Design Of A Proton Computed Tomography System For Application In Proton Radiation Therapy

Design Of A Proton Computed Tomography System For Application In Proton Radiation Therapy - Transcript

Design of a Proton Computed Tomography System for Applications in Proton Radiation Therapy
Reinhard Schulte Vladimir Bashkirov Tianfang Li Jerome Z Liang Klaus Mueller Jason Heimann Leah R Johnson Brian Keeney Hartmut F W Sadrozinski Abraham Seiden David C Williams Lan Zhang Zheng Li Steven Peggs Todd Satogata and Craig Woody
results of a design study for the development and implementation of pCT in a proton treatment center II DESIGN SPECIFICATIONS The requirements for a pCT system designed for applications in proton therapy are dictated by the needs for accurate and safe proton beam delivery The system must be integrated into a medical environment and therefore meet certain safety and practical constraints The design specifications of a pCT scanner for applications in proton therapy are presented in Table I
TABLE I DESIGN SPECIFICATIONS FOR A PCT SCANNER FOR THERAPEUTIC
APPLICATIONS

Abstract Proton computed tomography pCT has the potential to improve the accuracy of dose calculations for proton treatment planning and will also be useful for pretreatment verification of patient positioning relative to the proton beam A design study was performed to define the optimal approach to a pCT system based on specifications for applications in proton therapy Conceptual and detailed design of a pCT system is presented consisting of a silicon based particle tracking system and a crystal calorimeter to measure energy loss of individual protons We discuss the formation of pCT images based on the reconstruction of volume electron density maps and the suitability of analytic and statistical algorithms for image reconstruction
I ROTON

INTRODUCTION

radiation in Pradiation therapy duehas many proven toadvantageshigh to its capability deliver doses to well defined tumors or other targets close to critical normal structures For proton therapy to be successful the range of protons in tissue must be accurately known Up to now proton dose calculations have been performed using xray computed tomography xCT However the accuracy of xCT for proton treatment planning is limited due the difference in physical interactions between photons and protons which partially obviates the advantage of proton therapy There would be additional advantages of applying proton computed tomography pCT in proton therapy including the possibility to verify the correct delivery of a proton treatment plan while the patient is in treatment position This paper reports on the
Manuscript received October 29 2003 This work was supported by USPHS grant CA46295 and the National Medical Technology Testbed NMTB under the U S Army Medical Research Acquisition Activity DAMD17 97 2 7016 The views and conclusions contained in this presentation are those of the authors and do not necessarily reflect the position or the policy of the U S Army or NMTB R Schulte and V Bashkirov are with the Department of Radiation Medicine Loma Linda University Medical Center Loma Linda CA 92354 USA telephone 909 558 4243 email rschulte dominion llumc edu T Li Z Liang and K Mueller are with the Department of Physics and Astronomy Radiology and Computer Science State University of New York Stony Brook NY 11790 USA J Heimann L R Johnson H F W Sadrozinski A Seiden D C Williams and L Zhang are with the Santa Cruz Institute for Particle Physics University of California Santa Cruz CA 95064 USA Z Li S Peggs T Satogata and C Woody are with the Brookhaven National Lab Upton NY 11973 USA

Category Parameter Proton source Energy Energy spread Beam intensity Spatial resolution Electron density resolution Installation time Data acquisition time Reconstruction time

Accuracy

Value 200 MeV head 250 MeV trunk 0 1 103 105 protons sec 1 mm 1

Time Efficiency

Reliability

Safety

Detector radiation hardness Measurement stability Maximum dose per scan 5 cGy Minimum distance to 10 cm

10 min 5 min 15 min treatment planning 5 min verification 1000 Gy 1

The protons used for imaging must have sufficient energy to penetrate the body part to be imaged According to the NIST PSTAR data base 1 the continuous slow down approximation CSDA range of 200 MeV protons in A150 tissue equivalent plastic is 25 8 cm which is sufficient to penetrate an adult human skull nominal width of 20 cm in anterior posterior direction For 250 MeV protons the range

