arX

iv:1

411.

2932

v1 [

phys

ics.

com

p-ph

] 1

1 N

ov 2

014

Reconstruction of bremsstrahlung spectra from

attenuation data using generalized simulated annealing

O. H. Menina,b, A. S. Martineza,c, A. M. Costaa,d,∗

aDepartamento de Fısica, Faculdade de Filosofia, Ciencias e Letras de Ribeirao Preto,

Universidade de Sao Paulo, Av. Bandeirantes 3900, 14040-901, Ribeirao Preto, SP,

BrazilbInstituto Federal de Educacao, Ciencia e Tecnologia de Sao Paulo, Rua Americo

Ambrosio 269, 14169-263, Sertaozinho, SP, BrazilcInstituto Nacional de Ciencia e Tecnologia em Sistemas Complexos, Brazil

dInstituto Nacional de Ciencia e Tecnologia em Metrologia das Radiacoes em Medicina,

Brazil.

Abstract

The throughout knowledge of a X-ray beam spectrum is mandatory to assessthe quality of its source device. Since the techniques to directly measure-ment such spectra are expensive and laborious, the X-ray spectrum recon-struction using attenuation data has been a promising alternative. However,such reconstruction corresponds mathematically to an inverse, nonlinear andill-posed problem. Therefore, to solve it the use of powerful optimizationalgorithms and good regularization functions is required. Here, we present ageneralized simulated annealing algorithm combined with a suitable smooth-ing regularization function to solve the X-ray spectrum reconstruction inverseproblem. We also propose an approach to set the initial acceptance and visi-tation temperatures and a standardization of the objective function terms toautomatize the algorithm to address with different spectra range. Numericaltests considering three different reference spectra with its attenuation curveare presented. Results show that the algorithm provides good accuracy toretrieve the reference spectra shapes corroborating the central importance ofour regularization function and the performance improvement of the gener-alized simulated annealing compared to its classical version.

Keywords: X-ray, spectrometry, inverse problem, generalized simulated

∗Corresponding author. Telephone: +55-16-3602-4670 Fax: +55-16-3602-4887.Email address: amcosta@usp.br (A. M. Costa)

Preprint submitted to Radiation Physics and Chemistry November 12, 2014

annealing

1. Introduction

A diagnostic X-ray beam can be characterized by its photon fluence dis-tribution, which corresponds to the number of photons within each energyinterval in the spectrum range. Experimentally, this distribution is generallydetermined by direct measurement through the use of an energy dispersivesolid state detector together with appropriated electronics Cho et al. (2013).Since this technique is quite expensive and requires specialized labor, it hasbeen proposed an alternative approach to reconstruct X-ray fluence spectrafrom its attenuation curves data.

In its pioneer paper, L. Silberstein proposed a method based on theLaplace Transform to reconstruct X-ray spectra using an analytical approx-imation to the attenuation curve Silberstein (1933). In 1936, G. E. Belldiscusses Silverstein technique and tests its performance using experimentalcurves Bell (1936). Some decades later, J. W. Twidell presented a com-puter code to reconstruct iteratively X-ray spectra by error minimization be-tween experimental and numerical attenuation data Twidell (1970). In 1983,Kramer & Von Seggern published two reviews of the main techniques to solvethe X-ray spectra reconstruction problem and showed results obtained by theleast square method Kramer and Seggern (1983); Kramer (1983). One yearlater Rubio & Mainardi applied the Laplace Transform to reconstruct X-rayspectra with characteristic lines. Rubio and Mainardi (1984).

In the last years, several studies have been published dealing with ex-perimental and numerical aspects of the X-ray spectrum problem Delgado(2009); Raspa et al. (2010); Ali and Rogers (2012); Ali et al. (2012). Specif-ically, in 1998 Nisbet et al proposed the using of Classical Simulated An-nealing (CSA) algorithm to reconstruct X-ray spectra with energy between6MeV and 25MeV Nisbet et al. (1998) by the minimization of an objectivefunction. Typically, this function is defined as the error between the experi-mental and modeled data added to a regularization function through a factorλ. However, it was not clearly shown how the regularization function wasbuilt neither how each new solution has been generated from the previousones.

