Statistics-based position decoding for a block detector

We are developing a block detector to be used in in-beam PET for hardron therapy, which consists of a discrete scintillator array and four round-type PMTs. To improve positioning performance we applied Gaussian mixture model (GMM)-based positioning algorithm that was previously developed by our group. In order to maximize separability of light distributions among adjacent scintillator pixels and thereby optimize the positioning performance, we used partially segmented block scintillator proposed by Chung et al. In partially segmented block scintillator, length of light reflectors between two adjacent discrete scintillators varies depending on the locations of the scintillators in the array. We simulated 3D crystal array with variable length of reflectors so that we extract best combinations of reflector dimensions in the array. With these optimal values, we showed the performance of our positioning algorithms. The DETECT2000 simulation package was used to model a proposed detector. The designed the detector was made up of 13 × 13 array of 4 × 4 × 20 mm
³
LSO blocks. Four sides of each crystal was attached with different length of reflectors. We used 2 × 2 one inch PMTs(22 mm effective area) so that four PMTs can share the lights. In GMM-based positioning algorithm, the response of N detector channels is represented by a feature vector. Then it trains the feature vectors to obtain the optimal parameters of M Gaussian mixtures. In evaluation step, we decoded the spatial locations of incidence photons by evaluating the measured feature vector with respect to the trained mixture parameters. The results showed that the average bias were 0 mm. In addition, most of positions for the 13×13 scintillator block were clearly identified.