Machine learning positioning algorithms for long semi-monolithic scintillator PET detectors.

In this work, machine learning positioning algorithms are developed to improve the spatial resolutions of the semi-monolithic scintillator detectors in both monolithic () and depth of interaction () directions.Two long semi-monolithic scintillator detectors consisting of 12 lutetium yttrium oxyorthosilicate (LYSO) slabs of 0.96 × 56 × 10 mmand 14 LYSO slabs of 0.81 × 56 × 10 mmwere manufactured. The scintillator arrays were read out by a 4 × 16 silicon photomultiplier array. 27 × 5 () positions of each detector were irradiated via a collimatedNa pencil beam. Extreme gradient boosting (XGBoost) machine learning model was used to predict the interaction positions forand. The genetic algorithm (GA) or particle swarm optimization (PSO) algorithm was used to optimize hyperparameters for the XGBoost model. The results of the machine learning positioning algorithms were compared to analytical positioning methods.The GA and PSO algorithms provided similar results. Compared to the analytical methods, the machine learning positioning methods improved bothandspatial resolutions especially at both ends of the detectors. The averagespatial resolutions using the machine learning positioning methods were 0.92 ± 0.41 mm and 0.94 ± 0.44 mm as compared to those obtained with the squared center of gravity method of 1.38 ± 0.23 mm and 1.39 ± 0.25 mm for the two detectors, respectively. The averagespatial resolutions obtained with the machine learning positioning methods were 1.67 ± 0.41 mm and 1.68 ± 0.45 mm as compared to those obtained with inverse standard deviation method of 2.09 ± 0.82 mm and 2.14 ± 0.81 mm for the two detectors, respectively.With the machine learning positioning algorithms, the semi-monolithic scintillator detectors with submillimeter slab thickness evaluated in this work provide less than 1 mmspatial resolution and less than 2 mmspatial resolution.