Classification of Signal Events for Entanglement Studies with J-PET Using Machine Learning Techniques

Abstract

Understanding the entangled polarization of photons produced in positron–electron annihilation remains experimentally challenging. Recently, it has been proposed to utilize this correlation in positron emission tomography imaging to suppress the random background and as a diagnostic parameter. Conventional positron emission tomographs rely on the registration of the two annihilation photons from positron annihilation. However, measuring polarization correlations requires scattering of annihilation photons in the detector volume and the registration of their corresponding scatters. Such a 4-hit event topology inherently suffers from low efficiency and high background. Therefore, translating this quantum information into medical imaging requires a method to overcome the inherent limitations of conventional geometric cuts. We present a machine learning classification framework trained on simulated detector events. This approach enables a high-purity separation of polarization-correlated signal from background and simultaneously achieves a significant improvement in computational efficiency.