Professional Documents
Culture Documents
www.elsevier.com/locate/ndteint
Abstract
In this article, we present a generic software for the simulation of gamma-ray radiography. This software simulates the entire radiographic
system, from the source to the detector consisting of metallic screens and ®lms. In an industrial context where the goal is to detect structural
¯aws in material like cracks, this simulator allows to compute gamma-ray images for different system parameters. By this way, engineers can
choose an optimal set of parameters leading to the best image of ¯aws. We use Monte-Carlo techniques for the simulation of the whole
system composed of a source, an object to inspect and a detector. The main contribution of this paper is to show that simulated images are
coherent with real images although we use a simpli®ed model for particle transport. Besides, we propose an acceleration technique to
simulate the Markov chain of photon transport. Finally, an experimental design is performed leading to a linear model expressing the
in¯uence of the system parameters on image quality. q 2002 Elsevier Science Ltd. All rights reserved.
Keywords: Radiographic simulation; Nuclear plant control; Experimental design; Monte-Carlo computation; Argentic ®lm homogenization; Importance
sampling
0963-8695/02/$ - see front matter q 2002 Elsevier Science Ltd. All rights reserved.
PII: S 0963-869 5(02)00014-2
504 A. Bonin et al. / NDT&E International 35 (2002) 503±510
Fig. 4. Photons trajectories. Fig. 5. Collision parameters for the photon transport model.
506 A. Bonin et al. / NDT&E International 35 (2002) 503±510
law denoted R, derived from KN. These laws describe the the family of `importance sampling' techniques [10]. It
deviation fn an 2 an21 : consists in modifying the transition kernel so that photons
are more likely to reach the detector. Indeed, in most of our
p1 an uzn21 ; Cn Comp KN an 2 an21 ; ln21 ;
radiographic setup with the natural Markov chain, around
70% of photons do not reach the detector (Fig. 4). This
p1 an uzn21 ; Cn Rayl R an 2 an21 :
occurs either because photons are absorbed or because
R does not depend on the energy. For c [ {Comp; Rayl}; p 2 they exit from the object outside the detector. Modifying
is an exponential law E m ln 2 the transition kernel alleviates this drawback.
simpli®cation enables the software to generate realistic considered as a homogeneous material, photons collisions
images in acceptable time. In the ®lm, an electron blackens are simulated according to the simulation process described
all the grains that it crosses along its trajectory. Each photon in Section 2.1. Then, given these collisions, we have to
of the virtual image obtained at level I generates such a simulate the electron trajectories across the grains although
branching process. Finally, the latent image is obtained by they are not explicitly present in the homogenized ®lm.
counting on a regular grid the number of blackened grains in Because of the simpli®cation made earlier, for every elec-
each cell. tron we can assume that the positions where it hits a grain
Homogenization. To simulate the branching process in along its trajectory, are distributed as the grain positions
the ®lm, an obvious approach would be to simulate a repar- along this trajectory. Consequently, this distribution can
tition of spherical grains with random radii and then to be described by a model with parameters e g and t g. To do
simulate the branching process through this spatial con®gur- that, we decompose the length trajectory as
ation. This would be extremely time and memory consum-
X
I
ing due to the large number of grains. This is why we have `0 ` 0i ; with ` 0i t 1 e; 6
adopted an `homogenization' approach that we present in i1
the following.
where ` 0i is the distance between two collisions occurred
The mean number of grains by volume unit is very high
within two neighboring grains (we assume that only one
for commonly used ®lms so that a 5 mm 2 ®lm contains
collision can occur in a grain). The collision number I
several billions of grains. Manufacturers try to obtain a
depends on the electron energy. e is the random distance
homogeneous repartition for grains and we can assume
between two neighboring grains and t is the random length
that the grain repartition is uniform. This is why we consider
of the electron path in both grains. We suppose that ` 0i
the ®lm as a homogeneous material which will be later
follows an exponential law E tg 1 eg 21 whose expecta-
characterized by two parameters: e g, the mean distance
tion is E ` 0i tg 1 eg : To simulate an electron trajectory,
between two grains and t g, the mean grain diameter. Thus
the algorithm computes ` 0 by Eq. (5) and then draws
the detector is composed of three homogeneous materials:
successively ` 0i according to the exponential law under
two screens and a ®lm. Then the photon transport in these
the constraint (6). After each collision i, the electron loses
materials is similar to the one in the object, the cross-section
a part of its energy as it can be computed by inverting
coef®cients being those of the grains.
Eq. (5).