in A150 is 37 7 cm sufficient to penetrate an adult trunk nominal width of 34 cm excluding arms In order the meet the specified accuracy on electron density resolution the energy should be stable to within 0 1 or 0 25 MeV for a proton energy of 250 MeV The spatial and electron density resolutions of a pCT scanner are physically limited by multiple Coulomb scattering MCS and energy loss straggling The spatial and energy uncertainties of the pCT system should be considerably smaller than those imposed by the physical limitations in order not to compromise the overall performance of the CT system A clinically meaningful spatial resolution for therapy with protons is about 1 mm This resolution value is related to what is achievable in terms of target localization and patient positioning accuracy and is also matched by the steepness of the lateral and distal fall off of the high dose region For example in radiosurgery applications where large doses of radiation are delivered to intracranial targets that are often within a few millimeters of critical normal structures at risk dose localization accuracy requirements of the order of 1 mm are usually cited 2 The electron density resolution requirement is closely related to the need of a spatial resolution 1 mm In order to place the distal edge of a proton beam with 10 cm range with 1 mm 1 accuracy one needs to know the density of each tissue voxel along the beam path with better than 1 accuracy Human observer perception of tissue density differences of about 1 is also important in order to assist the physician in identification of tissue structures for treatment planning and verification purposes Density differences between various soft tissues are typically in the 0 5 to 1 range 3 A time efficient pCT system must have a short installation calibration scanning and removal time Installation and removal of the system may be required if the detectors are not sufficiently radiation hard to stay in the beam line permanently The time required for scanning a patient for a treatment planning study including installation and removal of a non permanent system should not be longer than 15 minutes excluding image reconstruction time When the system is used as an on line position and dose verification system during a treatment session fast image reconstruction within about 5 minutes is an essential requirement Detectors should be sufficiently radiation hard to maintain their function to within 1 of specified performance values for at least one year ideally for 5 years or more Furthermore the system should be relatively insensitive to changes in temperature humidity and magnetic fields present in the treatment room A reasonable compromise between the dose delivered and the accuracy of electron density determination must be found tailored to the clinical situation The typical dose delivered by existing CT scanners during a scan for treatment planning purpose 3 5 cGy may serve as a benchmark and should not be exceeded by the pCT scan under similar conditions If the scanner is used for pretreatment verification the dose per scan

times the number of treatment sessions should not exceed this benchmark value Thus the density resolution for treatment verification would be lower than that for treatment planning III PCT DESIGN CONCEPT The proposed approach to single proton tracking CT is illustrated in Fig 1 The object is traversed by a broad ideally but not necessarily parallel beam of protons of known energy Ein A proton tracking detector is arranged on both sides of the patient which records the entrance and exit points and angles of individual protons Protons are stopped in a scintillator array to measure their energy Technical considerations for these components of the pCT system will be presented below
x


z

t Ein

s
y



Eout

u Fig 1 Schematic of the proposed approach to pCT Protons with known entry energy Ein are recorded one by one in the detector reference system s t u as they traverse the image object from many different projection angles The recorded data include entry and exit positions and entry and exit angles as well as exit energy Eout in the energy detector

IV PROTON TRACKING SYSTEM To determine the most likely proton path entry and exit points as well as directions must be measured with a spatial accuracy better than the image pixel size 1 mm 1 mm This requires pairs of 2D position sensitive tracking systems on both sides of the patient In order to keep the scanning time reasonably short the particle tracking system should be able to detect individual protons at rates of 1 MHz or higher Several technical solutions are possible including gas based detectors silicon detectors and scintillating fibers Silicon Si detectors now widely used as vertex detectors in high energy physics are the preferred choice for pCT because of their unrivaled spatial resolution at high event rates There are two major Si detector structures for tracking applications Si strip detectors SSD and Si pixel detectors SPD For two dimensional 2D position sensitivity X Y double sided SSD D SSD and pixel detectors are the most frequently used detector structures for tracking purposes The advantages of D SSD are 1 2D position sensitivity 2 minimum readout channel number 2N N being the strip number on one side 3 minimum radiation length because only one wafer is used for 2D position sensitivity and 4 fast charge collection time 50 ns There are some major disadvantages for D SSD however 1 the two sided manufacturing process is about 3 4 times more complicated and expensive than the single sided process 2 the detector is relatively radiation soft due to the complicated structure on the n side 3 there is an ambiguity problem when a strip is hit

