Abstract
Charge sharing induces spectral and spatial distortions on photon counting detectors which must be corrected using methods such as charge summing circuitry. We propose a method of correction using a spatial-temporal convolutional neural network based on the CycN-Net design. We compare our results to an analytical scalar matrix correction and a U-Net. We show improvements in two energy channels set to 50 and 60 keV with a mean absolute percentage error reduced from 4.84% and 7.46% to 3.95% and 5.14%, respectively, when compared to the scalar matrix approach. We analyze the use of time offset projections and the incorporation of the arbitration counter as a prior in the CycN-Net and show their usefulness for accurate predictions. We also examine the effects of masking and the use of the analytical scalar matrix correction as preprocessing steps for the CycN-Net model. Our results show the potential of utilizing a spatial-temporal CNN approach for correcting charge sharing distortions in higher energy ranges.