See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/3137015

An Inter-comparison of three spectral-deconvolution algorithms for gamma-

ray spectroscopy

Article  in  IEEE Transactions on Nuclear Science · September 2000

DOI: 10.1109/23.872973 · Source: IEEE Xplore

CITATIONS READS

62 250

2 authors, including:

David Ramsden
Symetrica Security Ltd
147 PUBLICATIONS   952 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

A search for a compact monchromatic X-ray source View project

A 2-D tissue equivalent dosimeter for IMRT View project

All content following this page was uploaded by David Ramsden on 26 August 2015.

The user has requested enhancement of the downloaded file.

IEEE TRANSACTIONS ON NUCLEAR SCIENCE, VOL. 47, NO. 4, AUGUST 2000 1329

An Inter-comparison of Three Spectral-Deconvolution Algorithms

for Gamma-ray Spectroscopy

L.J. Meng and D. Ramsden

Department of Physics and Astronomy
University of Southampton, Southampton SO17 1BJ

Abstract Regularisation, ML and MEM, based on the use of an

industry-standard 3x3 inch NaI(T1) detector. The results
This paper presents a comparison of three deconvolution
clearly demonstrated the advantages of using the two iterative
techniques, Maximum Likelihood, Maximum Entropy and methods, ML-EM and MEM, but particularly the ML-EM
Linear Regularisation for the unconstrained deconvolution of
algorithm.
gamma-ray spectra. These convert the raw energy-loss spectra
obtained using a standard scintillation counter, into a good
representation of the incident gamma-ray spectrum. This work 11. THEMETHODS
is based on the use of an industry-standard 3x3 inch NaI
detector. Both simulated and measured data have been A. Theory
deconvolved using the three algorithms to provide a direct
Generally, the observed spectrum O ( E ) can be represented by
comparison between the qualities of the deconvolved spectra.
For applications in which it is important to derive an accurate the integral:
ca
estimate of the number of counts in a particular full-energy
peak, the Maximum Likelihood Method has been shown to be O(E)= ~R(E,Eo).Z(Eo).dEo (1)
superior. 0
where the [ ( E ) is the incident spectrum, and R(E,E,) is the
detector response function. This equation can be discretised
I. INTRODUCTION
as:
There is considerable interest in the possibility of improving
the accuracy of the interpretation of gamma-ray spectra
obtained using simple scintillation spectrometers through the
application of sophisticated data-processing techniques. In
particular, the power of modern personal computers now
makes it feasible to deconvolve the observed energy-loss
spectra to provide a good representation of the incident
gamma-ray spectrum. This task can now be carried out in the where the Ei(i=l,...,m) is the noise contribution and Ro is the
field to provide an operator with a simple, accurate and probability that an incident gamma-ray having an energy
quantitative assessment of the radiation environment. For falling into bin i will be detected in bin j. The task is to
example, recent work has shown that the measurement of the identify the most appropriate method to be used to predict the
natural radioactivity of concrete can provide a rapid indication incident spectra, given the observed energy-loss spectra and
of the quality of this important construction material [l].The the errors on those spectra.
preliminary work was carried out using cooled HPGe
An in-depth description of the methods that will be used in
detectors but there would be a considerable advantage if it
this comparison can be found elsewhere [SI. For convenience,
were possible to make an assessment of certain isotopic ratios
using a simpler, more efficient scintillation counter [2]. we only give very brief descriptions here.
Similarly, there arc applications in the assay of waste
1) Linear Regularisation Method
materials generated in the decommissioning of nuclear
installations in which one would like to improve the quality of The Linear Regularisation method, referred to as the Phillips-
the information provided by a scintillation counter in order to Townmey method, is a method of inversion with constraints.
obviate the need for the use of cooled detector systems. In order to avoid an oscillating solution from the
Several methods have been used in the unconstrained deconvolution process, Phillips chose the ‘straightest’
deconvolution of gamma-ray spectra, such as Singular-value solution by minimising the sum of the squares of the second
Decomposition (SVD)[3], Linear Regularisation 141, derivatives:
Maximum-likelihood Fitting by Expectation Maximisation
(ML-EM)[S] and Maximum Entropy Method (MEM)[6].
Bouchet [7] has provided a comprehensive review of de-
convolution methods for processing gamma-ray spectra. where the f ( E ) is the underlying function, in this case the
However, the methods mentioned in that work were only incident gamma-ray spectrum. On the other hand, The solution
tested on simulated and relatively feature-less spectra. In this flE)must also be consistent with the observed data, so chi-
paper, we present a comparison between three methods Linear squaredX2 and C are simultaneously minimised:

