Study Of High Spatial Resolution PET Detector Modules Using Neural Network-Based Position Estimators

Positron emission tomography is a nuclear imaging technique that based on the detection of gamma rays. It is a noninvasive technologies allowing tracing in vivo metabolic function of the organism at molecular level.In recent years, the interest in small animal positron emission tomography (PET) and dedicated PET (like PEM and Brain PET) has increased dramatically, stimulating the development of detectors with high spatial resolution. The small dimensions of these animals and the organs of human impose stringent requirements on the high spatial resolution and sensitivity of those PET systems. Current designs based on matrices of individual crystal pixels have achieved an intrinsic spatial resolution better than 2mm. But the detection efficiency in these pixelated detector designs is reduced due to the dead space introduced by the reflective material between the crystals. Additionally, the spatial resolution may be deteriorated by inter-crystal scatter and parallax errors because of depth of interaction (DOI) effects.Alternatively, detector based on large continuous scintillator blocks results in a higher sensitivity since there is no dead space due to material needed to optically separate individual pixels. Those detectors have the advantage of simple design, lower cost, better energy and time resolution, compared with the pixilated ones. The size of monolithic crystals can even be chosen such that they completely cover the photo detectors used to read out the scintillation light, including the packaging. The entry point of a detected 511 keV photon on the surface of the crystal, instead of the interaction position is determined from the distribution of the scintillation light using a non-anger position algorithm, such as neural network, nearest neighbor and statistics based positioning algorithm, etc.We are currently developing a practical implementation of prototype monolithic scintillator PET detector modules with neural network position algorithm. This thesis gives a detailed description of this PET detector prototype, including the measurement of the physical properties of the LYSO scintillator, the optimization of signal readout scheme from the MC-PMT, the design and test of the electronics system and the experimental PET prototype, the FPGA implement of neural network position estimator, etc.The following outlines the contents of each chapter in this thesis briefly. The first chapter is the preface, situates PET in the molecular imaging world and introduces the principle of PET imaging, imaging and scanner performance, two examples of application.The second chapter introduces the components of PET detector and the design scheme of several high resolution small animal PET systems.The third chapter depicts the basic principle of neural network and mainly introduces the structure and learning methods of MLP (Multiplier perceptron) and RBF (Radial Basis Function) networks when they are used to solve a regression problem (function approximation).The fourth chapter demonstrates the physical properties of the LYSO scintillator measured by ourselves, including emission spectra, decay time, energy resolution and light output.The fifth chapter illustrates the optimization and its affect on spatial resolution of four possible signal readout geometries through combinations of the 64 channels signal from PMT before the digitization by Monte Carlo simulation using Geant4.The sixth chapter gives a detail of the design of the experiments PET detector prototype, including the design and test of the hardware and software of the electrical system, the components and the primary measurements of the detector prototype.The seventh chapter presents an overview of the performance of monolithic scintillator detectors based on LYSO crystal and MC-PMT H7546B and several methods to chose the hidden neurons number of MLP NN. The investigation includes timing resolution, energy resolution and spatial resolution.The eighth chapter describes a performance and resource efficient architecture to realize on-line neural network position estimating for PET detector modules. The implementation of network is based on FPGA.The ninth chapter is the conclusion of this thesis. The innovation and future work are presented.