As it is widely known, simulated annealing is a stochastic optimiza-tion algorithm which has been employed to solve non convex maximiza-tion/minimization problems. Potential solutions are iteratively generated

2

through a suitable visitation distribution and whether accepted or not accord-ing to a predefined criterion. Its classical version (CSA) was formulated basedon the Metropolis acceptance criterion Metropolis et al. (1953) combined toa geometric or logarithmic temperature cooling schedule Kirkpatrick et al.(1983) and adopting a Gaussian visitation distribution. It has been mathe-matically shown that the logarithmic cooling temperature ensures the globalminimum convergence Geman and Geman (1984). Since such cooling sched-ule takes excessive computational time, a faster algorithm was proposed, theFast Simulated Annealing (FSA), which accomplishes a visitation distribu-tion based on Cauchy probability density function with the cooling scheduleas inverse of the iteration Szu and Hartley (1987).

By the introducing of two new parameters, qa and qv, Tsallis & Stariolodesigned the Generalized Simulated Annealing (GSA), which generalizes theMetropolis acceptance criterion, the visitation distribution and the coolingschedule, retrieving CSA and FSA as particular cases Tsallis and Stariolo(1996). Practical problems such as protein structure configurations andthe electrical impedance tomography inverse problem have also been tack-led using the simulated annealing algorithms Hansmann (2010); Menin et al.(2013).

Here we introduce an algorithm based on GSA combined with a suitablefirst derivative regularization function to solve the X-ray spectrum recon-struction inverse problem. A C language computer code has been imple-mented and numerical experiments have been carried out to survey the al-gorithm performance with respect to two main parameters, λ and qv, whichwe conveniently define as qv = qv − 1. Results show that our algorithm hassignificant performance improvement compared to the CSA version withoutregularization function and is able to retrieve the X-ray fluence distributionwith high accuracy. Moreover, results allow us to suggest a range of λ > 10−2

and qv > 0.6 that improves the algorithm performance.The paper is organized as follows: in Sec. 2, we state the X-ray fluence

distribution reconstruction inverse problem and present the generalized sim-ulated annealing algorithm. Also, we discuss the computational implementa-tion and present the numerical experiments. In Sec. 3, we show and discussthe results and, finally, in Sec. 4, we conclude stressing the effectiveness ofour method.

3

2. Experimental

In this section, we present the X-ray spectrum reconstruction inverseproblem formulation and our proposal for a regularization function. Then,we introduce the generalized simulated annealing algorithm fundamentals.Finally, we discuss the algorithm implementation and present the numericalexperiments.

2.1. X-ray spectrum reconstruction inverse problem

Consider a X-ray beam with unattenuated fluence distribution ΦE(E),as shown in Fig. (1-a). After crossing n attenuation plates with thicknesst1, t2, · · · , tn and densities ρ1, ρ2, · · · , ρn, as shown in Fig. (1-b), the attenu-ated fluence distribution Φ′

E(E) is

Φ′

E(E) = ΦE(E) exp

[

− (µ/ρ)n

∑

j=0

ρjtj

]

, (1)

where (µ/ρ) is the mass attenuation coefficient, which depends on E. Therelative transmission Tn after the X-ray beam cross n attenuation plates is

Tn =

∫ Emax

0ΦE(E)(µen/ρ) exp

[

−(µ/ρ)∑n

j=0 ρjtj

]

EdE∫ Emax

0ΦE(E)(µen/ρ)EdE

, (2)

where Emax is the spectrum energy upper limit and (µen/ρ) is the massenergy-absorption coefficient, which depends on E.

Then, one defines the X-ray spectrum reconstruction direct problem asthe determination of the transmission curve Tn from a known fluence distri-bution ΦE and the correspondent inverse problem as the determination ofthe fluence distribution ΦE from its transmission curve Tn. Discretizing thecontinuum fluence distribution ΦE(E) in a vector Φ =

[

φ(0), φ(1), · · · , φ(k−1)]

that represents a histogram with k energy bins, the integrals of Eq. (2)can be replaced by summations and the transmission becomes the vectorT =

[

T (0), T (1), · · · , T (m−1)]

, with components

T (n) =

k−1∑

s=0

an,sφ(s)

k−1∑

s=0

a0,sφ(s)

, for n = 0, 1, · · · , m− 1, (3)

4

(a)

(b)

Figure 1: (a) Typical diagnostic X-ray spectrum and (b) a X-ray beam with incidentfluence distribution ΦE(E) crossing n attenuation plates with thickness t1, t2, · · · , tn anddensities ρ1, ρ2, · · · , ρn. The attenuated fluence distribution is Φ′

