Event Localization in Continuous Crystal Scintillation Cameras using Distribution Matching Neural Networks

The localization of gamma interactions on solid scintillation crystals is normally estimated from the light distribution of the crystal's scintillation. The estimation of the interaction position is not an error nor bias free process, mostly due to light reflections within the crystal. Complex models exist to reduce this effect. These methods often rely on calibration measurements performed with collimated beams along the crystal's surface, making the process slow and troublesome. This paper presents a method to improve the interaction position estimation based on neural networks and an interaction distribution matching loss. The method requires only a flood acquisition with known interaction distribution. The neural network does not need a dataset with matched "light distribution-interaction position" data. The method is tested using an experimental acquisition and Monte Carlo simulation of a large scintillation camera composed of a 406.4 × 304.8 × 25.4 mm
³
continuous NaI(TI) crystal. The method improves the interaction localization and reduce the edge packing effects of the center of gravity algorithm, increasing the detector's effective area from 48.7% to 72.1%. Moreover the method is able to estimate depth of iteration.