Combination of MCNP and SimSET for Monte Carlo simulation of SPECT with medium- and high-energy...

668 IEEE TRANSACTIONS ON NUCLEAR SCIENCE, VOL. 49, NO. 3, JUNE 2002

Combination of MCNP and SimSET forMonte Carlo Simulation of SPECT With

Medium- and High-Energy PhotonsY. Du, E. C. Frey, Member, IEEE, W. T. Wang, C. Tocharoenchai, W. H. Baird, and B. M. W. Tsui, Member, IEEE

Abstract—Monte Carlo (MC) simulation has become an impor-tant tool for characterizing and evaluating imaging systems andreconstruction algorithms. In this work, we combined two MCcodes—Monte Carlo N-particle transport code system (MCNP)and simulation system for emission tomography (SimSET)—usingthe advantages of each to produce a simulation tool that is efficientenough to generate single photon emission computed tomography(SPECT) data. The new SimSET-MCNP method allows the use ofvoxelized three-dimensional (3-D) phantoms and models photonpropagation inside collimator-detector systems. This combinationprovides a tool for evaluating compensation methods applied toimaging of agents where medium- and high-energy photons areimportant. To validate the new tool, we compared simulated pro-jections of a sphere containing I-123 water solution embedded ina cylindrical water-filled phantom with experimentally measuredprojections and simulations using SimSET and MCNP alone.Using these data, we compared profiles through the projectiondata, energy spectra, and relative number of photons in fourprojection views. The agreement with experiment was good,with disagreements of the order of a few percent. In addition,for the MC simulations, we classified detected photons basedon whether they scattered in the phantom, whether they passedthrough the collimator holes, penetrated the septa, or scatteredin the collimator, and whether they resulted from 159-keV orhigh-energy photons. For all these classes of photons, there wasexcellent agreement between SimSET-MCNP and MCNP. Finally,we evaluated the new combination in terms of simulation timeand found it significantly more efficient than MCNP alone. Weconclude that the new simulation tool works and allows thegeneration of SPECT data using voxelized phantoms for caseswhen medium- and high-energy photons are important.

Index Terms—Collimator scatter, Monte Carlo (MC) sim-ulation, septal penetration, single photon emission computedtomography (SPECT).

I. INTRODUCTION

M ONTE CARLO (MC) simulation has become an im-portant tool for characterizing and evaluating imaging

systems and reconstruction algorithms [1]–[10]. In the past, anumber of MC codes have been used in nuclear medicine. These

Manuscript received November 25, 2001; revised February 24, 2002. Thiswork was supported in part by the National Institutes of Health under GrantsR01-HL61616 and R01-CA39463.

The authors are with the University of North Carolina at Chapel Hill, ChapelHill, NC 27599 USA (e-mail: [email protected]; [email protected];[email protected]; [email protected]; [email protected];[email protected]).

Publisher Item Identifier S 0018-9499(02)06176-2.

include both general-purpose codes designed for a range of ap-plications and special-purpose codes designed specifically fornuclear medicine. The general-purpose codes generally havevery complete physics and the ability to model complex geomet-rically defined objects. However, they are often not designed tohandle voxelized phantoms and are inefficient in modeling nu-clear medicine imaging systems. The specialized codes oftenhandle voxelized phantoms and use variance reduction tech-niques such as forced detection to allow improved efficiency.However, they often have incomplete physics treatments (e.g.,they may not model Pb X-rays and interactions in the colli-mator) or model the collimator septa as being opaque.

MC N-particle transport code system (MCNP) is a general-purpose MC code for calculating the continuous-energy trans-port of particles [11]. MCNP only allows modeling of objectsdefined by the union and intersection of quadratic surfaces; thismay be insufficient to describe the complex shapes seen in med-ical imaging. Modeling voxelized phantoms with MCNP is dif-ficult and inefficient. Parallel hole collimators and gamma cam-eras can, however, be easily modeled in MCNP. However, thebrute force tracking of photons in MCNP is very inefficient,and simulation of low-noise projections for complex objects, asneeded in studies evaluating penetration and collimator scattercompensation in single photon emission computed tomography(SPECT), requires extremely long simulation times.

Simulation System for Emission Tomography (SimSET) isa package that uses MC techniques to model the physical pro-cesses and instrumentation used in emission imaging. SimSETincorporates a variety of variance reduction techniques that areappropriate for nuclear medicine simulations. Further, it allowssimulation using voxelized activity and attenuation distributions[12]. However, in SimSET, the collimator is not modeled usingMC photon tracking but rather by using an analytic model ofthe geometric transfer function [13]. Thus, scattering and pen-etration within the collimator are not modeled. For some appli-cations, especially when medium- and high-energy photons areinvolved, this will produce inaccurate results.

