Reconstruction For Proton Computed Tomography - A Practical Approach

Reconstruction For Proton Computed Tomography - A Practical Approach - Transcript

Reconstruction for Proton Computed Tomography A Practical Approach
K Mueller Z Liang T Li F Xu J Heimann L Johnson H Sadrozinski A Seiden D Williams L Zhang S Peggs T Satogata V Bashkirov and R Schulte
imaged frequently or in sensitive tissue areas Proton CT is also the most appropriate imaging method to perform planning and verification of proton based radiation treatment since proton stopping power maps constructed by table based transformation of X ray CT images only render approximate stopping power estimates limiting the power of proton radiation therapy This paper discusses the fundamental aspects of proton CT starting with the current state of affairs and then proceeding to an overview of our new proposed approach A companion paper also submitted to this conference will give more detailed renditions on a preliminary study using a Monte Carlo simulated dataset 3 II BACKGROUND An increasing number of cancer patients is treated with proton radiation worldwide Proton radiation therapy is one of the most precise forms of image guided cancer therapy since the sharp dose peak called the Bragg peak produced by protons at the end of their range can be placed anywhere inside the patient s body and can deliver a destructive dose to the target while completely sparing tissues behind the target An important aspect of proton therapy is the accuracy of the range determination of protons in tissues When sensitive structures are in close proximity to the tumor the most appealing approach of stopping the proton beam immediately in front of the critical structure is often avoided by the radiation oncologist due to proton range uncertainties and the fear of side effects What is needed is an accurate map of the proton stopping power in the traversed tissue such that the Bragg peak can be delivered precisely to the tumor site Proton stopping power is linearly related to the electron density of tissue while X ray CT is dependent on attenuation coefficients of the tissues for a spectrum of photons Due to these fundamental physical differences the accuracy of mapping X ray Hounsfield numbers to proton stopping powers cannot be better than a factor of 3 and range uncertainties can exceed 10 mm in a human head 12 Cormack was probably the first to point to the possibility to perform proton CT in 1963 3 In early studies proton radiographs were taken in a fashion very similar to x ray radiographs a passive photographic plate behind the sample recorded projected images of an incident beam according to the transmission efficiency of the protons 8 The proton beam energy had to be tuned to center the Bragg peak on the photographic plate in order to get useful density contrast With

Abstract Proton Computed Tomography CT has important implications for both image guided diagnosis and radiation therapy For diagnosis the fact that the patient dose committed by proton CT compares favorably with that delivered by traditional X ray CT for the same density resolution and contrast may be exploited in dose critical clinical settings Proton CT is also the most appropriate imaging method to perform planning and verification of proton based radiation treatment since proton stopping power maps constructed by table based transformation of X ray CT images only render approximate stopping power estimates In proton CT sharp features become blurred by the phenomenon of multiple Coulomb scattering MCS resulting in a resolution of around 3 to 5 mm Studies showed that the spatial resolution of proton radiography and CT can be improved to about 1 2 mm by tracking individual protons in coincidence as they enter and exit the imaged object This paper describes a new practical implementation of this approach We first bin the captured protons into slots of similar tracks Optionally proton energy statistics can be collected within each bin to obtain further parameters for tissue characterization The envelope of path uncertainty due to MCS can be be modeled as a banana shaped curve The 3D reconstruction using either filter backprojection or iterative algorithms can be performed rapidly on graphics hardware using a slice blurring technique to model the MCS uncertainty curve

ROTON Computed Tomography CT has important implications for both image guided diagnosis and radiation therapy For diagnosis the fact that the patient dose committed by proton CT compares favorably with that delivered by traditional X ray CT for the same density resolution and contrast may be exploited in a number of clinical settings e g in cases in which a patient needs to be
Manuscript received Oct 29 2003 Asterisk indicates corresponding author K Mueller Z Liang T Li and F Xu are with the Department of Radiology and Computer Science Stony Brook University Stony Brook NY 11794 USA email mueller cs sunysb edu J Heimann L Johnson H Sadrozinski A Seiden D Williams and L Zhang are with the Santa Cruz Institue for Particle Physics UC Santa Cruz S Peggs and T Satogata are with the Collider Accelerator Department Brookhaven National Lab V Bashkirov and R Schulte are with the Department of Radiation Medicine Loma Linda University Medical Center

P

I

INTRODUCTION

this method sharp features become blurred by the phenomenon of multiple Coulomb scattering MCS resulting in a resolution of around 3 to 5 mm Later studies showed that the spatial resolution of proton radiography and CT can be improved to about 1 2 mm by tracking individual protons in coincidence as they enter and exit the imaged object 6 13 Our implementation uses this most accurate approach III A PRACTICAL RECONSTRUCTION APPROACH We shall assume that the incident angle a pair in 3D and the energy of the individual protons is available at good accuracy see Fig 1 A description of the equipment with which this can be achieved is subject of another paper 14 With regards to the 3D reconstruction there are a number of challenges that need to be addressed the large angular distribution of the particles must be accommodated the scattering envelope of the particles must be modeled the massive amount of projection data must be managed efficiently the reconstruction must be accomplished within a reasonable amount of time i e on the order of minutes The last item on this list determines the type of approach we feel is required to faciliate a practical implementation We accomplish the first and third task on the list by sorting the data into bins s where s is the location in the sinogram and are the discretized horizontal and vertical angles of ray inclination with respect to the flat detector plane The energy distribution within each bin can be used to compute probalistic measures for characterizing the object tissues encountered along the particle paths We shall use the average energy for now We then store the image data sorted by

