Application of neural networks to the identification of the compton interaction sequence in compton imagers

Compton cameras are well suited to image photons from a few hundred keV up to several MeV. However, one important data analysis step presents a significant challenge: the reconstruction of the Compton interaction sequence (event reconstruction). We present a new approach to event reconstruction based on a multi-layer perceptron neural network with a sigmoid activation function using back-propagation as the learning approach. Simulations show that this new method outperforms the classic event reconstruction approach and achieves roughly the same performance as the Bayesian approach to event reconstruction for events with two interactions and exceeds its performance for events with three interactions.