0018-9499/00$10.00 G 2000 IEEE

1330
L(j,a)= x 2 +ac (4)
to find the final solution. In this equation, A is the Lagrange
multiplier. The method, although simple to implement, suffers
some drawbacks, which limits its use in the some applications.
Firstly, the method generally prohibits the occurrences of
sharp features and also the solution derived is not positively
constrained. This leads to difficulty in defining the relevant
peak areas.
2) Maximum Likelihood Estimation using Expectation
Maximisation (ML-EM)
The ML-EM algorithm was initially developed for use in the
reconstruction of images in emission tomography [9].
Following Shepp and Viadi, the underlying function is
discretised into a series of elements fi, i=1,2,...N , each an
independent variable with a Poissonian distribution. The ML-
EM algorithm generates a sequence of estimated spectra
jr using:
j=l
where the f i * is the measured spectrum and Rg is the
probability that an incident gamma-ray with energy
corresponding to bin i is detected in bin j. Each new estimate
monotonically increases the likelihood function as defined in
Shepp and Vardi until a global maximum likelihood estimator
is reached.
3) Maximum Entropy Method (MEM)
The Maximum Entropy Method is a probabilistic method
(Jaynes 1957). As applied to spectral deconvolution, it
consists in maximising
where the term is the smoothing or regularising parameter.
The entropy term S ( f , m ) ensures the information contained
in the distribution f (the solution) with respect to an a priori
model m. There are several possible definitions of the entropy.
Here, we used the one suggested by Jaynes [ 101:
This expression is a maximum when the two distributions are
identical (f= m). The solutionfmust be both positive and
additive.
B. The Detector Modelling
The detector used in this study was a rather old Harshaw
Integral Line detector (Type: 12S12/3e), which consists of a
NaI(T1) scintillation crystal coupled directly to a
photomultiplier tube with a rigid, high refractive index, optical
coupling medium. The crystal and the matching
photomultiplier tube is hermetically sealed in a low mass
light-tight housing having an aluminium entrance window.
The housing consists of a thin aluminium can around the
crystal connected to a mu-metal shield which surrounds the
photomultiplier tube as shown in Figure 1.
082 .
r 2
Figure 1: The 3x3 inch NaI detector assembly. (From The Harshaw
/Filtrol SCINTILLATION PHOSPHOR catalogue)
The modelling of the detector’s energy-response function can
be a very computing-intensive task. It is usually carried out in
two ways
0 Numerical interpolation using a set of available spectra as
the basis [ 111.
Derivation of an empirical analytical function by
parameterising available response functions (e.g. Monte
Carlo). Several analytical forms already exist in the
literature [ 121.
In this work, we chose to use another approach. The detector
energy-response function is determined purely by Monte
Carlo simulation. Even though, the whole process takes a long
time to complete, it provides a better accuracy than has been
achieved using other methods.
The detector modelling was performed using the CERN high-
energy particle-transport code, GEANT-3. The detector
geometry was defined to match the characteristics of the
detector used. The system response-function consists of 1024
energy-loss spectra derived by irradiating the simulated
detector using mono-energetic point sources of radiation
located at a point lOcm on axis from the front surface of the
scintillation counter. The energy of this source was varied
between 0 to 3072keV in 3kev steps. For each energy, 2x106
events were generated and the calculated detector response
binned into 1024 channels.
Since the output from the GEANT simulation does not reflect
the Gaussian statistics in the process between the deposition of
the photon’s energy and the production of the photo-electroms
in the PMT, an extra program was used to apply an
appropriate broadening to the simulated energy-response. The