The goal of this work is to combine MCNP and SimSET,using the advantages of each to produce an MC simulation toolthat is efficient enough to generate SPECT data from voxelizedphantoms while modeling interactions in the collimator andcrystal. This will provide a tool for evaluating compensationmethods applied to imaging of agents where medium- andhigh-energy photons are important.

0018-9499/02$17.00 © 2002 IEEE

DU et al.: COMBINATION OF MCNP AND SimSET FOR MONTE CARLO SIMULATION OF SPECT 669

Fig. 1. Flowchart of the main steps of the combined SimSET-MCNP program.

II. M ETHODS

A. SimSET-MCNP Method

In the combined SimSET-MCNP method (Fig. 1), the firststep is to use SimSET to simulate photon propagation in a vox-elized 3-D phantom located in the object cylinder. In particular,SimSET is used to generate photons and transport them throughthe object until they exit the object cylinder. From there theyare propagated through a vacuum to the target cylinder. Theradial and axial extents of the target and object cylinders areuser-specified; they will usually be determined by the radial andaxial dimensions of the detection/collimation systems to be sim-ulated. Forced detection techniques are used to ensure that thephotons will be incident on the target cylinder. Weight controland stratification variance reduction techniques are used to im-prove the efficiency of the simulation. Information about pho-tons reaching the target cylinder, including the coordinates (,, ) of the photon on the surface of the target cylinder, the direc-

tion vector ( , , ) of the photon, the scatter order in the object,and the photon’s weight and energy are stored in a photon his-tory file. Each record of one photon is 68 bytes in size (8 doubleprecision real and one 4-byte integer values).

The history file is then processed by a modified version ofMCNP4b2 to generate a set of SPECT projection images. Thisversion of MCNP was previously modified to allow trackingof photon interactions and better modeling of a gamma camera[14]. In this work, we developed a source subroutine to read

Fig. 2. The off-center hot sphere in a cold cylindrical phantom (units are cm).

TABLE ITHE 15 MOST ABUNDANT I-123 PHOTONS

in the photon information stored in the SimSET photon historyfile and transform the photon’s coordinates and direction to thatused by MCNP. This source subroutine reads photon history in-formation one photon at a time. Every photon is used in MCNPonce for each projection view before reading in and startingpropagation of another photon. The photon’s coordinates anddirection are rotated for each projection view so that they arecorrect with respect to the view. The use of the same photon formultiple views introduces noise correlation between the views.However, for long simulations used to produce low noise pro-jections, these correlations should have little effect on the re-sulting images, especially when Poisson noise is added to them.The photon is then propagated through the collimator-detectorsystem by MCNP. Projection images or listmode data can besaved.

B. Simulations and Experiment

To verify this method (SimSET-MCNP), we simulated an off-center sphere containing an I-123-water solution embedded in awater-filled cylindrical phantom. The diameter of the cylindricalphantom is cm and the height is cm. Thediameter of the sphere is cm; its center is cmoff the phantom axis and cm away from the bottomof the cylindrical phantom (Fig. 2). Projections were computedusing a pixel size of 0.48 0.48 cm and a 15% energy windowcentered at 159.0 keV. We compared experimentally acquiredprojections to those simulated using SimSET, MCNP, and theSimSET-MCNP combination.

To ensure that the photon history file recorded all the photonsthat could hit the detector, the target cylinder was set to be thesame length as the detector. The 15 most abundant I-123 photonswere simulated (Table I).

Because SimSET can only simulate one photon energy eachtime, the 15 different energy photons were simulated separatelyfor SimSET and SimSET-MCNP and summed. For SimSETand SimSET-MCNP simulation, a voxelized three-dimensional


Fig. 3. Detector position in four views.

(3-D) phantom with voxel size of 0.27 0.27 0.27 cm wasused. A total of 128 projection views over 360 degrees weresimulated. For MCNP simulations, a geometric 3-D phantomwas used. Because MCNP can only simulate one view eachtime, only four views at 0, 90 , 180 , and 270 were sim-ulated (Fig. 3). Also, for the MCNP simulation, the 159-keVphotons were simulated separately from all the other photons.The energy resolution used in all these methods was 10% at159 keV and varied linearly with energy based on experimentalmeasurements. A 15% wide energy window centered at 159 keVwas used in all simulations. The ROR (radius-of-rotation) was15.5 cm. A total of two billion photons were simulated for eachmethod.

