Experimental feasibility of multi-energy photon-counting K
Transcription
Experimental feasibility of multi-energy photon-counting K
IOP PUBLISHING PHYSICS IN MEDICINE AND BIOLOGY Phys. Med. Biol. 53 (2008) 4031–4047 doi:10.1088/0031-9155/53/15/002 Experimental feasibility of multi-energy photon-counting K-edge imaging in pre-clinical computed tomography J P Schlomka1, E Roessl1, R Dorscheid2, S Dill2, G Martens1, T Istel1, C Bäumer3, C Herrmann3, R Steadman3, G Zeitler3, A Livne4 and R Proksa1 1 2 3 4 Philips Research Europe, Sector Medical Imaging Systems, Hamburg, Germany Philips Research Europe, Engineering & Technology, Aachen, Germany Philips Research Europe, Sector Medical Imaging Systems, Aachen, Germany Philips Healthcare, Global Research and Advanced Development, Haifa, Israel E-mail: jens-peter.schlomka@philips.com Received 11 March 2008, in final form 10 June 2008 Published 8 July 2008 Online at stacks.iop.org/PMB/53/4031 Abstract Theoretical considerations predicted the feasibility of K-edge x-ray computed tomography (CT) imaging using energy discriminating detectors with more than two energy bins. This technique enables material-specific imaging in CT, which in combination with high-Z element based contrast agents, opens up possibilities for new medical applications. In this paper, we present a CT system with energy detection capabilities, which was used to demonstrate the feasibility of quantitative K-edge CT imaging experimentally. A phantom was imaged containing PMMA, calcium-hydroxyapatite, water and two contrast agents based on iodine and gadolinium, respectively. Separate images of the attenuation by photoelectric absorption and Compton scattering were reconstructed from energy-resolved projection data using maximumlikelihood basis-component decomposition. The data analysis further enabled the display of images of the individual contrast agents and their concentrations, separated from the anatomical background. Measured concentrations of iodine and gadolinium were in good agreement with the actual concentrations. Prior to the tomographic measurements, the detector response functions for monochromatic illumination using synchrotron radiation were determined in the energy range 25 keV–60 keV. These data were used to calibrate the detector and derive a phenomenological model for the detector response and the energy bin sensitivities. 0031-9155/08/154031+17$30.00 © 2008 Institute of Physics and Engineering in Medicine Printed in the UK 4031 4032 J P Schlomka et al 1. Introduction The attenuation properties of matter for x-rays are energy- and material dependent. This allows, in return, deriving material properties by using a multitude of different x-ray energies for imaging. A simple realization of this technique is known as ‘dual-energy imaging’ (Alvarez and Macovski 1976, Kelcz et al 1979, Lehmann and Alvarez 1986). Using different incident spectra (Lehmann et al 1981) or energy discrimination on the detector side (‘dual crystal method’) (Barnes et al 1985, Carmi et al 2005), information about the average Z-number of the material under investigation can be derived. In medical imaging this method can be used to, e.g., suppress beam-hardening effects or to discriminate to some extent between calcium or iodine and lighter elements (Z < 20), which form body tissues. When heavier elements are used in contrast agents such as, e.g., gadolinium, gold or bismuth their K-absorption edges lying at 50.2 keV, 80.7 keV and 90.5 keV, respectively, become accessible for diagnostic x-ray imaging. By using more than two spectrally distinct measurements (tube spectra, filtrations, energy bins, etc), these K-edge features can be discriminated from the other contributions to the x-ray attenuation. To distinguish this acquisition method from ‘dual energy CT’ it is referred to as ‘spectral CT’ in this paper. Spectral CT in combination with a data analysis method, which allows extracting, from a single scan, separate images of the contrast agent on the one hand and of the anatomy on the other hand we refer to as ‘K-edge imaging’ (Lehmann and Alvarez 1986, Sukovic and Clinthorne 1999). K-edge imaging can be realized by using multi-bin energy-discriminating counting detectors. New developments in direct-conversion semiconductor detectors based on CZT (Cadmium–Zinc–Telluride) or CdTe (Cadmium Telluride) may enable the application of this technique also for medical CT imaging (Takahashi et al 2005, Kowase and Ogawa 2006). In a previous paper (Roessl and Proksa 2007) the theory of projection-based material decomposition from energy-binned photon-counting data was presented. The processing described there is based on an extension of the dual-energy method first introduced by Alvarez and Macovski (1976) to a higher number of basis components. Roessl and Proksa showed simulated images, which clearly allowed the differentiation between various basis components including images showing only the contrast agent. In that work the contrast agent was based on gadolinium, but also other high-Z materials seemed feasible. It was already mentioned that it might be possible to use even more than 1 contrast agent