curve relating the energy-resolution[r(E)] as a function of
energy E was derived by fitting a curve to the experimental
data in Figure 2.
r ( E ) = r(E')/(E'/ J q
using a value of e0.3.
0.125
\
0.12 . _ _ .
_..I
...........................................
.....................................................
s 0,115
2 0'11 ....................................................
0
f
v)
p
0 105 ....................................................
................................ ....................!....._
. 0.1 .............................................
w 0.095
0.09 ..................................................
Figure 2: Measured energy-resolution as a function of energy. The
solid curve is the estimated energy resolution using the power law:
r ( E ) = r(E')/(E'/ E)0.3
The simulated energy-response was compared with
experimentally measured spectra to prove the accuracy of the
modelling. Good agreement between these spectra is
demonstrated in Figure 3, where the discrepancy in the
Compton region is mainly due to the contribution of back-
scattering and X-ray emission from the surrounding material.
It is worthwhile to note that even though the simulated
spectrum looks very similar to the measured one, it does show
a higher Compton edge and shorter tail on the higher-energy
side of the photo-peaks than the measured one. These effects
may introduce some artefacts in the deconvolved spectra.
Another defect in the detector model is that since the number
of counts cannot be infinitely large, when these counts are
shared between a large number of energy-bins, the statistical
error (Figure 4) on the system response function will introduce
a broadening in the deconvolved spectrum.
0' 500 1000 1500 2000 2500 3000
Energy (keV)
Figure 3: Simulated energy response for different incident energies.
1331
c
1.11.0~0.149,
0 200 400 Nd Bm
ChamilNWii
Figure 4:Comparisons of simulated (blue) and measured (black)
energy loss spectra
.
!a---
111. RESULTS
OF THE COMPARISONS
A. Using simulated data
In order to compare the ability of these methods to resolve
closely spaced peaks, a simple test spectrum was generated. It
consists of two peaks, one at 180keV and the other at
180 + AE keV, with the AE was changed from 10 to 40 keV.
The spectrum was deconvolved using the Linear
Regularisation, MLEM and MEM methods respectively and
the results are shown in Figure 5.
The Linear Regularisation Method can only resolve the two
peaks when the AE is greater than 40keV. The relatively
poorer result obtained using this method is a consequence of
incorporating the regularising term C in the minimisation
principle. This tends to select the straightest curve available so
the sharp features can not be reproduced and an over-shoot
may be generated to compensate for the increasing value of C
which results from the pressure towards consistence with the
observed data. The other problem associated with this method
is that the solution is not positively constrained. Negative
values may be produced in the deconvolved spectrum. The
only advantage of the method over the ML-EM and MEM is
that the inversion of the matrix need only to be carried out
once so the deconvolution process is simply the multiplication
of a vector (the observed spectrum) with a matrix. This
process takes roughly the same time as that taken by a single
iteration in ML-EM and MEM method.
3500
Figure 5: Deconvolution of energy-loss spectra with two peaks at 180
and 180+AE (shown at the left-hand side). The methods used for the
de-convolutions are shown.
1332

A modified regularisation function has been suggested [ 131, very broadened peaks and severe artefacts around the peaks.
namely: This makes it difficult to determine both the peak-positions
1 and the peak-areas accurately. If a relatively accurate
s =Z
NJ=1 -(oj
0; -I j ) (9) deconvolution is required when speed is not important, we
recommend the use of either the ML-EM or the MEM method.
This function will be zero if the Z is exactly equal to the
original spectrum 0.Therefore, unlike the function C, this
function does not act uniformly over the whole spectrum. It
allows larger variations where variations are already present in
the observed spectrum and therefore the peaks are smoothed
out less. However, the matrix to be inverted depends on the
..................................
observed spectrum and therefore needs to be constructed for 3 10000
each 0 separately. This means that the speed advantage over
5
8
the iterative methods disappears. Furthermore, the
regularising function tends to select the solution most similar
to the measured one; therefore, the incident spectra will not be
E 5000

accurately reproduced when the detector has a relatively poor

energy-resolution,
The MEM provides a better resolution than LR and the 0
0 500 I000 1500 2000 2500 3000 3 10
solution is positively constrained. Unlike LR, the entropic Energy (key
regularisation function does not prohibit the occurrence of Figure 6: Simulated test spectrum
sharp peaks in a low, slowly changing, background. However,
the solution achieved presents lower peaks and higher valleys
than that achieved by ML-EM, because of the pressure
towards uniformity [ 141, Therefore, the method intrinsically
imposes some difficulties when the peak-areas need to be
measured calculated.
The best resolution was achieved using the ML-EM algorithm.
As will be shown later, the positive solution, conservation in
the total number of counts and the feasibility of working in a
very low-count situation, makes this a good method for use in
such an application. ..................
..............
Table 1: Lines and their intensities

