U bent hier


Opgelet! Dit event heeft al plaatsgehad.

Use of machine learning algorithms for gamma detection in positron emission tomography

woensdag, 29 april, 2009 - 17:00
Campus: Brussels Humanities, Sciences & Engineering campus
Faculteit: Science and Bio-engineering Sciences
Cedric Lemaître

Dedicated Positron Emission Tomography (PET) scanners such as
small animal PET, Positron Emission Mammography (PEM) or Brain
PET scanners, all require high spatial resolution and high sensitivity.
Most current designs use small scintillation crystals. The general
approach to improve the spatial resolution in such designs is to
decrease the crystal size. However, the decreased pixel size results in
loss of sensitivity because of the increased dead space between the
pixels. If the sensitivity of the PET scanner is insufficient, the
obtained images have to be smoothed to reduce the image variance.
Obviously, this results in a loss of image resolution and hence the
potential of the system is not fully exploited. To obtain a maximum
coincidence rate, the sensitivity of the detectors in the PET system
has to be optimized. This can be achieved by increasing the thickness
of the scintillators used to stop the 511 keV annihilation photons and
by minimizing the dead spaces in the detector design. However, these
design changes should not degrade the spatial resolution of the

To achieve this goal we developed detectors based on monolithic
scintillator blocks that are read out by avalanche photodiodes
(APDs). This increases the sensitivity due to the absence of optical
separation material between the individual scintillation pixels used in
current PET designs. The position information within the scintillator
block is embedded in the shape of the scintillation light distribution.
This principle of light spreading allows the scintillator block to be
larger than the sensitive area of the photo detector, avoiding dead
space due to the packaging of the photo detector. This again
enhances the sensitivity. In this perspective, the detector module had
to be based on new technologies. For the scintillator part, Lutetium
Orthosilicate (LSO) was chosen because of its high light yield, good
stopping power and short decay time. The S8550 APDs were chosen
as photo detector. These presented a number of advantages relatives
to position sensitive photomultiplier tubes (PSPMTs) in the
applications of interest.

In this thesis, the characteristics and implementation of the
monolithic LSO scintillator blocks in combination with a machine
learning positioning algorithm were evaluated, via simulations as well
as experimentally on a bench set-up and on a prototype scanner.
First three different positioning algorithms were tested experimentally
on the bench set-up. To this end, following positioning algorithms
were evaluated: Levenberg-Marquardt Neural Networks (LM-NN),
Neural Networks trained with an algebraic method (Alg-NN) and
Support Vector Machines (SVM). The position information is
extracted from the measured scintillation light distribution generated
in monolithic LSO blocks of various shapes and read out by the
Hamamatsu S8550 APD array.

The data acquired for the positioning algorithm evaluations were
done on an “academic” bench set-up. In order to evaluate the block
detectors in a real compact PET environment, a prototype PET
demonstrator was built. The demonstrator consists of only two
20x10x10mm3 LSO detector modules. To simulate a full-ring scanner,
the detector modules are mounted on separate rotating platforms
which allow the movement of both detector modules, also relative to
each other. In addition, since the detector characteristics may change
in time, it is also appropriate to acquire new training data from time
to time. The use of an auxiliary bench set-up for this calibration
procedure implies the removal, calibration and re-mounting of all
detector modules of the scanner. This would be a time consuming and
tedious task. That’s why an automated acquisition method of training
data for the positioning algorithm is investigated. The
implementation and validation of this procedure was done on the
demonstrator set-up.

A slight difference in the spatial resolution between the bench set-up
and the demonstrator set-up was noticed. In order to study the origin
of this difference and which instrumentation parameters limit the
performance of the whole system, a GATE based Monte Carlo
simulation was developed. Training data were simulated using the
parameters that represent the experimental set-ups. After training,
the NN was evaluated using simulated data generated with the same
parameters except that the photon beam is assumed to be perfect
now, i.e. a zero beam width. The resulting resolution will hence only
reflect the influence of the detector components and the data
acquisition method, i.e. it represents the intrinsic detector resolution.

Finally; the spatial resolution in 2D reconstructed images of the
monolithic front-end detectors in combination with the trained LMNNs
is also examined. The radial and tangential resolutions as
function of the radial source position were tested. To conclude, a
mini-Derenzo phantom filled with FDG showed very encouraging
results and corresponds with the expectations according to the
outcomes of the studied point sources.