simultaneously, provided sufficient energy discrimination on the detectors side. In the present paper we report on the experimental proof of this technique. For this task, a research spectral CT scanner based on a rotating gantry was built, which enables convenient multi-slice scanning. In the measurements presented here we used a technical phantom containing several materials including two different conventional medical contrast agents. Thus, a four-basis component decomposition was required. Separate images of the anatomy and the contrast agents are presented. A prerequisite for the spectral analysis is the knowledge of the spectral response of the detector. We measured this function using monochromatic synchrotron radiation at several different primary energies and derived a phenomenological description of the complete spectral response behavior of the detection system. This paper is structured as follows: in section 2 we introduce the theoretical background of spectral CT, i.e., the physical models of the attenuation and the detection of the x-rays as well as the maximum-likelihood processing method performed prior to image reconstruction. Section 3 introduces the experimental spectral CT system used in the studies and briefly describes the synchrotron set-up used for detector characterization. In section 4 we present experimental results focusing on first spectral CT phantom images. A discussion and outlook Experimental feasibility of multi-energy photon-counting K-edge imaging in pre-clinical CT 4033 section completes this paper. The detector energy scale calibration and the response function model are briefly explained in the appendix. 2. Theory The theoretical formalism of K-edge imaging using energy-binned detection and maximumlikelihood processing of the projection data has been introduced in an earlier paper (Roessl and Proksa 2007) and will therefore only briefly be reproduced here for completeness. 2.1. X-ray attenuation and detection The energy dependence of the linear attenuation coefficient µ(E, x ) of materials in the diagnostic energy range ∼15–150 keV can be approximated fairly well by a linear combination of the photoelectric and the Compton cross sections fph (E) and fKN (E), see Alvarez and Macovski (1976) and references therein. If no absorption edges are located within the used energy range, the photoelectric cross section can be approximated by its E −3 energy dependence, and we obtain 1 + aCo ( x )fKN (E). (1) E3 The vector x describes the space dependence of the attenuation and the energy dependence of the total Compton cross section fKN (E) was first derived by Klein and Nishina (1929) and can be found in, e.g., Knoll (2000). In the presence of elements with high atomic number Z, the above description of the attenuation properties of matter has to be modified to correctly describe the attenuation of the sample. In the experimental set-up presented below, we used the elements iodine and gadolinium. Both elements are widely applied in medical contrast agents. For spectral CT iodine has the disadvantage that its K-edge is at the rather low energy of 33.2 keV and can therefore hardly be discriminated from other tissue in thicker objects such as, e.g., human patients by means of its K-edge only. On the other hand, for pre-clinical applications iodine is very well suited due to the much lower tube voltages and less severe photon starvation below the iodine K-edge. Gadolinium is mainly known as an MRI contrast agent due to its magnetic properties but is also used in some medical x-ray imaging applications (RemyJardin et al 2006). It should be emphasized that the two contrast materials were mainly taken because of their easy accessibility. K-edge imaging in the medical case is not limited to just these materials. Provided toxicity issues can be resolved, all materials with K-edge energies above ∼40 keV, e.g., gold or bismuth, could potentially be used in human applications. For the sake of clarity, however, we will use the indices ‘Gd’ and ‘I’ for the two K-edge materials, noting that they are potential placeholders for other future K-edge materials. To incorporate the attenuation of iodine and gadolinium, equation (1) now has to be extended by the energy-dependent attenuation functions of the two elements gadolinium and iodine (Lehmann and Alvarez 1986, Sukovic and Clinthorne 1999, Roessl and Proksa 2007): 1 x ) 3 + aCo ( x )fKN (E) + aGd ( x )fGd (E) + aI ( x )fI (E) µ(E, x ) = aph ( E 4 ≡ aα ( x )fα (E). (2) µ(E, x ) = aph ( x) α=1 aGd ( x ), fGd (E), aI ( x ), fI (E) denote the local density and the mass attenuation coefficient of gadolinium and iodine, respectively. In the second line of equation (2) the values α = 1, 2, 3, 4 4034 J P Schlomka et al represent the photoelectric effect, the Compton effect and the gadolinium and iodine-based contrast-agent components. The mass attenuation coefficients of the two K-edge elements account for all attenuation effects including the Compton effect and the photo-effect with the K-edge discontinuity of the photo cross section. While in dual-energy systems (at least) two spectrally different measurements per x-ray path have to be performed in order to determine x ) and aCo ( x ), the extended model equation (2) relies the line integrals of the coefficients aph ( on a minimum of four measurements. Our x-ray detector was operated in photon-counting mode applying energy discrimination by six threshold values for each individual pixel k = 1 . . . 