by two or more protons simultaneously and 4 two polarities of readout electronics are required The maximum rate of DSSD is limited by the charge collection time the multi hit ambiguity problem and the shaping time in readout electronics With fast shaping time the rate can be 1 10 MHz detector chip which would be sufficient The maximum detector chip size is limited by the maximum wafer size in the high resistivity detector industry which is currently at 150 mm diameter resulting in a maximum chip size of about 10 6 cm 10 6 cm Therefore a modular design is needed to cover a larger anatomical area The advantages of SPD are 1 2D position sensitivity 2 single sided process 3 no multi hit ambiguity and 4 fast charge collection time 50 ns The main disadvantages of SPD are 1 large readout electronic channel number N2 N being the number of rows or columns 2 complicated and difficult bumper bonding technology for interconnections between the detector chip and readout electronics chip 3 additional radiation length in the readout electronics chip 4 position resolution limited by the minimum size of bumper bonding pads to 20 m not critical for our application and 5 high overall cost due to the large number of electronic channels and cost of bumper bonding The maximum rate and detector chip size of SPD are similar to that of SSD A novel detector structure called Si stripixel detector SSPD was recently developed at BNL 4 and would be suitable for the purpose of pCT In the SSPD structure each pixel is divided into an X and Y cell which are connected and read out by projective X and Y strips respectively To get 2D position sensitivity it is essential that for each proton hit the resulting charge is shared between the X cell and Y cell in a given pixel Since the charge spreads during the drifting time tens of ns due the diffusion process the charge cloud has a finite size on the order of 20 m In order to get charge sharing between the X cell and Y cell each pixel should be divided in such way that the cells are interleaved with maximum distance between the two parts 20 m 4 The main advantage of the novel SSPD is that it achieves 2D position sensitivity with the simpler single sided process Thus it combines the advantages of both SSD and SPD technology The remaining although not limiting disadvantage is that it is essentially a strip layout i e it has the same multi hit ambiguity as the DSSD V ENERGY DETECTOR The residual energy of protons traversing the image object is the single most important quantity in pCT The accuracy of this measurement determines the density resolution of the pCT method Although one may contemplate using the position sensitive silicon detectors to determine the residual energy via dE dx measurements in the depletion layer of these detectors this method is inherently inaccurate due to the weak dependence of stopping power on proton energy Therefore a separate energy detector is required

To measure the residual energy of a single proton with better than 1 accuracy using a compact detector is not trivial Several solutions have been suggested such as magnetic spectrometers 5 or a range telescope 6 Another possibility to measure energy is a calorimeter consisting of a single or multi array of crystal scintillators which convert radiationinduced ionizations into scintillation light collected by photomultiplier tubes or large area photodiodes Photodiodes would be preferable in the case of pCT because they are not as sensitive to the magnetic fields present in the proton gantry Calorimeters have been widely used during last decade for charged particle detection and identification in experiments where good energy resolution is an essential requirement see 7 10 and references therein
TABLE II CHARACTERISTICS OF CRYSTAL SCINTILLATORS SUITABLE FOR PCT
APPLICATIONS

P arameter Density g cm 3 R ange cm of 250 MeV protons Refractive index Radiation hardness Decay time ns fast slow Light outputa fast slow
a

CsI pure 4 51 14 9 1 8 good 10 36 1000 0 1 0 02

Bi4Ge3O12 BGO 7 13 9 0 2 2 modest 300

Lu 2SO5 Ce LSO 7 4 8 4 1 82 good 40

0 15

0 65

relative to NaI

Many options for scintillating crystals exist Table II compares the physical characteristics of three types of crystal scintillators The statistical uncertainty of the calorimeter output arises from the variation of photoelectron production for a given quantum efficiency The light yield of virtually all crystal scintillators is sufficient to give a statistical contribution to the energy resolution of less than 1 for 200 MeV protons To achieve this resolution one needs a minimum of 104 photoelectrons which with reasonable light collection efficiency and quantum efficiency of the readout device can be easily achieved with any of the scintillators under consideration Due to its high density fast decay and high light yield LSO may be preferred for pCT application VI IMAGE RECONSTRUCTION A Physics of Proton Image Formation The pCT reconstruction problem differs in some respects from that of xCT PET and SPECT but underlying principles are the same In the latter image modalities data collection is usually considered as the Radon transform of the object source function For x CT for example the object data represent the attenuation coefficient map and the projection data the log values of the detected x ray counts

