Event Classification in Heterostructured Scintillators With Limited Readout Information Using Neural Networks

To improve coincidence time resolution (CTR) in time-of-flight positron emission tomography (TOF-PET), various approaches have been explored, including the use of novel materials like heterostructured scintillators. These scintillators combine different materials with complementary properties like Bismuth Germanate for its high detection efficiency and EJ232 for fast timing. By layering these materials on a micrometer scale, energy sharing between them becomes possible, enabling fast timing, while maintaining high detection efficiency. For TOF-PET applications, scalable electronics are essential. While earlier models characterized heterostructured scintillators in analog, single-pixel setups, the digital and scalable systems required for full positron emission tomography (PET) scanners present additional challenges due to increased signal complexity. In this study, we explored neural networks to characterize heterostructured scintillators using parameters available in scalable systems. We trained one neural network to identify photoelectric events and another one to estimate the amount of energy sharing between the two materials. The method demonstrated promising results using multiple combinations of the aforementioned parameters, with prediction accuracy for photoelectric events ranging from 91.6% to 96.8%, and a mean average error in the energy sharing estimation between 7.7 and 43.9 keV. This suggests the potential application of heterostructured scintillators in scalable readout electronics for full TOF-PET systems.