Performance study of neural network position estimators for the monolithic scintillator PET detector modules

We have developed a monolithic scintillator PET detector module based on 25.5mm × 25.5mm × 10mm LYSO crystal and multi-anode PMT H7546B, and built an experiment setup to test its performance. The impinging position of a 511 keV photon onto the detector is calculated by pre-trained neural networks using the information of the detected scintillation light distribution. A very attractive advantage of the neural network positioning method is that good resolutions can be achieved even for oblique irradiation. We have carried out experimental tests to verify the effect using the same neural network structure for different incident photon beams. The preliminary processing results show that when the incident angles changes from 0° to 40°, the spatial resolutions (FWHM) changes from 1.82mm to 2.37mm (beam width not subtracted). This means the DOI effect in the PET detectors based on neural network position estimators is obviously eliminated, and the reduced parallax errors in a PET scanner could be expected.