Electron transport. After the Compton collision of a
photon with energy l , the emitted electron has energy l 0
and deviation angle f 0 satisfying (cf. Fig. 2 and Ref. [17] 3. Experimental results and optimization
kl 1 2 cos f
l0 l ; In Section 1, we have mentioned that the radiographic
1 1 kl 1 2 cos f
4 image quality depends on the choice of the radiographic
1 setup: source type, source diameter, ®lm type, front and
f 0 arctan ; back screen thickness, source/object distance, and on the
1 1 kltan 1=f
object parameters: thickness, inclusion position, material
where f is the photon deviation angle. The mean free path type, etc. We try now with our simulation software to under-
of the electron is de®ned in Ref. [20] stand the in¯uence of these parameters on the image quality
0:407 01:38 y measured in terms of contrast and blur in the case of
`0 l if l 0 , 0:8 MeV parallelipedic inclusions. The contrast is the gray level
r
difference between the inside and outside of the ¯aw in
1
0:542l 0 2 0:133 if l 0 . 0:8 MeV; 5 the simulated image. The blur is the measured line spread
r function of the ¯aw edge. It means that we are able to extract
r being the volumic density of the material. such values from every image.
In the screens, only electrons diffused towards the ®lm Experimental design methodology is a well suited tech-
contribute to the image formation, so only back-scattered nique to analyze and optimize the in¯uence of parameters
electrons are useful in the back-screen (the front screen (also called factors) (cf. Ref. [18] among many others). We
plays a reinforcing role as it liberates numerous electrons have identi®ed eight main factors. For sake of clarity, we
that will reach the ®lm because of the low screen thickness). restrict our presentation to three factors with two modalities:
In the back screen, back-scattering predominates and we source diameter (0.1 mm, 3 mm), front screen thickness
model it directly. For an electron, the probability to be (low, standard), back screen thickness (low, standard).
back-scattered is determined thanks to dedicated cross- These factors are, respectively, denoted by A, B, C and are
sections, and the back-scattering angle f 0 follows the coded by {21,1}. To observe all factor combinations we
Rutherford's law [5]. We now describe the particle transport need to perform 2 3 experiences, that is eight simulations.
in the ®lm. Fig. 7 shows simulated images corresponding to six com-
Simulation process. The ®lm (gelatin and grains) being binations in the same radiographic setup as Fig. 3.
508 A. Bonin et al. / NDT&E International 35 (2002) 503±510
Fig. 7. Simulated images (with N 400 millions of photons) from a two- Surprisingly, the simulation software is able to reproduce
level factorial design 2 3, the factors being: A source diameter, B front ®ne effects that we observe on real radiographic images. For
screen thickness, C back screen thickness. For the ®rst line: A 0.1 mm
instance, using our simulation software, we have obtained
and for the second line: A 3 mm. For each line, the factor modalities
correspond to: (a) ®lm `alone', (b) ®lm with front screen, (c) ®lm with an image density curve in terms of the factor B (front screen
front screen and back screen. thickness). This curve is coherent with the real experiments
(Fig. 8).
The results of these simulations respect the expert knowl- Thus, we determine that the optimal front screen thick-
edge. Without screen of signi®cant thickness, the radio- ness is 0.75 mm, which is exactly what real experiments
graphic image has a very poor quality whatever the source give for this con®guration. In this context we can try to
size: 0.1 mm (Fig. 7(a)/line 1) or 3 mm (Fig. 7(a)/line 2). optimize the system parameters by handling more than
The main qualitative effect is due to the front screen factor B one factor, for instance A and B. Such a problem can be
(Fig. 7(b)). Given B, the source factor A has major in¯uence solved by using a model similar to Eq. (7) but with quanti-
but the back screen factor C has a minor one. tative variables A and B instead of binary ones. It is the
This experimental design can be more deeply analyzed by well-known surface response technique where one tries to
statistical techniques [18]. If we consider that the image optimize the system response y in terms of a and b, [18]. Let
quality measure y is the occurrence of a random variable us add that the computation of a simulation without impor-
Y, this analysis is based on a linear model which gives a tance sampling, and using 100 million of photons for a
decomposition of the expectation E(Y). For each triplet 5 mm 2 ®lm digitized on a 100 2 grid takes less than one
a; b; c [ { 2 1; 1}3 ; this expectation is denoted by E Y hour on a Pentium III 650.
