Mitigating the Adverse Effect of Compton Scatter on the Positioning of Gamma Interactions in Large Monolithic PET Detectors

In a typical monolithic PET detector setup, scintillation light is captured by an array of photodetectors from which the first interaction position is estimated. This is necessary to draw an accurate line of response. However, a majority of gamma rays undergo one or more Compton interactions before photoelectric interaction. For these events, it is more difficult to recover the first interaction position. In this study we use optical simulation data and neural networks to understand and mitigate the degrading effect of Compton scatter on positioning accuracy. A neural network was trained to predict the 3D first interaction position. Additionally, a network was trained to classify events into three classes: events scattered over a 2D distance smaller than 1 mm (class 0), between 1 mm and 5 mm (class 1) and further than 5 mm (class 2). Finally, a pipeline was designed where events are first classified with the scatter detection network and subsequently discarded (class 2) or positioned with separate networks trained for class 0 and 1. With one neural network trained for all events, an average 3D positioning error of 1.5 mm and FWHM of 0.49 mm is achieved. The scatter detection network achieves an overall accuracy of 65%. Through the combination of scatter detection and separate positioning neural networks for class 0 and 1, the average 3D positioning error reduces with 0.29 mm. Hence, we show that an improvement of about 20% can be achieved through the inclusion of Compton scatter detection. The ultimate goal is to apply the presented methodology to experimental data.