E(E).

where m is the number of attenuation measurements and

an,s =

∫ E(s+1)

E(s)

(µen/ρ)(E) exp

[

−(µ/ρ)(E)

n∑

j=0

ρjtj

]

EdE. (4)

Thus, in our discretized X-ray spectrum reconstruction inverse prob-lem, we know the transmission data T and we must determine the flu-ence distribution Φ. Comparing the experimental transmission data S =[

S(0), S(1), · · · , S(m−1)]

with the numerical one T =[

T (0), T (1), · · · , T (m−1)]

generated from a trial fluence distribution Φtrial, we define an objective func-tion as

f (Φtrial) =1

m

m−1∑

n=0

[

T (n) − S(n)]2

+ λR(Φtrial), (5)

5

where R(Φtrial) is a regularization function and λ is the regularization pa-rameter.

The suitable choice to the regularization function R(Φtrial) is of funda-mental importance. In our algorithm, it has been constructed to guaranteethe smoothness of the spectrum penalizing large fluence differences betweentwo neighboring components of Φtrial. Thus, we have considered it as thefirst derivative of the standardized trial fluence distribution

R(Φtrial) =1

k − 2

k−1∑

s=1

∣

∣

∣φ(s)trial − φ

(s−1)trial

∣

∣

∣, (6)

where φ(s)trial are the elements of the standardized trial fluence

Φtrial =Φtrial

max[Φtrial]. (7)

Our propose for such standardization is important to makes the regular-ization function R(Φtrial) to have the same order magnitude of the absolute

mean square errorm−1∑

n=0

[

T (n) − S(n)]2/m of Eq. (5) and facilities the choice

of the regularization parameter λ, which represents the threshold betweenthe data fitting and the spectrum smoothness. Moreover, since T (n) and S(n)

are relative transmission, our standardization of the trial fluence distributionmakes the objective function Eq. (5) dimensionless.

Summarizing, in the X-ray reconstruction spectrum inverse problem, onemust find out a fluence distribution that yields the objective function globalminimum. Mathematically, the solution of the optimization problem Φ∗

trial

must satisfy

f(Φ∗

trial) < f(Φtrial), ∀ Φtrial 6= Φ∗

trial, (8)

which we are concerned in the following.

2.2. Generalized Simulated Annealing Algorithm

The algorithm starts from a random initial solution. At each iteration,a new guess solution Φnew is generated and compared with the previous oneΦold. Before describing the way new solutions are generated, let us addressthe acceptance criterion. According to Metropolis criterion, the probabilityp(Φold ← Φnew) to accept the new solution is

6

p(Φold ← Φnew) =

{

1, if ∆f ≤ 0e−∆f/Ta , if ∆f > 0

(9)

where ∆f = f(Φnew) − f(Φold) is the variation between the new and oldsolutions objective function and Ta is the acceptance temperature, whichcorresponds to a stochasticity control parameter.

To generate a new solution Φnew, each component φ(s)old of the current

solution Φold is modified, one at a time, generating a new component φ(s)new

through the visitation distribution

φ(s)new = φ

(s)old + TvGqv(φ), for s = 0, 1, · · · , k − 1, (10)

where Tv is the visitation temperature, which is related to the visitationlength jump, and

Gqv(φ) =Γ (1/qv)

√qv

Γ (1/qv − 1/2)√2π

exp−qv

[

−12φ2

]

(11)

is the standard q-Gaussian probability distribution function (pdf) within

expq(x) =

{

0, se qx < −1limq′→q (1 + q′x)1/q

′

, se qx ≥ −1 (12)

being the generalized exponential function Martinez et al. (2009).The attractive feature of the generalized Gaussian visitation is its ver-

satility. The suitable choice of qv parameter allows the algorithm transitfrom a global search to a local one and vice-versa. The 2nd moment ofthe q-Gaussian pdf (11) is only finite for qv < 2/3, which allows us to in-fer that for qv ≈ 0.7, one has the best agreement between local and globalvisitation Martinez et al. (2009). Tsallis & Stariolo, for example, estimatedthat the GSA best performance is reached to qv ≈ 1.7 Tsallis and Stariolo(1996), while Deng et al proposed an approach to generate q-Gaussian devi-ate and show numerically that the optimal GSA performance is to qv ≈ 1.3Deng et al. (2005).