The experimental data were collected with a Siemens E.CAMSPECT system using a low-energy high-resolution (LEHR) col-limator. The phantom geometries were the same as those used inthe simulations. Listmode data were acquired for photons in therange 120–600 keV. Projection data were acquired at 128 viewsover 360 . Full energy spectra and projection images at 159 keVwith a 15% energy window were obtained by processing the list-mode file.

C. Evaluation and Comparison

We evaluated and compared the experimental and simulatedprojection data in the following ways. First, from the projec-tion images, we extracted image profiles across a ROI coveringthe phantom. The total counts in the ROI of the projection im-ages were also computed. For the MC simulations, we classifieddetected photons and made comparisons of detected counts forthe following photons: photons not scattered in the phantom andphotons scattered at least once in the phantom. For both of thesecategories, we further classified photons into the following cat-egories: geometrically collimated photons; photons penetratingthe collimator; and photons scattered in the collimator. For eachphoton class, the detected photons were also classified as thoseoriginating from 159-keV photons and those from higher energyphotons. By comparing the total counts in the various photonclasses, we can verify that the new method correctly simulatesphoton propagation inside the phantom and photon propagationinside the collimator as compared to SimSET and MCNP alone.

The energy spectra of the experiment and SimSET-MCNPsimulation were obtained from listmode data and compared.

The times required to simulate 500 million photons at pro-jection view3 as shown in Fig. 3 and 128 projection views usinga 1-GHz Pentium III CPU were compared for SimSET-MCNP

Fig. 4. The projection at view3. (a) The experiment result. (b) The MCNPsimulation. (c) The SimSET-MCNP simulation. (d) The SimSET simulation.Images are displayed in logarithmic gray scale. Note the experimental imagehas a different field of view.

Fig. 5. The penetration at view3 with different septa thickness in the horizontaldirection. The left image shows the penetration with normal septa thickness.The right image is the penetration with septa increasing by 50% in horizontaldirection. Images are displayed in a logarithmic gray scale.

and MCNP. The simulation times were compared separately for159-keV and high-energy photons.

III. RESULTS

Fig. 4 shows simulated and experimental projection imagesat the projection view3 as shown in Fig. 3. In Fig. 4, the projec-tion images from MCNP and SimSET-MCNP simulation lookvery similar to that from the experiment except that one of thepenetration “spokes” is dimmer in the experimental image. Thismay be due to the fact that one of the septa in the foil collimatoris formed by gluing two strips of lead, resulting in a differentthickness for that wall. To demonstrate the effect of a thickerseptum in one direction, we simulated the phantom after in-creasing the thickness of the septum in the horizontal directionby 50%. Fig. 5 shows the effect of this increase on the pene-tration component of the projection at view3. Note that thereis a reduction in the intensity of one of the spokes similar to


Fig. 6. Five row-wide profiles through the images in Fig. 4 across the hot sphere. They are normalized by the number of geometrically collimated photons.

TABLE IIRELATIVE TOTAL COUNTS IN FOUR PROJECTIONVIEWS

that observed in the experimental images. However, it is dif-ficult to quantitatively model this because the exact change inseptal thickness is hard to obtain. As expected, because SimSETdoes not simulate the interactions in the collimator no penetra-tion “spokes” are present.

Profiles of these four images taken through the area coveringthe hot sphere are shown in Fig. 6. Since SimSET can accuratelysimulate the geometric response of the collimator, these profilesare normalized by the number of geometrically collimated pho-tons. From Fig. 6, we can see that the profiles through the projec-tions images are in good agreement for MCNP, SimSET-MCNP,and the phantom experiment. The discrepancy between SimSETand the other curves illustrates the effect of penetration and col-limator-scattered photons. Table II shows the comparison of therelative total counts in four projection views. These counts arenormalized to view2. There is good agreement between MCNPand SimSET-MCNP. The difference between SimSET-MCNPand experiment in view4 is about 3%. For experimental data,the counts in view1 are fewer because a phantom support underthe phantom attenuates some photons in this view. In view3 ex-perimental counts are about 7% higher than SimSET-MCNP.This maybe due to the backscatter from the opposite detector,

TABLE IIIRELATIVE COUNTS OFGEOMETRICALLY COLLIMATED PHOTONSSCATTERED

INSIDE THE PHANTOM DIVIDED BY THOSEUNSCATTERED INPHANTOM IN

PROJECTIONVIEW1 FROM THREE SIMULATIONS

the floor, and the phantom support, which were not modeled inthe simulation.