t a; b; c; and the model reads as
t a; b; c e 1 a eA a 1 b eB b 1 c eC c 1 ab eAB a; b
4. Conclusion
1 ac eAC a; c 1 bc eBC b; c 1 abc eABC a; b; c;
We have presented a simulation tool for representing the
radiographic process from the source to the detector. The
where eA, eB and eC are the main effects, eAB,¼,eABC are the particle transport simpli®cation and the Markov chain simu-
interaction effects and e isPthe mean effect.P This model lation acceleration lead to acceptable computation times in
naturally
P assumes that: a eA a 0; ¼; b eAB a; b the industrial context of non-destructive evaluation. This
b eAB a; b 0; ¼ The statistical analysis of the eight simulation tool is highly con®gurable and simulated images
y values extracted from the simulated images con®rms are sensitive to parameter modi®cation in the same way as
that ®rst, B and second, A are the main principal effects. real images. Besides, the experimental design through
Furthermore, A has a slight in¯uence through the inter- surface response methodology and the resulting model is
action AB. On this limited experimental design, the of great interest for experts to assist them in their analytical
effect of C is not statistically signi®cant. So, the model is approach of radiography.
reduced to: The qualities of this simulation software must, however,
t a; b; c e 1 a eA a 1 b eB b 1 ab eAB a; b: 7 be tempered by the fact that the ®lm development process is
not modelized. This leads to a difference between real and
For instance, the triplet (1, 2 1,1) gives: simulated images as shown on Fig. 3. Development seems to
t 1; 21; 1 e 1 eA 1 2 eB 21 2 eAB 1; 21: introduce an additional blur. We work on estimating this
blur through a point spread function estimated from the
For this linear model, the effects are estimated by minimiz- comparison of real and simulated images. Fig. 3(c) shows
ing the least-square criteria between y and t . the introduction of such a blur. However, this blur does not
A. Bonin et al. / NDT&E International 35 (2002) 503±510 509
modify the relative in¯uence of parameters on the radio- number of occurrences m~ , m: In the literature, this
graphic images, and consequently it is not a drawback for procedure is called `importance sampling'. So, let {zni ; 0 ,
factor analysis. n # J 1 1}i1;¼m~ be m~ independent occurrences of the
Markov chain {Zn ; 0 , n # J} with kernel K: ~ The new
estimator is:
5. Annex: variance reduction "j 1 1 #
1 Xm~ Y dK Z i ; Zni
Following Ref. [7], the two next sections recall the clas- P~ j z0 n21
m~ i1 n1 dK~ Zn21
i
; Zni
sical framework for achieving variance reduction using
importance sampling technique. To gain an advantage, 1Sj i [ M; Sj11
i i
[ v1Lj11 [ l
most successful applications of the method rely on exploit-
ing the peculiarities of the particular problem at hand. In the 1 X m~
third section, we propose a original and generic approach to C i 1Sj i [ M; Sj11
i i
[ v1Lj11 [ l :
m~ i1 j
determine speci®c laws for importance sampling.
Natural Markov chain. Let us assume that the grid G is Clearly, this is an unbiased estimator: EK~ j11 P~ j z0
composed of elementary cubes v and that the range of z0 ;´
energy is partitioned into intervals l of center l . For Pj z0 : This property is due to the weight C j i :
every V v; l ; we have to estimate the probability Biased kernel estimation. The importance sampling tech-
Q(z0) Prob(z0; V) that a photon reaches V from its initial nique consists in replacing the natural laws p 0, p 1 and p 2
state z0. This probability is written: which de®ne the kernel K by new ones in order to de®ne the
biased kernel K: ~ Let us present our original approach for the
X
1
Q z0 Pj z0 deviation law in the case of Compton collision.
j0 To do that, let us consider m0 independent occurrences of
the natural Markov chain, with m0 p m: ~ Let D
where Pj(z0) is the probability that a photon reaches V from { fi ; li ; i 1; ¼; m 00 } be the set of deviations and energies
the state z0 after j and only j collisions in M after z0: obtained during the simulation of these occurrences (we
Pj z0 PZj11 [ V; Zj Ó VuZ0 z0 : have m0 , m 00 since there are at least one collision per
photon). For every l , this set gives a sample of the random
Let us truncate this series: variable (F ul ) of law KN(f , l ). Now, let D~ be the restric-
X
J tion of D to the occurrences having reached the grid G.