The main goal of pCT for therapy applications is the determination of the volume electron density e by measuring the energy loss of protons after traversing the object Ionization and atomic excitation are the main processes for the energy loss of protons The mean rate of the loss per unit track length i e the stopping power is given by the Bethe Bloch equation dE r e r F I r E r 1 dx where e r is the electron density relative to water r represents the spatial location I r is the mean ionization potential of the medium E r is the proton energy which changes with r as the proton travels through the body and F is a known function of I and E defined by the Bethe Bloch equation 11 With reasonable assumptions and approximations the non linear differential equation 1 can be integrated as Eout dE e r dl 2 L Ein F I water E which would be in the format of the Radon transform if the proton path were a straight line the right side is the line integral of the relative electron density along the proton path L and the left side is a unique function of the proton energy difference In equation 2 I r was replaced by the mean ionization potential of water Iwater 61 77 eV This is justified because in human tissues the variation of I is not very large and the dependence of the function F on I is relatively weak In pCT multiply scattered protons traversing the object travel along a curved zigzag path which may deviate significantly from a straight line see Fig 2 Furthermore protons usually do not get absorbed but traverse the object completely Thus the photon counting rate used in x ray CT PET and SPECT has to be replaced by the energy loss measurement for protons traveling along tracks L that lead to the same image pixel Given the known proton entrance energy and the measured exit energy the energy integral of equation 2 can be computed resulting in the projection data The image reconstruction problem for pCT is then to obtain the best estimate for the relative electron density map from the measured proton data





Fig 2 Monte Carlo simulated tracks of 500 MeV protons traversing an object of water density with 50 cm diameter The solid lines correspond to the most likely path of protons with an exit displacement of 0 1 3 and 5 mm while the dashed lines represent the one sigma envelopes caused by MCS

B Reconstruction Algorithms Since proton paths are not straight lines the exact inversion of the pCT integral equation is impossible The wellestablished analytical reconstruction algorithms based on the inversion of the Radon transform such as the filtered back projection FBP algorithm are therefore of limited use On the other hand it has been shown in pCT simulation studies that the straight line approximation can lead to pCT images with reasonable spatial and density resolution 12 The FBP algorithm which is fast and robust may be sufficient if pCT is used for pretreatment verification of patient position In order to find a more exact solution of the object function e r one may resort to iterative statistical reconstruction algorithms In general terms the discrete data acquisition model for pCT may be expressed as yi E rin rout in out M ij rin rout in out j j
ni rin rout in out bi

3 where yi is the integral described by the left side of equation 2 for the energy loss E Ein Eout the matrix element Mij is the product of the probability that a proton detected in pixel i passed the voxel j and the path length through that voxel approximated by the voxel size j is the relative electron density of the voxel j ni is the noise in pixel i and bi is the background signal of the energy detector in pixel i The arguments rin rout and in out are the entry and exit positions and directions of the protons respectively Note that the pCT image noise which is determined by a combination of detector measurement uncertainty and energy straggling is more

complicated than the Poisson distributed noise in xCT SPECT and PET Monte Carlo simulations of radiation transport may be the ideal tool to model the noise in pCT however the computing effort would be very large Iterative algorithms are usually based on the maximization of the maximum likelihood ML or the maximum a posteriori probability MAP For the Poisson noise model used in PET and SPECT the ML expectation maximization ML EM 13 14 and MAP EM 15 16 reconstruction techniques have been extensively investigated Equivalent algorithms for pCT have yet to be developed With the ordered subset OS strategy 17 the computational burden for both ML EM and MAP EM is acceptable and the same strategy may be applicable to pCT reconstruction The ultimate approach to pCT reconstruction may be listmode data acquisition which records individual proton entrance and exit positions and directions exit energies and loss due to nuclear interactions as an event The pCT image is then reconstructed from each recorded event It is clear that this will result in the most accurate image because the listmode data contain the largest amount of information about the imaged object It is also obvious that this reconstruction will consume most computing time because the reconstruction has to process each of many million events individually rather than sorting millions of events into a few hundred projection data sets and processing each projection dataset as a group simultaneously However using modern hardware acceleration it is likely that reasonable reconstruction times can be achieved C Hardware Acceleration Clinical utility of pCT demands fast reconstruction within minutes Software based implementations of the statistical reconstruction algorithms described above may be too slow to achieve this goal Alternatively one may perform the reconstruction on commodity graphics hardware boards GPUs 18 The reconstruction is accomplished by using the texture mapping facilities of the hardware for performing the interpolations required in the projection and back projection operations 19 The other arithmetic operations that occur in the reconstruction algorithms may also be performed in this hardware In addition one may incorporate some Monte Carlo simulation capabilities directly into the hardware reconstruction platform for modeling of the MCS processes during the 3D reconstruction VII CONCLUSIONS The development of pCT is desirable to improve the accuracy of planning and delivery of proton radiation treatments and to fully exploit the advantages of the proton therapy A state of the art pCT system should use single proton registration with silicon based track detectors and a crystal calorimeter for energy measurement Effective iterative reconstruction algorithms taking into account the peculiarities of proton transport through the object have yet to be optimized

