Ensemble of neural networks for 3D position estimation in monolithic PET detectors

We propose an ensemble of multilayer feedforward neural networks to estimate the 3D position of photoelectric interactions in monolithic detectors. The ensemble is trained with data generated from optical Monte Carlo simulations only. The originality of our approach is to exploit simulations to obtain reference data, in combination with a variability reduction that the network ensembles offer, thus, removing the need of extensive per-detector calibration measurements. This procedure delivers an ensemble valid for any detector of the same design. We show the capability of the ensemble to solve the 3D positioning problem through testing four different detector designs with Monte Carlo data, measurements from physical detectors and reconstructed images from the MindView scanner. Network ensembles allow the detector to achieve a 2-2.4 mm FWHM, depending on its design, and the associated reconstructed images present improved SNR, CNR and SSIM when compared to those based on the MindView built-in positioning algorithm.