Directional Pixelated Gamma-Ray Spectrometer Utilizing Convolutional Neural Network and Simulated Training Data

We have developed a pixelated detector that measures the bearing to a detected source in the limit of 100–1000 events, practically representing just a few seconds of live-streaming data collection. The detector records the event mode intensity and location of energy deposition in the 2-D plane of the detector. We train a convolutional neural network (CNN) against simulated data to process this information into an estimation of the direction to the source, as well as assess background levels and identification of the emitting radioisotope. Corrections that account for optical light sharing between pixels were required to bring simulation and measurement into useful alignment. Laboratory measures are presented, showing that the resulting neural network model provides a satisfactory measure of the angle to the source for a selected range of representative radioisotopes.