0 500 1000 1500 2000 2500 3000 3500

Energy (keV)
Figure 7: Deconvolved spectrum using Linear Regularisation
algorithm

.... ..........................................

In order to quantify the performance of these methods, a

relatively complex spectrum was generated using the same
detector model. The source used was a simulated U-ore point
source in which its 12 strongest gamma-ray lines were
included. The intensities of these lines are shown in Table 1.
The number of counts in the spectrum was reduced in steps
from 1x106 to 16,250. The test spectra were deconvolved each
of the three methods. The results shown in Figure 6-9 clearly Energy (keV)
demonstrate that the incident spectrum is well reproduced Figure 8: Deconvolved spectrum using MEM algorithm
when the ML-EM and MEM methods were used. Again, the
deconvolved spectra obtained using the LR method shows
 1333

entropic constraint. Therefore, if the peak-area is of the

primary importance, ML-EM maybe a better choice.
E
The peak-positions and the peak-areas were also calculated as
-
h

2 6
a function of the size of the data-set used. Four independent
data-sets were randomly generated from the original data-set
Y
E containing 1x106events by dividing the number of counts by
E 4 successive factors of four as shown in Table 3. An important
-
c

conclusion from this data is that the errors in the calculated

2 peak-areas remain close to that caused by the Gaussian
0' 500 1000 1500 2000 2500 3000
Energy (keV)
Figure 9: Deconvolved spectrum using ML-EM algorithm
In order to compare how well the incident spectrum can be
reproduced using the ML-EM and MEM methods, the
estimated peak-positions and peak-areas of three particular
peaks-of-interest (295keV, 609keV and 112OkeV) were
measured in the deconvolved spectra. A test spectrum
containing 1x106 events was used in this simulation. The
results are presented in Table 2. Both methods predicted the
position of the peaks very accurately. The ML-EM algorithm
also provided a very accurate estimate of the peak-area,
usually within 0.5%. The MEM method tends to give slightly
lower values than the real peak-areas by about-3% for the
three peaks selected. As discussed before, this is the result of
the pressure towards uniformity which is imposed by the
fluctuations. This indicates that the accuracy of the peak-area
measurements have not been degraded by the deconvolution
3500 process. Clearly, relatively inexpensive scintillation detectors,
used in conjunction with a deconvolution technique may be
able to provide a better result that an expensive HPGe detector
for some applications. This is a consequence of the power of
the method to move counts appropriately from the Compton
region into the photo-peaks.
B. Using Experimental Data
The deconvolution methods were also tested using
experimentally measured spectra. The 3x3 inch NaI(T1)
detector was irradiated by point sources on the axis of the
detector and lOcm away from its entrance window. Both
simple (22Na+'37Cs and 6oCo) and complex spectra (U-ore)
were used in these tests. The energy-loss and the deconvolved
spectra are shown in Figures 10 and 11. Some important
results are listed below:
0 The best peak-resolving power was achieved using the

Table 2: Estimated peak positions and peak areas using ML-EM and MEM (1 million counts)
Methods MLEM MEM MLEM MEM MLEM MEM
Peak I Real 295keV 609keV ll2OkeV
p0si.s
Estimated 295.4 295.1 607.8 607.79 1118.0 1119.2
(keV) + 2.13
t 0.044 + 0.074 + 0.44 + 0 045 + 1.26
] Peak
Areas
1 Real I 153930 I 280762 I 79993 I
153713 150872 279788 275665 80799 79009
(counts) Estimated
f0.17% f0.47% f0.2% t0.26% k0.6% +1.2%