1024. Physically, the thresholds are realized by voltages u(i,k) T , i = 1, . . . , 6, which are fed into the pulse-height comparator circuits. The pulse height obtained from the detector is nearly proportional to the energy of the detected photon, but gain and offset vary from bin to bin and from pixel to pixel. Therefore, the relationship between the measured photon energy UT(i,k) (in keV) and the threshold voltage u(i,k) T (in mV), which triggers the corresponding event, had to be determined through a calibration process described in the appendix. Upon successful completion of this calibration process of all individual pixels corresponding to a given value of i had to the threshold voltages u(i,k) T be set to different but well-defined values to ensure the realization of homogeneous threshold settings in the measured energy UT(i,k) despite the variations in the gain and offset of the detector electronics. Thus the index k, representing the pixel dependence in the measured energy can be omitted. In practice, slight variations of the measured energy remain, but were not taken into account in the present processing scheme. A photon with energy Eph was counted in one of the six adjacent bins Bi , if its measured energy equivalent Uph fulfilled UT(i) Uph < UT(i+1) . The uppermost bin B6 had no upper threshold, i.e. UT(7) = ∞. A discrepancy between the actual photon energy Ep and the measured energy Up arises from the detection process. Physical effects such as electronic noise, incomplete charge collection, energy loss due to K-escape, or other sources contribute to a degradation of the energy resolution of the detector. As a consequence, two adjacent energy bins are not sharply separated by the energy threshold at their common boundary. Their sensitivities are described by so-called bin sensitivity functions Si (E) and the effects contributing to the degradation of the energy resolution lead to an overlap of the bin-sensitivity functions. These functions were determined experimentally as described in section 4 and the appendix. Having described the attenuation and the detection process we can now introduce a statistical description of the counting process via the expectation value λi of the number of x ): photons in the energy bin Bi and the line integrals Aα of the coefficients aα ( ∞ 4 Si (E)(E) exp − fα (E)Aα D(E) dE, i = 1, . . . , 6, (3) λi (A1 , A2 , A3 , A4 ) = 0 with α=1 Aα = aα ( x ) ds, α = 1, . . . , 4. (4) In equation (3), (E) denotes the spectral x-ray photon fluence, and D(E) the detector absorption efficiency. 2.2. Maximum-likelihood processing of projection data For our application, the number of energy bins exceeds the dimension of the attenuation basis. The system in equation (3) is therefore overdetermined, and, in general, no solution exists for A1, A2, A3 and A4. To overcome this difficulty, the improved spectral sampling of the Experimental feasibility of multi-energy photon-counting K-edge imaging in pre-clinical CT 4035 attenuation process can be combined with adequate methods of parameter estimation, such as the least-squares method or the method of maximum likelihood. In our processing method we used the maximum-likelihood approach, because the corresponding estimators turned out to be efficient and unbiased in many practical situations (Cowan 1998, Eadie et al 1971). Mi (Aα ), i = 1, . . . , N (with N = 6 in our case) denote the random variables describing the number of photons detected in the energy bin Bi and λi (Aα ) denote their mean values. The random variables and their corresponding mean values depend on the line integrals Aα . Assuming that the Mi form a set of independent Poisson random variables, we can calculate the likelihood function as the probability of an arbitrary measurement result (M1 = m1 , . . . , MN = mN ), given a particular composition of the object parameterized by Aα : P (m1 , . . . , mN |λ1 (Aα ), . . . , λN (Aα )) = N λi (Aα )mi i=1 mi ! e−λi (Aα ) . (5) Instead of directly maximizing the likelihood function P , it is usually more convenient to minimize the negative log-likelihood L which, re-expressed as a function of the parameters Aα , becomes L(m1 , . . . , mN |Aα ) = −ln[P (m1 , . . . , mN |λ1 (Aα ), . . . , λN (Aα ))] N = [λi (Aα ) + ln mi ! − mi ln λi (Aα )] i=1 ∼ = N [λi (Aα ) − mi ln λi (Aα )]. (6) i=1 In the last step, we dropped terms independent of Aα and, therefore, irrelevant for the minimization. By minimizing the above expression equation (6) given the measurement results (m1 , . . . , mN ) with respect to the four parameters Aα , we obtain the maximum-likelihood estimates AML α . The optimization problem can only be treated numerically. A large variety of powerful numerical algorithms exist, which are specialized to the task of likelihood maximization. We used the simplex method of Nelder and Mead for our calculations (Nelder and Mead 1965, Press et al 1992). For the processing of spectral projection data we determined the line integrals Aα as described above for all measured paths through the object under study. Once the maximumlikelihood processing for all paths was complete, four conventional reconstruction steps using standard filtered-back projection techniques generated four basis images (Kek and Slaney 2001). 