The stochastic search is controlled by the temperature cooling schedule.The process starts with high temperatures allowing a wide spread visitationand high probability to accept a new solution that increases the objectivefunction. As the process approaches the end, the visitation becomes more lo-cal and the probability to accept bad new solutions tends to vanish. We have

7

chosen two schedules to decrease the temperatures Ta and Tv as a functionof the iteration t,

T (t)a = αtT (0)

a , (13)

where α ∈ (0, 1) is a cooling rate, and

T (t)v = T (0)

v

lnqv(2)

lnqv(t+ 1), (14)

with lnq(x) = limq′→q

[

(xq′ + 1)/q′]

being the generalized logarithm function,which is the inverse of the presented generalized exponential function Eq.(12).

It is important to note that as qv → 0, Eq. (11) and Eq. (14) become,respectively, the usuals Gaussian distribution and logarithm cooling schedule,retrieving the CSA. Also it is possible verify that for qv → 1 we obtainthe FSA with Cauchy visitation distribution and inverse of iteration coolingschedule.

One of the present paper contributions is the initial temperatures set-ting approach. To guarantee initial large probability to detrap from a lo-cal minimum basin the initial acceptance temperature T

(0)a must be greater

than the depth of deepest local minimum landscape. To estimate T(0)a , an

initial random search is performed in which µ solutions Φ1,Φ2, · · · ,Φµ aregenerated and its correspondent objective functions f(Φ1), f(Φ2), · · · , f(Φµ)are calculated. Evaluating the greatest difference of the objective function∆fµ = max [f(Φ)]−min [f(Φ)], we define

T (0)a = − ∆fµ

ln(p0), (15)

where p0 ∈ (0, 1) is a predefined initial probability to accept new solutionsthat increase the objective function and is set near to 1 to give the algorithminitial high probability to escape from local minima basin.

To ensure an efficient wide landscape exploration, the initial visitationtemperature, T

(0)v , must be greater than the largest actual spectrum fluency.

To simplify the T(0)v setting, we modeled the actual spectrum shape as a

triangle with base Emax − Emin and high T(0)v , where Emin and Emax are the

minimum and maximum spectrum energy. As the collision air kerma Kair

can be experimentally measured and it represents the area underneath of themodeled spectrum (triangle), we set

8

T (0)v =

2Kair

Emax − Emin. (16)

Summarizing, consider the following GSA algorithm, being tmax the max-imum number of iterations and r the number of new trial solutions at eachiteration.

1. set the simulation parameters: tmax, r, α, qv e λ;

2. set the initial temperatures T(0)a and T

(0)v using Eqs. (15) and (16);

3. generate a random initial solution Φold;

4. evaluate the objective function f (Φold) using Eq. (5);

5. for t from 1 to tmax:

(a) repeat r times the following steps:

i. generate a new solution Φnew using Eq. (10);ii. evaluate the new objective function f (Φnew) using Eq. (5);iii. Considere the Metropolis criterion:• if f (Φnew) ≤ f (Φold) , then Φold = Φnew;• else, if e[f(Φold)−f(Φnew)]/Ta > X ∼ U(0, 1), then Φold =Φnew;

(b) cooling down the temperatures Ta and Tv using Eqs. (13) and(14);

6. output results

2.3. Implementation and numerical experiments

A C language computer code to solve the X-ray spectrum reconstructioninverse problem have been implemented and numerical tests were carried outto assess its performance. Since our contributions are mainly related witha suitable regularization function and the use of GSA instead of CSA, thenumerical experiments have dealt specifically to the algorithm accuracy withrespect to the parameters qv and λ.

The implemented computer code receives, as input data, the collisionair kerma, Kair, the minimum and maximum spectrum energies, Emin andEmax, and the experimental transmission data, S. Thus, the GSA algorithm(Sec. 2.2) is throughout performed N times starting from different randominitial guess solutions, yields N numerical spectra Φ1,Φ2, · · · ,ΦN . The finalsolution is the numerical fluence distribution Φnum defined as

9

φ(s)num =

1

N

N∑

i=1

φ(s)i for s = 0, 1, · · · , k − 1, (17)

where φ(s)num is the numerical mean fluence for each energy bin s of Φnum and

k is the number of energy bins. To generate the qv-Gaussian deviate, we haveused the generalized Box-Muller method Thistleton et al. (2007).

