Single-Scattered events Imaging in ToF-PET via Machine Learning-Based Classification

Scattered coincidences are typically rejected in conventional imaging in Positron Emission Tomography (PET) despite carrying potentially useful spatial information. We present a novel algorithm for PET imaging from single-scattered (inside tissue) events with known Time-of-Flight (TOF) and without energy information, particularly useful for plastic scintillator. The 3D annihilation loci corresponding to individual single-scattered (SS) events are modeled as spindle-torus probability distributions constrained by TOF. These event-wise probability volumes are merged using a MATLAB-based reconstruction framework to estimate the source location. In this study, we used GATE-based Monte Carlo simulations of PET acquisitions to generate realistic emission-event data, including true, SS, and multiple-scattered coincidences. Supervised machine learning classifiers, namely Random Forest (RF) and Extreme Gradient Boosting (XGB), were employed only for the identification of SS events from event-level geometric and timing-related features extracted from the simulated data. The identified SS events were subsequently processed using a dedicated 3D geometric TOF-based reconstruction framework to estimate the annihilation probability distribution and generate SS images. Both RF and XGB achieved SS-event precision of approximately 68-71%, with recall ranging from 56-79% across the evaluated scanner configurations. Using the identified SS events, we created a 3D emission image and validated its quality in terms of resolution, contrast, and uniformity. The proposed framework was also evaluated on the NEMA Image Quality (IQ) phantom under realistic imaging conditions. The results demonstrate that SS events retain meaningful spatial information and that their combination with geometric TOF-based reconstruction provides a feasible pathway for scatter-aware imaging in plastic scintillator PET systems.