Evaluation of machine learning algorithms for localization of photons in undivided scintillator blocks for PET detectors

Neural Networks trained with error back propagation Levenberg-Marquardt training (LM), Neural Networks trained with an algebraic method and Support Vector Machines (SVM) were evaluated to extract the position information from measured light distributions generated by the interactions of 511 keV photons in monolithic scintillator blocks. All three algorithms can achieve a similar average resolution (~1.6 mm FWHM in a 20 x 10 x 10 mm LSO block) but the LM trained neural networks do so most efficiently. When the incidence angle of the photons increases to 30[degrees], the resolution degrades slightly to 2.0 mm FWHM. A small mismatch (< [+ or -] 5[degrees]) between the true incidence angle and the angle for which a neural network was trained can be tolerated without significant resolution loss. Increasing the thickness to 20 mm and using a top-bottom readout of the block yields an average resolution of 2.2 nun FWHM. Index Terms--Monolithic scintillator, positron emission tomography, neural network.