Data sets 111 114 1/16 1I64

Methods MLEM I MEM MLEM I MEM MLEM 1 MEM MEM I MEM
Peak Real 609keV
posi.
Estimated 607.8 607.79 608.21 607.8 608.16 607.84 608.0 608.18
(keV) + 0.44 + 0,045 + 0.97 + 0.084 + 0.99 + o 327 + 1.19 + 1.59
Real 280762 70190 17548 4387

Peak Estimated 279788 274665 69853 68564 17436 17116 4325 4237
Areas
(counts) Enor f 0.2% f 0.26% k 0.53% t 0.56% k 0.87% k 0.9% k 1.8% k 1.9%
1 Gaututm I t 0.18% 1 t0.4% 1 k0.76%
1 f 1.5%
I
1334

ML-EM algorithm. This also provided best peak-to-valley
ratio and narrowest peak-widths.
The energy-resolution of the detector is improved from
12% to <3% for the 662 keV line by the use of ML-EM
algorithm.
0 The number of counts in the 1332 keV peak in the
deconvolved 6oC spectrum, is approximately eight times
greater than in the original energy-loss spectrum.
Apart from the artefacts introduced as a result of using the
LR algorithm, some small artefacts are generated in the
ML-EM and MEM deconvolved spectra. These are
mainly due to the discrepancy between the real and
simulated energy-response of the detector.
IV. CONCLUSIONS
The Linear Regularisation method is linear and it is easier to
understand the relationship between the incident and output
spectrum. The implementation of this method can be based on
6 5
,x 10
Measured
U-ore spectra
some well-known linear algebra packages, such as LAPACK
[15]. By using this method, events can be collected and
processed individually, so that the real-time monitoring of the
radioactivity is possible. However, due to some limitations
intrinsic to the algorithms, the deconvolved spectral quality is
relatively poor compared with the other two methods. The
ML-EM and MEM generally give better deconvolved spectra
and are capable of working in situations the spectra to be
processed have a very low number of counts per channel.
Another important advantage of these two techniques is the
are intrinsically positively constrained. On the other hand,
these methods require a large number of iterations to be
performed before convergence is reached. The computing-
time needed is typically n times longer than the time to
required by non-iterative methods such as SVD and the
Regularisation Method where n is the number of iterations
required. However, this problem is not serious for
spectroscopy applications since the spectra usually have only
few thousand channels or less. Some conclusions we reached
so far are list below:
.,x 10
.-
Deconvolved
U-ore spectra
LR

"0 500 1000 1500 2000 2500 3000 0 500 1000 1500 2000 2500 3000
Energy (keV)
IC16
' 1x5 '

Decopnvolved
I U-ore spectra
MEM

500 1000 1500 2000 2500 3 3

Energy (keV)

Figure 10: Comparison of deconvolved Uranium Ore spectra using different methods
 1335

ML-EM ..-

‘
A

I ............. .......................................

I If
0 500 1000 1500 2000 500 1000 1500 200
Energy (kev) Energy (kev) Energy (kev)

10‘
15000,
J ............................... > ......................
ML-EM I MEM .........................
2.5.-
*CO
?.
I !................. A

E 2--- ......................
5 3 ..................................
v

f g,5 ....................................................
z2..
a.
+
-
....................................... a’
I

c ,.....................................................
,..... ..........................................

0-
T
n
d,
37L
I

0 500 1000 1500 2000

Energy (kev)

Figure 11: Deconvolution of measured spectra using different methods. The blue lines are measured spectra and the black ones are
deconvolved spectra

