You are here

Warning message

Attention! This event has already passed.

Artificial neural network based position estimation in positron emission tomography

Thursday, 16 December, 2010 - 15:00
Campus: Brussels Humanities, Sciences & Engineering campus
Faculty: Science and Bio-engineering Sciences
Mateusz Wedrowski
phd defence

Gamma detection in nearly all commercial positron emission tomography (PET)
scanners is based on the use of block detectors with a large number of small scintillator
pixels. One of the important factors limiting the spatial resolution in PET
scanners using this approach is the uncertainty of the exact depth of interaction
of the gamma ray in the crystal. Using one large and undivided piece of scintillator
where the position of the gamma interaction is extracted from the light
distribution in the block, allows overcoming this limitation. At the same time, the
sensitivity of the scanner can be increased by avoiding inter-crystal dead spaces.
However this approach tends to produce poor results if simple minded methods,
such as centre of gravity, are used to determine the position of the gamma interaction
in the crystal block. On the other hand, if machine learning algorithms such
as artificial neural networks are used, excellent performance is obtained. With
the dramatic increase of computing power in recent years, this is now a realistic
approach to real-time gamma detection in PET scanners.

Usage of avalanche photo diodes (APD) is an alternative to photomultiplier
tubes (PMT). Generally a PMT is cheaper and has a higher output signal, so
that in standard PET applications PMTs are still more common. Nevertheless, a
PET scanner by itself, as a separate device, is getting less and less competitive
with respect to integrated multi-modality solutions. In the case of PET coupled
to magnetic resonance imaging (MRI) the detectors are exposed to intense magnetic
fields. This excludes the usage of PMT and makes APD competitive. The
recently developed silicon photo-multiplier (SiPM) detectors seem to have even
more potential.

The goal of this thesis is to investigate the intrinsic detector spatial resolution
behaviour of 20×20×10mm3 monolithic Lutetium Orthosilicate (LSO) scintillator
block PET detector based on Hamamatsu S8550 APD light sensor with the artificial
neural network (ANN) applied as the estimating algorithm. The conditions of
measurements and analysis are based on realistic scanner operation. The robustness
of the neural network is studied on several parameters as the incidence beam
angle, random fraction in data, APD high voltage and temperature fluctuations.
Finally a comparison with alternative light sensors for a monolithic block detector
design is done. The data from 64-multichannel PMT and 16-channel SiPM based
detectors are studied individually to apply the same resolution reconstruction conditions.

The research is done in the framework of the Crystal Clear Collaboration in cooperation
between CIEMAT Madrid/Spain, Forschungszentrum Jülich/Germany
and Vrije Universiteit Brussels. The data from alternative detector designs are
analysed by courtesy of the group from the University of Technology Delft and
from The University of Science and Technology of China (UTSC).

In chapter 1 the basic physics processes and the main PET characteristics are
explained. First the concept of tracers is explained. Then the PET principles as
annihilation detection with other effects are described. In the later part of the
chapter, the most important PET factors are looked into i.e. spatial resolution,
sensitivity, noise and energy resolution.

In the first part of chapter 2 the scintillator material is described, its characteristic,
mechanism of operation and most common types. The second part is
devoted to a discussion of light sensors such as PMT and photo diodes. The biggest
emphasis is put on APD detector, its principles and main parameters. Promising
SiPM detector is described.

Chapter 3 is the introduction to the presentation of the result. The advantages
of the monolithic scintillator blocks based on APD design and the details of the aim
of the project are explained. The detector set-up structure and signal processing
idea are described and the artificial neural network algorithm is referred to.

Chapter 4 summarises the investigation done for APD based detectors. The
first part describes the data structure and data acquiring process. All calibrations
and preparation for measurement are specified. Then the ANN structures applied
for the spatial resolution estimation are explained. The last part shows the result
of the ANN robustness studies.

The comparison between data obtained with different detectors is presented in
chapter 5. In this chapter, the first three sections present an analysis of the data
obtained with obtained with with two detectors using of PMT and one detector
using of SiPM. The last section presents conclusion from a comparison between
these detectors.