Three experimental spectra data with peak voltages of 50 kVp, 60 kVpand 70 kVp were selected as reference to test the suitability of our algorithmAnkerhold (2000). The choice was made to avoid characteristic X-ray inaddition to the bremsstrahlung spectrum. The reference spectra minimumenergy, Emin, maximum energy, Emax, and number of energy bins k areshown in Tab. 1. Attenuation curves for aluminum considering m = 9different thickness were numerically generated and used as input data inour numerical experiments. The normalized fluence distributions for eachreference spectra and its relative transmission curves are shown in Fig. (2).

To compare the numerical solution Φnum to the actual one Φactual, wehave chosen the absolute mean error

ε =1

k − 1

k−1∑

s=0

∣

∣

∣φ(s)actual − φ(s)

num

∣

∣

∣. (18)

where φ(s)actual and φ

(s)num are the component s of the actual and numerical

normalized fluence distributions given, respectively, by

Φactual =Φactual

max[Φactual]and Φnum =

Φnum

max[Φnum]

Table 1: X-ray reference spectra with its energy range and number of energy binsAnkerhold (2000)

Spectrum Emin (kev) Emax (kev) k1 10 52 2112 9 62 1073 9 72 127

For each reference spectra, the program has run setting N = 25, tmax =200, r = 50 and α = p0 = 0.9 constants, but varying qv and λ in the rangeqv = 0, 0.2, 0.4, · · · , 2.0 and λ = 10−6, 10−5, · · · , 104. To each pair (qv,λ), the

10

0 10 20 30 40 50 60 70 800

0.2

0.4

0.6

0.8

1

Reference Spectrum 1

E(keV)

Φ/Φ

max

0 1 2 3 4 5 6 7 80

0.2

0.4

0.6

0.8

1

n

Sn

0 10 20 30 40 50 60 70 800

0.2

0.4

0.6

0.8

1

Reference Spectrum 2

E(keV)

Φ/Φ

max

0 1 2 3 4 5 6 7 80

0.2

0.4

0.6

0.8

1

n

Sn

0 10 20 30 40 50 60 70 800

0.2

0.4

0.6

0.8

1

Reference Spectrum 3

E(keV)

Φ/Φ

max

0 1 2 3 4 5 6 7 80

0.2

0.4

0.6

0.8

1

n

Sn

Figure 2: Normalized fluence distributions Φ/Φmax and the correspondent transmissionSn curves for the reference spectra.

numerical fluence distribution Φnum was compared to the actual one Φactual

yielding 11× 11 = 121 absolute mean errors calculated by Eq. (18).

3. Results and discussion

Now we present and analyze the main results obtained from the numer-ical experiments. First, we present in Fig. (3) the absolute mean error εbehavior as function of parameters λ and qv. Results does not allow us findout an unique “optimal” pair (λ, qv). However it is notable that the errordecreases strongly for λ > 10−2 to all reference spectra, which corroboratesthe fundamental importance of our regularization function. Also, one cansee considerable improvement on the results for qv > 0.6. Indeed, as we havediscussed in Sec. 2.2, the 2nd moment of pdf (11) diverges for qv > 2/3 andthe visitation becomes more global and efficient.

Beyond that, to show how the numerical fluence distributions fit to thereference actual spectra, we present in Fig. (4) the results obtained usingCSA (qv = 0) without regularization function (λ = 0) and the best fluencedistributions reached by GSA combined with our regularization function.Despite these best results have occurred for different values of qv and λ for

11

each reference spectra, it is quite evident that for qv > 0.6 and λ > 10−2 thealgorithm provides excellent fitness to the actual data.

Finally, although the Metropolis criterion might be also generalized, wehave chosen its classical version since such generalization is still under in-vestigation. Indeed, one year after the original paper of Tsallis & StarioloTsallis and Stariolo (1996), Xiang et al proposed an rule to change the pa-rameter qa to speed up the convergence of GSA Xiang et al. (1997). In 2009Crokidakis et al have used a modified Metropolis criterion based on gener-alized Boltzmann-Gibbs entropy to analysis phase transitions of the Isingmodel Crokidakis et al. (2009) and three years later, Silva et al Silva et al.(2012) have retrieved the detailed energy balance and locality of the gen-eralized Metropolis criterion using the generalized operators Arruda et al.(2008).