3. Experimental set-up 3.1. CT set-up The tomographic measurements were carried out using an experimental spectral CT scanner based on a direct-drive rotating gantry (figure 1). On the drive motor a rotating disc was mounted and the components were mounted onto this disc. As an x-ray source a microfocus tube (KEVEX PXS10-65 W) operating at voltages between 70 kVp and 130 kVp was used. Energy spectra were measured prior to installation on the gantry using a germanium detector with energy resolution of a few hundred eV. The focal spot size depends on the applied 4036 J P Schlomka et al Figure 1. Schematic drawing and photograph of the pre-clinical spectral CT system used for this study. The key components on a rotating gantry are a micro-focus x-ray tube (top) and a single-line energy-binning photon-counting detector (bottom). electrical power starting at 10 µm for low power increasing to 100 µm at a maximum power of 65 W. The measurements presented below were taken at a 100 kVp beam voltage and a 100 µA beam current, which resulted in a beam spot of ∼12 µm (according to the manufacturer) and a measured x-ray flux of about 3 × 105 cts pixel−1 s−1. The x-ray tube has a beryllium exit window. No additional energy-spectrum filter or bow-tie filter was used. Slits at the tube exit collimate the x-rays into a narrow fan beam. As detector, a single-line photon-counting CdTe array manufactured by GammaMedicaIdeas (Northridge, Ca., USA) was employed (Tajima et al 2004). The x-rays enter the detector through a 2 mm aluminum housing. Concerning the x-ray dose, it would be more appropriate to reduce the detector window thickness and instead use pre-object filtering. This is planned in a future system upgrade. The detector is equipped with 1024 pixels with 0.4 mm average pixel pitch. The active area of each pixel is 0.38 mm × 1.6 mm with a crystal thickness of 3 mm. A lead slit placed in front of the detector crystals collimates the illuminated detector width to 1.2 mm. For each pixel, six independent comparator threshold levels can be set via software. One 16 bit counter is provided for each of the comparators. Read-out of all counters takes 25 µs during which no data can be acquired. The system geometry is adjustable providing the capability for high-resolution measurements with a small field of view and medium resolution measurements with a larger field of view. For this purpose, the tube and the detector are mounted on slides and can be repositioned easily. The geometry parameters for two settings are given in table 1. For the measurement below, the medium resolution set-up was applied. The maximum rotation speed of the system is limited to 0.33 revolutions s−1 to avoid excessive centrifugal forces on the mounted components. For the measurements presented here, however, the gantry rotated much slower to enable good photon statistics while maintaining a rather low photon flux onto the detector. Angular positions are measured by an optical encoder, which directly triggers the PC-based data acquisition on the Experimental feasibility of multi-energy photon-counting K-edge imaging in pre-clinical CT 4037 Table 1. Geometry parameters for two configurations of the spectral CT experimental scanner. To switch from one configuration to the other, source and detector can be translated independently on slides. Spatial resolution was estimated by measuring the full width at half maximum of a 50 µm thick tungsten wire in reconstructed images. High resolution set-up Distance source–isocenter 100 mm Distance isocenter–detector center 500 mm Magnification 6 Field of view 60 mm Spatial resolution ∼100 µm Large field-of-view set-up 300 mm 300 mm 2 200 mm ∼250 µm gantry. A master PC controls the data acquisition PC remotely via an ethernet link through a slip ring. 3.2. Synchrotron set-up Detector calibration measurements were carried out at the DESY/Hasylab facility (Hamburg, Germany) at beam-line X1 (Hasylab 2007, Baumer et al 2008). The beam line is positioned at a bending magnet of the DORIS ring, equipped with a Si (5 1 1) double crystal monochromator and provides monochromatic x-ray photons in the energy range from 25 keV to 80 keV with an energy bandwidth lower than 100 eV. The cross section of the beam was 6 mm × 0.25 mm illuminating about 15 adjacent pixels at a time. Due to this small footprint of the synchrotron beam, it was required to translate the detector perpendicular to the beam to cover the full active area of the detector. Measurements were carried out with photon fluxes between 105 and 106 photons pixel−1 s−1. 4. Results 4.1. Detector calibration and characterization at the synchrotron The spectral response was derived from so-called threshold scans. For such a scan, the detector was illuminated under constant flux conditions with monochromatic photons of energy E. The count rate c for a given pixel k = 1 . . . 