Abstract:
The objective of this thesis is to improve the resolution and linearity of a continuous detector for positron emission mammography (PEM) imaging, by using an algorithm based on artificial neural networks. Another aim of this work is to investigate the effect of crystal thickness on the resolution and bias of the detector. A continuous scintillation detector is chosen, in order to overcome the difficulties observed in light collection and manufacturing of pixellated crystals and to reduce the cost. In this study, the detector is composed of 49 mm x 49 mm continuous LSO crystal where its thickness changes from 3 mm to 24 mm with increments of 3 mm. The photosensor chosen is Hamamatsu H8500 flat panel multi-anode photomultiplier consisting of 8 x 8 anodes. The interactions of narrow beams of 511 keV photons impacting the detector surface and the photosensor output are simulated using DETECT2000 simulation platform. The 64 outputs of the PMT is reduced to 4 and these outputs are used as the input vectors of the multilayer perceptron network for each interaction. Two sets of simulations are performed for each thickness of the scintillation crystal. One set to generate the training set and another set to create the test set. By fixing the parameters of the network and the number of iterations, the effect of crystal thickness and energy threshold on the intrinsic spatial resolution and bias are investigated. Our simulations confirmed the bias problem of the Anger algorithm and the necessity of using a biasfree positioning algorithm for scintillation coordinate estimation. Using artificial network based positioning algorithm better results are observed when compared to Anger algorithm. Results obtained show an intrinsic resolution of 0.329 mm and 0.690 mm for a crystal thickness of 3 mm and 24 mm in the center of the crystal, respectively. The systematic errors calculated are better than those obtained with Anger algorithm.|Keywords: Positron emission mammography (PEM), continuous scintillation crystal, positioning algorithm, artificial neural networks, Anger algorithm.