4. Conclusions

A technique based on generalized simulated annealing algorithm com-bined to a first derivative regularization function to solve the X-ray spec-trum reconstruction inverse problem was presented. A computer code thatreceives a X-ray attenuation data and returns an approximation for its fluencedistribution was implemented. Numerical experiments using three referencespectra with 50 kVp, 60 kVp and 70 kVp were carried out to assess the algo-rithm performance mainly with respect to the parameters λ and qv. Resultsshow high accuracy improvement for λ > 10−2 and qv > 0.6, corroboratingthe GSA better performance compared to CSA and showing the fundamentalimportance of our smoothness regularization function. Moreover, our tech-nique seems to be quite reliable since its numerical reconstruction fluencedistributions present excellent fitness to the reference spectra. Finally, thepresented technique is promising to assist the tricky practical problem of de-termination of a X-ray spectrum beam from its attenuation data, which isconsiderably less expensive and easier to accomplish than the current directmeasurement approaches.

Acknowledgements

OHM thanks the IFSP and CAPES. ASM thanks the CNPq (305738/2010-0 and 485155-2013-3). AMC thanks the CNPq and FAPESP.

12

References

References

Ali E S M, McEwen M R and Rogers D W O 2012 Unfolding linac pho-ton spectra and incident electron energies from experimental transmissiondata, with direct independent validation Med. Phys. 39, 6585–6595.

Ali E S M and Rogers D W 2012 An improved physics-based approach forunfolding megavoltage bremsstrahlung spectra using transmission analysisMed. Phys. 39, 1633–1675.

Ankerhold U 2000 Catalogue of X-ray spectrum and their characteristic data-iso and din radiation qualities, therapy and diagnostic radiation qualities,unfiltered X-ray spectra Technical report PTB - Report DOS-34.

Arruda T J, Gonzales R S, Tercariol C S A and Martinez A S 2008 Arithmeti-cal and geometrical means of generalized logarithmic and exponential func-tions: Generalized sum and product operators Phys. Lett. A 372, 2578–2582.

Bell G E 1936 Spectral distribution in the continuous X-ray spectrum andthe specification of a X-ray quality Brit. J. Radiol. 9, 680–688.

Cho H M, Kim H J, Choi Y N, Lee S W, Ryu H J and Lee Y J 2013 Theeffects of photon flux on energy spectra and imaging characteristics in aphoton-counting X-ray detector Phys. Med. Biol. 58, 4865–4879.

Crokidakis N, Soares-Pinto D, Reis M, Souza A, Sarthour R and Oliveira I2009 Finite-size analysis of a two-dimensional Ising model within a nonex-tensive approach Phys. Rev. E 80, 051101–1–051101–8.

Delgado V 2009 Determination of X-ray spectra from Al attenuation databy imposing a priori physical features of the spectrum: Theory and exper-imental validation Med. Phys. 36, 142–148.

Deng J, Chang C and Yang Z 2005 An exact random number generator forvisiting distribution in GSA I. J. of Simulation 6, 54–61.

Geman S and Geman D 1984 Stochastic relaxation, Gibbs distributions, andthe bayesian restoration of images IEEE Trans. Patt. Anan. Mach. Int.

PMA1-6(6), 721–741.

13

Hansmann U H E 2010 Temperature random walk sampling of protein con-figurations Physica A 389, 1400–1404.

Kirkpatrick S, Gellat C and Vecchi M 1983 Optimization by simulated an-nealing Science 220, 671–680.

Kramer H M 1983 The determination of X-ray spectra from attenuatin data.Part II: Experimental results Nucl. Instrum. Meth. Phys. 214, 445–450.

Kramer H and Seggern H V 1983 The determination of X-ray spectra fromattenuation data. Part I: The potentials of various methods Nucl. Instrum.

Meth. Phys. 213, 373–380.

Martinez A S, Gonzalez R S and Tercariol C A S 2009 Generalized probabilityfunctions Advances in Math. Phys. 2009, 1–13.

Menin O H, Rolnik V and Martinez A S 2013 Boundary element method andsimulated annealing algorithm applied to electrical impedance tomographyimage reconstruction Rev. Bras. Ens. de Fıs. 35, 2304–1–2304–7.

Metropolis N, Rosenbluth A, Rosenbluth M, Teller A and Teller E 1953Equation of state calculation by fast computing machines J. Chem. Phys.