1024 and comparator i = 1, . . . , 6 was recorded while varying the comparator threshold voltages u(k,i) T . From the derivative of this function (k,i) (k,i) d (k,i) c u the spectral response R E, U was obtained after making use of the (k,i) T T duT and UT(k,i) through the energy scale calibration (see appendix). relationship between u(k,i) T The aim of the calibration procedure was to set the thresholds UT(k,i) of all detector pixels k = 1 . . . 1024 to one common value UT(i) by using individual values u(k,i) T for each comparator. Since we assumed equal spectral response of each pixel, the index kcould also be omitted for the spectral response R (i) E, UT(i) . The spectral response of a bin i, R (i) E, UT(i) , describes the probability of a photon of energy E to be counted at a measured energy UT(i) . From the response function the bin sensitivity functions S (i) (E) was calculated for given thresholds UT(i) : S (i) (E) = UT(i+1) UT(i) R (i) (E, U ) dU i = 1, . . . , 6 UT(7) = ∞. (7) 4038 J P Schlomka et al 0.12 25 keV 30 keV 35 keV 40 keV 45 keV 50 keV 55 keV 60 keV Normalized count rate 0.1 0.08 0.06 0.04 0.02 0 15 20 25 30 35 40 45 50 55 60 65 70 Measured photon energy/keV Figure 2. Spectral response of one threshold in one detector pixel resulting from monochromatic illumination with photon energies ranging from 25 keV to 60 keV. The count rate in the detector was about 105 photons pixel−1 s−1. The data are normalized to an integrated count rate of 1 cts s−1 above 20 keV. The bin sensitivity function gives the probability of a photon of energy E to be measured in bin i. Figure 2 shows exemplary, but representative experimental results for one detector pixel and one energy threshold measured at a count rate of 105 photons pixel−1 s−1. We used eight distinct incident energies ranging from 25 keV to 60 keV. The maximum response occurring at the incident energy—the photo-peak—could be well described by a Gaussian with increasing width at higher energies. This broadening could not be explained by the statistics of the electron–hole pair generation within the crystal. Incomplete charge collection contributed dominantly. In addition to the photo-peak, some spectra exhibited a K-escape peak (Sato et al 2005). This peak appeared about 25 keV below the photo peak. An incident photon contributes to the K-escape peak if immediately after its photoelectric absorption a K-photon is emitted from the host atom and leaves the active pixel volume. This results in a reduction of the deposited energy by the energy of the K-escape photon. For Cd and Te, K-fluorescence energies lie between 22.7 keV and 31.7 keV, in good agreement with the observed average of 25 keV. Emission of the fluorescence photon from the excited atom and escape of this photon from the active detector pixel volume are prerequisites for the reduction in deposited energy measured in the pixel. The latter process is energy dependent and therefore influences the intensity of the escape peak. For higher energies, lower photo absorption cross sections and the deeper average penetration of the primary photons will lower the K-escape probability and consequently the escape peak intensity. Finally, besides the photo- and the K-escape peak, a rather constant background in the response function was found between the noise band and the photo-peak. We attribute this to charge sharing events, i.e., the charge generated by a single incident photon is measured in two neighboring pixels with reduced energy in each pixel. For energies above the incoming photon energy hardly any intensity was measured, which showed that for a rate of 105 photons pixel−1 s−1 pile-up did not play a significant role (Knoll 2000). This behavior changed when the flux was increased to 106 photons pixel−1 s−1 leading to significant pile-up. To avoid the degradation of the spectral response in this feasibility study by pile-up, we measured with Experimental feasibility of multi-energy photon-counting K-edge imaging in pre-clinical CT 4039 10000 Measurement 0.22mm Sn foil @ 70 kV 9000 Measurement 0.22mm Sn foil @ 130 kV 8000 number of photons Measurement 1.2mm Pb @ 130 kV 7000 simulations 6000 noise measured with 30mm Pb @ 130 kV 5000 4000 3000 2000 1000 0 10 30 50 70 90 110 130 Measured photon energy/keV Figure 3. Measured and calculated transmission of 70 kVp and 130 kVp x-ray tube spectra through 0.22 mm of tin and 1.2 mm of lead, respectively. Measurements were performed through threshold scans. For the calculations the response function described by equations (A.2) and (A.3) was used. the low flux of 105 photons pixel−1 s−1. Count rate issues will be treated in more detail in section 5. The observation of the shape of the spectral response motivated the phenomenological model described in the appendix. Since the shapes of the spectral responses of the individual pixels and bins did not vary strongly, we used only one function for the description of all pixels. To verify the applicability of the modeled response, we measured the transmission of two known polychromatic spectra (acceleration voltages of 70 kV and 130 kV) through tin and lead. The two materials exhibit their K absorption edge within the x-ray range considered here. Thus the transmission spectra show very distinct features. These spectra were measured by threshold scans and obtained in parallel by calculations using the above spectral response model. Figure 3 shows the good agreement of measurement and calculation. 