Fig. 7 shows a comparison of the SimSET-MCNP simulatedspectrum and experimental spectrum of I-123 between 120–600keV. The spectra are normalized for equal area. There is somediscrepancy at higher energies, which indicates that the form ofenergy dependence of the energy resolution assumed in the MCsimulation is not accurate at higher energies.

In a SPECT study, photons scattered inside the patientplay an important role in the image quality. Table III lists therelative counts of geometrically collimated photons unscatteredin the phantom and those scattered in phantom in projectionview1 for the three MC simulation methods. All the data arenormalized to the total counts of geometrically collimated


Fig. 7. Comparison of energy spectra from experiment and SimSET-MCNP simulation.

TABLE IVRELATIVE COUNTS FROM DIFFERENTCOLLIMATOR COMPONENTSFROM

MCNP AND SIMSET-MCNP SIMULATIONS IN PROJECTIONVIEW3

Note: 159 means 159-keV photons; HE means higher energyphotons; all means all the 15 energies photons.

TABLE VTHE TIME IN HOURSREQUIRED TOSIMULATE 500 MILLION PHOTONS

photons which were not scattered inside the phantom. All threemethods are in good agreement. The differences are on theorder of 4% between the two independent methods SimSETand MCNP. SimSET-MCNP lies halfway between the SimSETand MCNP results. This indicates that the discrepancy betweenMCNP and SimSET is equally due to the scattering objectand the collimator-detector system. Part of the difference

Fig. 8. Projection images at view3 for the simulation of 500 million 159-keVphotons. The images are shown in a logarithmic gray scale.

between MCNP and SimSET-MCNP maybe due to the use of avoxelized phantom. Another advantage of the SimSET-MCNPand MCNP methods over SimSET is that they perform MCtracking of photon propagation inside the collimator. Thisprovides information about not only the geometrically colli-mated photons but also photons that penetrate the collimatorand photons scattered inside the collimator. Table IV lists therelative counts for different classes of the detected photons inthe projection view3. They are normalized to the geometricallycollimated photons which are not scattered in phantom. There isgood agreement between these two methods for all the photonclasses. For both methods, higher energy photons made upabout 30% of total detected counts even though they represent


Fig. 9. Energy dependence of the simulating time for SimSET and SimSET-MCNP for different energy. Note that the number of photons simulated is 1 millionfor SimSET-MCNP and 5 million for SimSET.

only 3.06% of emitted photons. These photons are mostly de-tected after penetrating through or scattering in the collimator.This demonstrates the importance of simulating higher energyphotons and collimator interactions in simulations of I-123imaging.

In Table V, the first row gives the hours required forSimSET-MCNP and MCNP to simulate 500 million 159-keVphotons at projection view3. The second row gives the hoursrequired for these two program to simulate 128 views for 500million photons of all the 15 energies. Here the time requiredby MCNP for 128 projection views was calculated as 128times the average time of the four views simulated. We can seethat, for only one projection view, SimSET-MCNP requires25% more time than MCNP to simulate the same number of159-keV photons. In this case the SimSET part of simulationtakes about 10 hours and the MCNP part takes only 5.35 h.However, for 128 views SimSET-MCNP uses less than halfthe time of MCNP to simulate all 15 photon energies. This isbecause in SimSET-MCNP simulation the photon generationand propagation in the phantom is simulated only once for allthe views, while in MCNP simulation this step is repeated foreach view. Also, inclusion of higher energy photons slows theMCNP simulation. Note that one thing not factored into thisanalysis is the effect of variance reduction in SimSET.

Fig. 8 shows the images of different classes of detected pho-tons at projection view3 from these two methods by simulating500 million 159-keV photons. We can see those images haveapproximately the same level of noise.

Fig. 9 shows the energy dependence of the simulation timefor the SimSET and SimSET-MCNP simulations. It indicatesthat the simulation time is a function of the photon energy for

SimSET-MCNP. The higher the photon energy, the more time isconsumed because the higher energy photons scatter more timesin the collimator.