1) Linear Regularization uses a priori information to smooth

the resulting spectrum and suppress the noise. However, V. ACKNOWLEDGMENTS
the use of some linear regularisation functions smoothes
The authors would like to thank Dr. I.D. Jupp for his help in
the spectrum and inhibits the occurrence of sharp features
in the spectrum. This results in a poor peak-resolving understanding the algorithms and developing the computer
programs for the work. The authors would also like to thank
characterisitc and introduces artefacts due to the pressure
Dr. G. Daniel for his general help in the implementation of the
to conform to the a priori model. The solutions from
Maximum Entropy Method in the work. Our interest in the
these methods are usually not positively constrained.
application of spectral deconvolution techniques was triggered
Even though simple to implement and less demanding of
by Dr. R.J. Evans demonstration of its ability to improve the
computing power, this is not a recommended technique
quality of data derived from a CsI(T1)-photodiode array that
for application in the processing of gamma-ray spectra.
he used to record neutron-activation gamma-ray spectra [ 161.
The ML-EM algorithm has proved to be very effective in L.J. Meng would like to thank Rutherford Appleton
medical imaging applications particularly when dealing Laboratory for partially supporting his PhD studentship.
with low signal-to-noise ratio data. It provides the scope
for modelling the Poisson statistics in each detection VI. REFERENCES
element. The problem of slow convergence is less serious
in the spectroscopy application since the size of the data- [l] C.G. Rowbottom, W.B. Gilboy and D.J. Hannant,
set is relatively small. The resolution achieved in the “Determination of cement content of cement blends using
feasibility study using both simulated and real data is Gamma-ray spectroscopy”, Cement and Concrete
promising. Research, Vol. 27, No. 11, ppl68l-1690, 1997.
3) The Maximum Entropy Method for solving the inverse [2] The HarshawFiltrol SCINTILLATION PHOSPHOR
problem falls into the general category of Bayesian catalog by HarshawFilrtol Partnership. 608 1 Cochran
Methods. The a priori model chosen is the one which is Road, Solon, Ohio 44139 U.S.A
maximally noncommittal about unavailable information. [3] W.H. Press, S.A. Teukolsky et al. “Numberical Recipes in
Compared with the Linear Regularisation method, such a C”, Volume 2, The Press Syndicate of the University of
priori information leads to a non-linear regularisation Cambridge.
function and it is then difficult to implement and also
[4] D.L. Phillips, Journal of Associate Computing, Vol. 84,
becomes computationally intense. The Maximum Entropy
March 1962.
method has been used successfully in many fields and has
proven to be a very powerful technique. For this [ 5 ]L. A. Shepp and Y. Vardi, “Maximum Likelihood
particular application, it gives similar results to that Reconstruction for Emission Tomography”, IEEE MI1,
achieved using ML-EM. No. 2, 1982.
 1336
[6] Edited by J. Skilling, “Maximum Entropyand Bayeisian
Methodsin Science and Engineering”, Vol 1.
[7] L. Bouchet, “A Comparitive study of deconvolution
methods for gamma-ray spectra”, Astronomy &
Astrophysics Supplement Series. Ser. 113, pp167-183,
1995.
[8] M. J. Berger and S. M. Seltzer, “Response function for
Sodium Iodie scintillation detectors”, NIM, Vol. 104,
~ ~ 3 1 7 - 3 3 1972.
2,
[9] L. A. Shepp and Y. Vardi, “A statistical model for positron
emissiontomography”, Journal of American Statistical
Association,Vol. 80, No.389, pp.8-20, 1985.
[lo] E.T. Jaynes, “How the the brain do plausible reasoning?’,
in Maximum-Entropy and Bayesian Methods, Edited by G.
J. Erickson and C. R. Smith. Kluwer Academic Press.
[ 111 R. R. Kiziah and J. R. Lowell, “Experimental
responsfunction spanning the gamma-ray energy range of
123.6 to 11.67MeV and response matrix generation for
View publication stats
Bismuth-Germinates Scintillation detectors” , NIM, Vo1305,
pp92-, No.1, 1991.
[ 121 C.E. Moss et al, “unfolding the Bismuth-Germinates
pulse-height distributions to determine gamma-ray flux
spectra ”, NIM, Vol. 219, No. 3, pp558-564, 1984.
[ 131 C. Sukosd, W. Galster et al, “Spectrum unfolding in high
energy gamma-ray detection with scintillation detectors”,
Nuclear Instruments and Methods in Physics Research
A355, ~ ~ 5 . 5 2 - 5 51995.
8,
[ 141 Edited by B. Buck and V. A. Macaulay, “Maximum
Entropy in Action”, Oxford Science Publications, 1994.
[ 151 LAPACK User’s Guide - Release 2.0, The Society for
Industrial and Applied Mathematics. SIAM, 3600
University City Center, Phildiaphia, PA 19 104-2688.
[ 161 R. J. Evans, PhD Thesis, University of Southampton,
1999.