4.2. Tomographic measurements For the feasibility study of K-edge spectral CT a phantom was used that exhibited certain features of clinical relevance. As described in section 1, one major application is the distinction of the contrast agent from calcifications in a single CT scan. Furthermore, applications using more than 1 type of contrast agent may also become relevant, e.g., in the simultaneous, selective enhancement of different regions such as the vascular system on the one hand using a blood-pool agent and soft plaque on the other hand using an agent targeting, e.g., fibrin. We therefore designed the PMMA phantom shown in figure 4. It hosted 12 contrast agent filled tubes surrounded by calcium-hydroxyapatite ([Ca5-(PO4)3(OH)], bone mineral) in a gel matrix. Furthermore, a central reservoir was filled with water. The detailed layout is given in the figure. As contrast agents two conventional medical contrast agents were tested, one based on iodine, PeritrastTM5 , and one based on gadolinium: MagnevistTM6 . The concentrations of the heavy metals in the inserts are given in table 2. 5 6 PeritrastTM , Köhler Chemie, Alsbach-Hähnlein, Germany. MagnevistTM , Schering, Berlin, Germany. 4040 J P Schlomka et al PMMA Gd-3 Gd-2 I-1 water I-2 I-3 Gd-1 Figure 4. Schematic drawing of the ‘coronary phantom’ used for the tomographic feasibility test of K-edge imaging. The small holes were filled with two different contrast agents (gadolinium and iodine based) in three different concentrations each, surrounded by the bone mineral calciumhydroxyapatite in a gel matrix. The diameter of the phantom was 8 cm, the calcium-hydroxyapatite inserts had an outer diameter of 8 and 10 mm and an inner diameter of 5 mm. Concentrations are given in table 2. Table 2. Nominal and measured contrast agent concentrations of the inserts in the ‘coronary phantom’ schematically drawn in figure 4. Standard deviations were measured by analyzing regions of interest in the center of the inserts. Inserts 1 (low concentration) Inserts 2 (medium concentration) Inserts 3 (high concentration) Nominal Measured Nominal Measured Nominal Measured Iodine concentration (µmol ml−1) Gadolinium concentration (µmol ml−1) 37.5 32 ± 25 75 68 ± 25 150 132 ± 30 30 32 ± 25 60 70 ± 25 120 116 ± 25 Per rotation 450 projections were acquired with frame duration of 250 ms. This resulted in a total measurement time of 112.5 s per slice and a dose equivalent of 11.3 mAs. The energy thresholds were set at 25 keV, 30 keV, 40 keV, 50 keV, 60 keV, and 80 keV. As a first pre-processing step two-sided energy windows were calculated from the original data. This was done by subtracting the number of photons measured above the upper threshold from the number of photon counts measured above the lower threshold. This procedure can possibly be problematic for high count rates when baseline shifts can occur in the detector electronics, which was, however, not the case at the low rates used in this experiment. For the uppermost bin no subtraction was needed and the bin remained single sided. As a next step, we performed the normalization using an air-scan measured prior to the object measurement. Furthermore, gain correction was done on each pixel and bin using prior transmission data through a polycarbonate board of constant 4 cm thickness. The measurement was modeled and the quotient between measurement and simulation was used as a correction factor for the observed intensities. In the pre-processing algorithm this correction factor was weighted with the measured attenuation. The weight was zero for no attenuation and 1 when the attenuation matched the attenuation of the polycarbonate board. For other attenuation values data were inter- or extrapolated linearly. As a final step, and to further improve the reconstruction quality, Experimental feasibility of multi-energy photon-counting K-edge imaging in pre-clinical CT 4041 Figure 5. Reconstructed CT images of the six individual energy bins of the photon-counting detector. The lower energy thresholds of the six bins were 25, 30, 40, 50, 60 and 80 keV (From left to right, from top to bottom). A tube voltage of 100 kVp and a beam current of 100 µA were used. The rotation speed was 112.5 s per turn. All slices were normalized to Hounsfield units measured in the central inset. Level and window of the images are 150 HU/1700 HU. a deringing method following the algorithm described in Niederlöhner (2006) was applied in projection space. 