21, 1087–1091.

Nisbet A, Weatherburn , Fenwick J D and McVey G 1998 Spectral recon-struction of clinical megavoltage photon beams and the implications ofspectral determination on the dosimetry of such beams Phys. Med. Biol.

43, 1507–1521.

Raspa V, Knoblauch P, Lorenzo F D and Moreno C 2010 Plasma focus basedflash hard X-ray source in the 100 keV region with reproducible spectrumPhys. Lett. A 374, 4675–4677.

Rubio M and Mainardi R T 1984 Determination of X-ray spectra includ-ing characteristic line intensities from attenuation data Phys. Med. Biol.

29, 1371–1376.

Silberstein L 1933 Spectral composition of an X-ray radiation determinedfrom its filtration curve Phil. Mag. 15:98, 375–394.

14

Silva R, de Felıcio J R D and Martinez A S 2012 Generalized Metropo-lis dynamics with a generalized master equation: an approach for time-independent and time-dependent Monte Carlo simulations of generalizedspin systems Phys. Rev. E 85, 066707–1–066707–9.

Szu H and Hartley R 1987 Fast simulated annealing Phys. Lett. A 122, 157–162.

Thistleton W, Marsh J A, Nelson K and Tsallis C 2007 Generalized Box-Mller Method for generating q-Gaussian random deviates IEEE Trans.

Inf. Theory 53, 4805–4810.

Tsallis C and Stariolo D A 1996 Generalized simulated annealing Physica A

233, 395–406.

Twidell J W 1970 The determination of X-ray spectra using attenuationmeasurements and a computer program Phys. Med. Biol. 15, 529–539.

Xiang Y, Sun D Y, Fan W and Gong X G 1997 Generalized simulated an-nealing algorithm and its application to the Thomson model Phys. Lett.A 233, 216–220.

15

1E−61E−4

1E−21E+0

1E+21E+4 0

0.40.8

1.21.6

20

0.05

0.1

0.15

0.2

qvλ

ε

1E−6 1E−4 1E−2 1E+0 1E+2 1E+40

0.4

0.8

1.2

1.6

2

λ

qv

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

(a)

1E−61E−4

1E−21E+0

1E+21E+4 0

0.40.8

1.21.6

20

0.05

0.1

0.15

0.2

qvλ

ε

1E−6 1E−4 1E−2 1E+0 1E+2 1E+40

0.4

0.8

1.2

1.6

2

λ

qv

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

(b)

1E−61E−4

1E−21E+0

1E+21E+4 0

0.40.8

1.21.6

20

0.05

0.1

0.15

0.2

qvλ

ε

1E−6 1E−4 1E−2 1E+0 1E+2 1E+40

0.4

0.8

1.2

1.6

2

λ

qv

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

(c)

Figure 3: 3-D surface viewing (left) and projection on λ− qv plane (right) of the absolutemean error ε as function of qv and λ for reference (a) Spectrum 1, (b) Spectrum 2 and (c)Spectrum 3. The error decreases drastically for λ > 10−2 and for q > 0.6.

16

0 20 40 60 800

0.2

0.4

0.6

0.8

1

E (kev)

Φ/Φ

max

qv = 0

λ = 0

0 20 40 60 800

0.2

0.4

0.6

0.8

1

E (kev)

Φ/Φ

max

qv = 0.8

λ = 1E+0

(a)

0 20 40 60 800

0.2

0.4

0.6

0.8

1

E (kev)

Φ/Φ

max

qv = 0

λ = 0

0 20 40 60 800

0.2

0.4

0.6

0.8

1

E (kev)

Φ/Φ

max

qv = 1.6

λ = 1E+4

(b)

0 20 40 60 800

0.2

0.4

0.6

0.8

1

E (kev)

Φ/Φ

max

qv = 0

λ = 0

0 20 40 60 800

0.2

0.4

0.6

0.8

1

E (kev)

Φ/Φ

max

qv = 1.6

λ = 1E−1

(c)

Figure 4: The actual fluence distribution (- - -) and the worst (left) and the best (right)numerical fluence distribution (◦) for the reference (a) Spectrum 1, (b) Spectrum 2 and(c) Spectrum 3. It is evident the better fitness of GSA combined to our regularizationfunction (λ > 10−2 and q > 0.8) compared to CSA without regularization (λ = 0 andq = 0).

17