Exploring Various Neural Networks for Interaction Localization in Semi-Monolithic Scintillators

Semi-monolithic scintillators for PET detectors offer an excellent trade-off between pixelated and monolithic scintillators. Interaction positioning techniques have been explored in these detectors, including machine learning approaches that have demonstrated high prediction accuracy by analyzing the signal distribution among the photosensor array. In this work, we explore whether this accuracy can be further improved using different neural network architectures and number of input signals. To this end, we performed measurements of semi-monolithic scintillator arrays with a 64-channel SiPM readout used in the total-body IMAS PET scanner, and in a forthcoming brain-dedicated scanner (UHB). We compare the positioning performance of a multilayer perceptron (MLP) fed with 16 signals obtained from a summation of the rows and columns of the readout to a convolutional neural network (CNN) fed with all $64(8 \times 8)$ signals. The CNN seems able to exploit some of the light sharing observed to occur between the semi-monolithic slabs, resulting in an improved accuracy compared to the MLP. This study contributes to quantifying the performance disparities among machine learning models when provided with varying levels of signal information.