Q z0 < Pj z0 : For every l , D~ gives a sample of the biased law KN; f
j0 which contributes to de®ne K: ~ Practically, we choose a
parametric expression KN ga for KN; f and then, the biased
The estimate of Q(z0) will be of the form:
law is estimated according to the pseudo maximum like-
X
J lihood principle [3]. By setting D~ l { fi ; li : li [ l };
Q^ z0 P^ j z0 : for every l ; we compute the pseudo maximum likelihood
j0
estimate of a :
Let us detail how the estimate P^ j z0 is obtained. Let Y
a^ l arg max ga fi li :
KN
{zni ; 0 , n # J 1 1}i1;¼;m be m independent occurrences a
fi ;li [D~ l
of the Markov chain {Zn ; 0 , n # J 1 1} with kernel
K z0 ; ´: By noting that
"j 1 1 #
Z Y
Pj z0 dK zn21 ; zn 1sj [ M; sj11 [ v1lj11 [ l Acknowledgements
z1 ;¼;zj 1 1 [EJ 1 1 n1
Biased Markov chain. To reduce the computation time, titative Non Destruct Eval 1995;16.
[3] Besag J. Spatial interaction and the statistical analysis of lattice
we do not use anymore the natural Markov chain but a new systems. J R Stat Soc 1974;B-36:192±236.
Markov chain with kernel K~ for which the associated esti- [4] Bonin A, Chalmond B. Moderato: a monte carlo radiographic simula-
mator P~ j z0 has an equivalent accuracy, but for a smaller tion. Rev Prog Quantitative Non Destruct Eval 1999;19.
510 A. Bonin et al. / NDT&E International 35 (2002) 503±510
[5] Cauchois Y, HeÂno Y. Cheminement des particules chargeÂs. Paris: [16] Lapeyre B, Pardoux E, Sentis R. Monte-Carlo techniques for transport
Gauthier Villars, 1964. equation and diffusion. Berlin: Springer, 1998 in French.
[6] Chalmond B, Coldefy F, LavayssieÁre B. Tomographic reconstruction [17] Leo WR. Techniques for nuclear and particle physics experiments: a
from non-calibrated noisy projections in non-destructive evaluation. how-to approach. 2nd ed. Berlin: Springer, 1994.
Inverse Probl 1999;15:399±411. [18] Myers R, Montgomery D. Response surface methodology: process
[7] Chalmond B. Modeling for image analysis. Berlin: Springer, 2000 in and product optimization using designed experiments. New York:
French. Wiley, 1995.
[8] Della Rocca A, Ferriani S, La Porta L. Computer simulation of the [19] Schumm A. Ultrasonic ray models for complex geometries, Rev Prog
radiographic image formatting process: implementation and applica- QNDE, vol. 19. New York: Plenum Press, 1997.
tions. NDT&E Int 1995;28(3):7163±70. [20] Segre E. Nuclei and particles. 2nd ed 1977.
[9] Elsha®ey I, Gray JN. Optimization tool for X-ray radiographic NDE. [21] Spanier J, Gelbard EM. Monte Carlo principles and neutron transport
Rev Prog Quantitative Non Destruct Eval 1996;15 New York. problem. Reading, MA: Addison-Wesley, 1969.
[10] Fishman GS. Monte Carlo. Berlin: Springer, 1995. [22] Tillack GR, Bellon C, Artemiev V, Naumov AO. Simulation of
[11] GlieÁre A. Sindbad: from CAD model to synthetic radiographs. Rev scattering process in radiation techniques exploiting the theory of
Prog Quantitative Non Destruct Eval 1998;17 New York. Markovian processes with random structure. Rev Prog Quantitative
[12] Gray JN, Inanc F, Shull BE. Three dimensional modeling of Non Destruct Eval 1997:16.
projection radiography. Rev Prog Quantitative Non Destruct [23] Tillack GR, Arteniev VM, Naumov AO. Statistical linearization
Eval 1989;8A. method for analytical simulation in nonhomogeneous media.
[13] Hubbell JH. Photon cross sections, attenuation coef®cients and energy Bundesanstalt fur Materialforschung. Research Report, May
absorption coef®cients from 10 to 100 GeV. National Bureau of Stan- 1997.
dards NSRDR-NBS 1969;29 Washington. [24] Williamson FF. Monte-Carlo simulation of photon transport phenom-
[14] Inanc F. Analysis of X-ray and gamma ray scattering through compu- ena: sampling techniques. In: Morin RL, editor. Monte-Carlo simula-
tational experiments. J Nondestruct Eval 1999;18(2). tion in the radiological sciences, Boca Raton: CRC Press, 1988. p. 53±
[15] Jenkins TM. Monte Carlo transport of electrons and photons. New 101.
York: Plenum Press, 1988.