IV. DISCUSSION ANDCONCLUSION

In this work we have developed a new MC simulation toolthat uses voxelized objects and models the photon propagationin the collimator and detector. We compared simulations fromthe new method with those from experiment and MCNP andSimSET simulations. From these we concluded that there isgood agreement in terms of the shape of projections, ratios ofcounts in various photon classes, and energy spectra between thevarious methods. Also, SPECT simulation times are substan-tially reduced for the new method compared to MCNP alone.Results also indicate that the higher energies photons make up to30% of the total detected counts in the I-123 projection images.The combination of SimSET and MCNP is useful for evaluationstudies where medium and high energy photons are important.

In the current method, SimSET and MCNP are running se-quentially, sharing photon data through a file. We are in theprocess of combining the two using pipes so that large (severalhundred MB) photon history files are not required. Also, to fur-ther improve the practicality of simulating SPECT for situationswhere collimator interactions are important, variance reductiontechniques can be used in MCNP to increase the efficiency ofthe simulation.

REFERENCES

[1] D. E. Raeside, “Monte Carlo principles and applications,”Phys. Med.Biol., vol. 21, no. 2, pp. 181–197, 1976.


[2] J. W. Beck, R. J. Jaszczak, R. E. Coleman, C. F. Sterner, and P. Nolte,“Analysis of SPECT including scatter and attenuation using sophisti-cated Monte Carlo modeling methods,”IEEE Trans. Nucl. Sci., vol.NS-29, pp. 506–511, 1982.

[3] M. Ljungberg and S. E. Strand, “A Monte Carlo program for the sim-ulation of scintillation camera characteristics,”Comput. Methods Prog.Biomed., vol. 29, pp. 257–272, 1989.

[4] J. C. Yanch, A. B. Dobrzeniecki, C. Ramanathan, and R. Behrman,“Physically realistic Monte Carlo simulation of source, collimator andtomographic data acquisition for emission computed tomography,”Phys. Med. Biol., vol. 37, pp. 853–870, 1992.

[5] M. S. Rosenthal and L. J. Henry, “Scattering in uniform media,”Phys.Med. Biol., vol. 35, pp. 265–274, 1990.

[6] H. Wang, R. J. Jaszczak, and R. E. Coleman, “Solid geometry-basedobject model for Monte Carlo simulated emission and transmission to-mographic imaging systems,”IEEE Trans. Med. Imaging, vol. 11, pp.361–372, 1992.

[7] D. J. de Vries, S. C. Moore, R. E. Zimmerman, S. P. Mueller, B. Fried-land, and R. C. Lanza, “Development and validation of a Monte Carlosimulation of photon transport in an Anger camera,”IEEE Trans. Med.Imaging, vol. 8, pp. 430–438, 1990.

[8] R. L. Harrison, S. B. Gillespie, S. D. Vannoy, M. S. Kaplan, and T.K. Lewellen, “A public domain simulation system for emission tomog-raphy: Photon tracking through heterogeneous attenuation using impor-tance sampling,”J. Nucl. Med., vol. 34, p. 60P, 1993.

[9] M. Ljungberg, S. E. Strand, and M. A. King,Monte Carlo Calculationsin Nuclear Medicine. Bristol, U.K.: IOP, 1998.

[10] H. W. A. M. de Jong, E. T. P. Slijpen, and F. J. Beekman, “Accelerationof Monte Carlo SPECT simulation using convolution-based forced de-tection,” IEEE Trans. Nucl. Sci., vol. 48, pp. 58–64, 2001.

[11] MCNP-A Genral Monte Carlo N-Particle Transport Code, Version 4B,J. F. Briesmeister, Ed., 1997.

[12] R. L. Harrison, S. D. Vannoy, D. R. Haynor, S. B. Gillispie, M. S. Kaplan,and T. K. Lewellen, “Preliminary experience with the photon historygenerator module of a public-domain simulation systems for emissiontopography,” in1994 IEEE Nuclear Science Symp. Medical ImagingConf., Norfolk, Va, Nov. 1994.

[13] B. M. W. Tsui and G. T. Gullberg, “The geometric transfer function forcone and fan beam collimators,”Phys. Med. Biol., vol. 35, no. 1, pp.81–93, 1990.

[14] W. T. Wang, E. C. Frey, and B. M. W. Tusi, “Parameterization of PbX-ray contamination in simultaneous Tl-201 and Tc-99m dual isotopeSPECT,” in1999 Conf. Rec. Nuclear Science Symp. Medical ImagingConf., Seattle, WA, 1999, to be published.

Combination of MCNP and SimSET for Monte Carlo simulation of SPECT with medium- and high-energy...

Documents

Transcript of Combination of MCNP and SimSET for Monte Carlo simulation of SPECT with medium- and high-energy...