4.2.1. Analysis of the energy-windowed data. In figure 5 the reconstruction results of the projection data of the five two-sided energy bins and the upper single-sided bin are shown. Already in these images the K-edge properties of the contrast agents—at least for the high concentrations—are visible. The iodine inserts exhibit strong attenuation in the bin 30 keV– 40 keV, whereas gadolinium increases strongly in attenuation when the energy exceeds 50 keV. Hydroxyapatite shows strong decrease in attenuation with increasing energies. This behavior is typical for medium-Z materials, where attenuation is dominated by photo electric absorption at lower energies. These materials are therefore best visible in the lower energy bins. The attenuation of the light materials water and PMMA is dominated by Compton scatter and therefore the dependence on energy is much weaker. They show slightly different energy dependence due to the small difference in their average Z number. 4.2.2. Spectral analysis. The results of the spectral decomposition applying the procedure described in section 2.2 are shown in figures 6 and 7. Four basis functions describing four attenuation processes were used for the procedure: photoelectric absorption, Compton scattering, attenuation by iodine and attenuation by gadolinium. The basis functions describing the attenuation by the heavy metals contain the total attenuation caused by these elements, originating from the photo effect including the K-shell contribution and the Compton scattering. Their photo-effect and Compton contributions therefore did not show up in the other, unspecific basis function images. The photo- and the Compton attenuation images of figure 6 were calculated at an x-ray energy of 60 keV from the first and the second term of 4042 J P Schlomka et al Figure 6. Reconstructed images after material decomposition. From left to right: photo-electric attenuation image (gray scale level/window (L/W): 3.75 × 10−3 mm−1/7.5 ×10−3 mm−1 at 60 keV photon energy), Compton scattering attenuation image (L/W 0.021 mm−1/0.042 mm−1 at 60 keV photon energy), iodine concentration image and gadolinium concentration image (L/W 50 µmol ml−1/150 µmol ml−1). Note the narrow line at the bottom of the iodine image. It can be attributed to a sticky tape containing barium, which has its K-edge only 4.3 keV higher than iodine. Figure 7. Left/center: Calculated monochromatic images at photon energies of 34 keV and 51 keV. Gray scale images range from µ = 0, . . . , 1.0 mm−1 and µ = 0, . . . , 0.8 mm−1, respectively. Right: color overlay image of the gadolinium concentration (green) and iodine concentration (red) on a monochromatic 51 keV image. (This figure is in colour only in the electronic version) equation (2), respectively, using the energy-independent values aph ( x ) and aCo ( x ) obtained through the spectral processing. As can be seen in the images, the decomposition into the four basis functions separated the individual materials. In the photo image the bone mineral is most prominent, the ‘soft tissues’ water and PMMA are well visible in the Compton image, and the two contrast agents are separately shown in the individual images almost without any anatomical background. Some ring artifacts remain in the image mainly due to small pixel-to-pixel variations of the threshold positions. Furthermore, during the processing all detector pixels were described by one and the same response function, which can at best be an approximation. Differences in the crystal quality, for example, resulted in different electronic properties of the material, e.g. variation in charge carrier lifetime and mobility, and thus differences in the spectral response. Additionally, ‘cross-talk’ between the two contrast agent images is visible. At the positions of the high iodine concentrations a small negative shift occurs in the gadolinium image, and at the positions of the high gadolinium concentrations a small positive contribution also occurs in the iodine image. Furthermore, all images show small offsets. In the photo- and Compton basis images, the offsets are negative. In the iodine and gadolinium images a positive offset outside the object changes into a negative offset inside the object. These artifacts have to be attributed to a combination of an inaccurate description of the spectral response by our Experimental feasibility of multi-energy photon-counting K-edge imaging in pre-clinical CT 4043 model function, incomplete knowledge of the primary x-ray tube spectrum as well as limited accuracy of the threshold positions. 4.2.3. Quantification. Besides the possibility to separate individual materials, K-edge imaging also offers the capability for quantification. Using the two right-hand side images of figure 6 the iodine and gadolinium concentrations were measured by averaging over regions of interest within the 12 contrast agent filled inserts. Since two inserts were always filled with the same concentration, the two values were averaged. The error bars were derived from the image noise and the constant offset was taken into account by subtracting an averaged concentration value measured near the center of the object. The quantification results are summarized in table 2 showing that very good agreement is found between the measured values and the nominal concentrations. For the given values of dose and spatial resolution, a contrast-to-noise ratio of about 1 was reached at a concentration of 25 µmol