Convolutional neural network-based reconstruction for positronium annihilation localization

Abstract

A novel hermetic detector composed of 200 bismuth germanium oxide crystal scintillators and 393 channel silicon photomultipliers has been developed for positronium (Ps) annihilation studies. This compact 4π detector is capable of simultaneously detecting γ-ray decay in all directions, enabling not only the study of visible and invisible exotic decay processes but also tumor localization in positron emission tomography for small animals. In this study, we investigate the use of a convolutional neural network (CNN) for the localization of Ps annihilation synonymous with tumor localization. Two-γ decay systems of the Ps annihilation from 22Na and 18F radioactive sources are simulated using a GEANT4 simulation. The simulated datasets are preprocessed by applying energy cutoffs. The spatial error in the XY plane from the CNN is compared to that from the classical weighted k-means algorithm centroiding, and the feasibility of CNN-based Ps annihilation reconstruction with tumor localization is discussed.