Neural Network-based Inter-crystal Scatter Event Positioning in a PET System Design Based on 3D Position Sensitive Detectors

We demonstrate a simulation-based proof-of-concept for superior inter-crystal scatter (ICS) event positioning using a feed-forward neural network (NN) architecture compared to traditional winner-takes-all (WTA) and loser-takes-all (LTA) algorithms. Using a GATE Monte Carlo simulation of a 3D position-sensitive scintillation detector module comprising long crystals read out from the side using SiPMs with 3×3×3 mm3 effective detector voxels, we observe NN ICS event positioning accuracies of 0.753 to 0.680 when the number of interactions per annihilation photon ranges from 2 to 5: significantly more robust compared to 0.726 to 0.367 for LTA and 0.613 to 0.251 for WTA methods over the same range. We then scale the single-detector simulation into a 25 cm diameter PET brain imaging system and reconstruct contrast and resolution phantoms for image quality analysis. The NN model outperformed both WTA and LTA, with image normalized Mean Absolute Errors of 0.030 and 0.122 for contrast and resolution phantoms compared to 0.046, 0.178 and 0.034, 0.140 for WTA and LTA. The NN demonstrated 6.04 to 8.95% higher Contrast Recovery (from resolution phantom), 0.53 to 2.85% larger Contrast Noise Ratio (from contrast phantom), and 2.13 to 6.34% higher Modulation Transfer Function values (from resolution phantom) compared to LTA, which performed second-best. The upper bound for these NN relative improvements occurred with features near the spatial resolution limit of the simulated system (2 mm). Our results indicate the NN positioning approach we examined improves most image quality and quantitation figures of merit.