Compton Decomposition and Recovery in a Prism-PET Detector Module

Identifying the correct Line-of-Response (LOR) in positron emission tomography (PET) requires accurate localization of the first interaction between incident gamma ray and detector. Improving the accuracy of this localization typically entails offsetting losses in detector efficiency, complexity, or cost. In this paper, we propose a solution that can localize scattered gammas without losses in detector efficiency, utilizing Prism-PET - a single-sided detector module with pixelated light guide. Using Monte Carlo simulations of gamma-detector interactions, we train a Convolutional Neural Network to predict the first interaction position of gammas incident on a detector block.