incomplete projection data It may be the case that some angles or bins are not selected by particles during an imaging session which will give rise to undesirable reconstruction artifacts caused by the initial filtering stage In these cases we may resort to slower iterative methods that generally perform better in these scenarios With respect to the significant Coulomb scattering MLE is also an appropriate choice to maximize the fit of the reconstruction to the data We suggest to use Ordered Subsets EM OS EM 7 for faster convergence Clinical utility requires fast reconstruction within minutes This is best afforded by performing the reconstruction exploiting commodity graphics hardware boards GPUs which are currently evolving in their capabilities A reconstruction is performed by using the texture mapping facilities of the hardware for performing the interpolations required in the projection and backprojection operations 2 10 The other arithmetic operations that occur in the reconstruction algorithms can also be performed in this hardware Finally the probablity envelope can be modeled by performing a slice blurring before the projection summing or volume update during projection and backprojection respectively We determine the amount of blurring as a function of the width of the envelope at the slice position Using our hardware reconstructor we were able to perform a cone beam reconstruction 16 cone angle of a 1283 volume from 80 projections in 3 SART iterations in 50 seconds which is clearly in the clinical range Fig 2 shows a slice of a reconstructed CT head b and the original a

a

b

Figure 2 CT head dataset a original b reconstructed from 80 high quality simulated projections acquired from a

Figure 1 Two sample envelopes of possible proton trajectories due to multiple Coulomb scattering t For reconstruction we may either use a Filtered Backprojection method such as the one by Feldkamp et al 4 or an iterative method such as MLE 11 or SART 1 5 The former is generally faster but does not deal well with

IV CONCLUSIONS The approach we have outlined here forms the general data processing pipeline for a proton CT application past the projection data pre processing stage As we experiment more with both Monte Carlo simulated as well as real data we hope to come up with effective and accurate models for the probability envelope as a function of We also hope to

incorporate some simple Monte Carlo simulation capabilities directly into our hardware reconstruction platform in order to better model the Coulomb scattering processes during the 3D reconstruction V DISCLAIMER This work was supported by Calspace and the National Medical Technology Testbed Inc NMTB under the U S Department of the Army Medical Research Acquisition Activity Cooperative Agreement DAMD17 972 7016 and by the U S Department of Energy contract No DE AC0298ch10886 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 VI REFERENCES
1 A H Andersen and A C Kak Simultaneous Algebraic Reconstruction Technique SART a superior implementation of the ART algorithm Ultrason Img vol 6 pp 81 94 1984 2 B Cabral N Cam and J Foran Accelerated volume rendering and tomographic reconstruction using texture mapping hardware 1994 Symp Volume Vis pp 91 98 1994 3 Cormack Representation of a function by its line integrals with some radiobiological applications J Applied Physics vol 34 pp 2722 27 1963 4 L A Feldkamp L C Davis and J W Kress Practical cone beam algorithm J Opt Soc Am pp 612 619 1984 5 R Gordon R Bender and G T Herman Algebraic reconstruction techniques ART for three dimensional electron microscopy and X ray photography J Theoretical Biology vol 29 pp 471 481 1970 6 K Hanson J Bradbury T Cannon R Hutson D Laubacher R Macek M Paciotti C Taylor Computed tomography using proton energy loss Physics in Medicine and Biology vol 26 no 6 pp 965 83 Nov 1981 7 H Hudson R Larkin Accelerated Image Reconstruction Using Ordered Subsets of Projection Data IEEE Trans Medical Imaging vol 13 pp 601 609 1994 8 A Koehler Proton beam radiography in tumor detection Science vol 179 pp 913 14 1973 9 T Li Z Liang K Mueller et al Reconstruction for Proton Computed Tomography A Monte Carlo study submitted IEEE Medical Imaging Conference 2003 10 K Mueller and R Yagel Rapid 3D cone beam reconstruction with the Algebraic Reconstruction Technique ART by using texture mapping hardware vol 19 no 12 pp 1227 1237 IEEE Transactions on Medical Imaging 2000 11 L Shepp Y Vardi Maximum likelihood reconstruction for emission tomography IEEE Trans on Medical Imaging vol 1 pp 113 122 1982 12 Schneider U Pedroni E Proton radiography as a tool for quality control in proton therapy Medical Physics vol 22 pp 353 63 1995 13 U Schneider E Pedroni Multiple Coulomb scattering and spatial resolution in proton radiography Medical Physics vol 21 no 11 pp 1657 63 Nov 1994 14 R Schulte V Bashkirov et al Design of a Proton Tomography system for applications in proton radiation therapy submitted Instrumentation for Radiation Medicine Workshop 2003