We evaluate the performance of monolithic positron emission tomography (PET) detector elements depending on its scintillator crystal plate thickness (6 and 12 mm) and the surface finish (rough or polished) using a neural network to reconstruct events. A GEANT4 PET detector model was used for this study. It consisted of a LYSO crystal with a 57.6x 57.6 mm face and a 64-channel Sensl ARRAYC-60035-64P-PCB photomultiplier. Separate runs were made with varying crystal parameters: thickness (6 and 12 mm) and surface finish (rough and polished), resulting in four separate event pools. A feed-forward neural network was used to reconstruct the point of 511 keV gamma interaction. The number of layers and neurons per layer were varied. The best resolution was achieved with a 6 mm thick detector with a rough finish with an average of 0.57± 0.01 mm for the XY plane and an average 0.89± 0.01 mm for the Z coordinate (depth of interaction), and a dR of 1.19 ± 0.01 mm. Keywords: scintillating crystals, gamma radiation, positron emission tomography, neural networks. DOI: 10.21883/0000000000