and may require hardware acceleration methods in order to be suitable for clinical applications VIII REFERENCES
1 2 3 4 5 6 National Institute for Standards and Technology PSTAR database http www physics nist gov PhysRefData Star Text PSTAR html M H Phillips K J Stelzer T W Griffin M R Mayberg and H R Winn Stereotactic radiosurgery a review and comparison of methods J Clin Oncol vol 12 5 pp 1085 1099 1994 M A Weinstein P M Duchesneau and W J MacIntyre White and gray matter of the brain differentiated by computed tomography Radiology vol 122 3 pp 699 702 1977 Z Li Novel Silicon Stripixel Detector Concept Simulation and Design BNL Tech Rep 67527 June 2000 Y Takada K Kondo T Marume K Nagayoshi I Okada and K Takikawa Proton computed tomography with a 250 MeV pulsed beam NIM A vol 273 1 pp 410 422 1988 P Pemler J Besserer J de Boer M Dellerta C Gahna M Moosburger et al A detector system for proton radiography on the gantry of the Paul Scherrer Institute NIM A vol 432 2 3 pp 483 495 1999 A Wagner W P Tan K Chalut R J Charity B Davin Y Larochelle et al Energy resolution and energy light response of CsJ Tl scintillators for charged particle detection NIM A vol 456 pp 290299 2001 M Moszynski J Zalipska M Balcerzyk M Kapusta W Mengesha and J D Valentine Intrinsic energy resolution of NaI Tl NIM A vol 484 pp 259 269 2002 W G Gong Y D Kim G Poggi Z Chen C K Gelbke Lynch WG et al Resolution tests of CsI Tl scintillators read out by pin diodes NIM A vol 268 pp 190 199 1988 J R M Annand J C McGeorge G J Miller V Holliday and L Isaksson A large solid angle CsI Tl Proton spectrometer NIM A vol 292 pp 129 134 1990 H Bichsel Passage of charged particles through matter American Institute of Physics handbook vol 42 B H Billings D E Gray et al Eds New York McGraw Hill 1972 pp 8 189 T Li Z Liang K Mueller J Heimann L Johnson H Sadrozinski et al Reconstruction for proton computed tomography a Monte Carlo study presented at the 2003 IEEE Nuclear Science Symposium and Medical Imaging Conference M10 343 L Shepp and Y Vardi Maximum likelihood reconstruction for emission tomography IEEE Trans Med Imaging vol 1 pp 113 122 1982 K Lange and R Carson EM reconstruction algorithms for emission and transmission tomography J Comput Assist Tomog vol 8 pp 306 316 1984 Z Liang and H Hart Bayesian image processing of data from constrained source distributions I non valued uncorrelated and correlated constraints Bull Math Biology vol 49 pp 51 74 1987 E Levitan E and G Herman A MAP EM algorithm for image reconstruction in emission tomography IEEE Trans Med Imaging vol 6 pp 185 192 1987 H Hudson and R Larkin Accelerated image reconstruction using Ordered Subsets of Projection Data IEEE Trans Med Imaging vol 13 601 609 1994 F Xu and K Mueller A unified framework for rapid 3D Computed Tomography on Commodity GPUs presented at the 2003 IEEE Nuclear Science Symposium and Medical Imaging Conference M10 331 K Mueller and R Yagel Rapid 3D cone beam reconstruction with the Algebraic Reconstruction Technique ART by using texture mapping hardware IEEE Trans Med Imaging vol 2 pp 1227 1237 2000

7

8 9 10 11 12

13 14 15 16 17 18 19