study the performance of a series of variable preprocessing methods (namely: selection, scaling and projection) in combination with artificial neural networks (ANNs) to estimate the 2-D position of photon impacts on the surface of a dual head PET detector based on continuous scintillators coupled to multi-anode photomultipliers (MA-PMT). We have used a dynamic readout structure based on weighted sums to code the 64 input signals of the photodetectors into the first 8 moments of the input distribution. The preprocessing methods have been applied to these 8 moments by means of a global optimization guided by a genetic algorithm (GA) that aims to minimize a nonparametric noise estimator called Delta Test (DT). The mean systematic error and mean spatial resolution have been evaluated in all cases. The results show that the lowest systematic error is obtained by the combination of scaling and projection, while the best spatial resolution was achieved with scaling alone.