ml−1. To obtain more accurate concentration measurements or to measure lower concentrations, more dose equivalent than the 11.3 mAs used here had to be delivered. Furthermore, the system had not yet been optimized, e.g., the threshold positions could be tuned to the K-edges of the contrast materials. From the basis component images, monochromatic images for arbitrary photon energies could be obtained. The four images were combined in such a fashion that the summed image represents the attenuation at exactly the chosen energy, as if the image was recorded with monochromatic radiation. This is presented in the two images on the left in figure 7. Using this type of representation has the advantage that anti-correlated noise, as it can occur in the material separation procedure, is strongly reduced. Such artificial images are free of any beam-hardening effects and the energy can be chosen such that the contrast agent contribution to attenuation is enhanced. The two examples in figure 7 were chosen at energies right above the K absorption edges, such that the iodine and the gadolinium contributions were maximal in the respective image. Spectral CT sometimes is promoted as ‘Color CT’. In fact, the material-separated images can be combined into a single image with color coding and thus produce a color image. One such example is given in the right image of figure 7. The gadolinium and iodine concentrations are color coded in green and red, respectively, and drawn as overlay on top of the monochromatic 51 keV image. 5. Discussion and outlook In this paper we demonstrated that K-edge spectral CT is experimentally feasible, confirming simulations presented in a previous paper (Roessl and Proksa 2007). Images of the different contrast mechanisms were reconstructed individually and the concentrations of contrast agents were quantified. In the experiments it was shown that even two heavy metals could be distinguished and quantified independently from a single scan. The images do not have medical diagnostic quality since they suffer from some ring artifacts and offsets. The causes for these artifacts are, however, already mostly known. The offsets occurred because our model of the spectral response, the primary spectrum and the threshold positions of the individual pixels did not perfectly describe the experimental conditions. The ring artifacts mainly occurred due to the low-threshold positioning accuracy inherent to the current system, which could only be adjusted to about 1.3 keV standard deviation. Furthermore, variations in crystal quality slightly changed the spectral response from pixel to pixel (Sato et al 2005). Since in the current version of our processing, all pixels 4044 J P Schlomka et al are described by identical threshold positions and spectral response, pixel-to-pixel variations in the spectral decomposition do occur, which as a consequence lead to the observed ring artifacts. In future work, improved calibration and processing schemes will take these variations into account and will lead to a clear suppression of the observed artifacts. Immediate applications of K-edge imaging can be found in pre-clinical imaging, where spectral CT can become an important imaging technique to provide material-specific quantitative information in combination with high spatial resolution imaging. Naturally, one major pre-clinical research field for spectral CT will be contrast agent research, aiming at development of functional, disease-specific, targeted contrast agents. The experiments presented here were taken under conditions of very low x-ray flux accommodating the fact that state-of-the-art, high-rate photon-counting detectors cannot cope with count rates as high as those present in human CT. In a conventional medical CT scanner with a source–detector distance of about 1 m, high power x-ray tubes guarantee a fast data acquisition with photon fluxes of above 108 s−1 mm−2 (without the use of a bow-tie filter). For current-mode (energy-integrating) readout through a combination of a scintillator and a photodiode, as is common practice in conventional x-ray CT, such high photon fluxes pose no particular problem. With photon-counting detectors based on direct-conversion semiconductors such as CZT or CdTe, counting photons at such high rates constitutes a formidable task. Typical charge collection times in these materials are around 50 ns and thus the possible maximum count rates will lie somewhere around 5 × 106 s−1, taking into account pulse shaping and pulse height discrimination. In a recent paper (Iwanczyk et al 2007) a measured count rate of 6 × 106 s−1 was reported using a detector with 30 ns peaking time. Our measurements were carried out at much lower rates, since at high rates pile-up will occur and the spectral response will degrade. It is therefore an important task to develop correction schemes in the form of additional hardware components in the detector itself or software correction schemes, which will allow good spectral performance at these rates. Frey et al (2007a, 2007b) already investigated this issue and presented a pile-up correction scheme. Even when pile-up correction can be achieved, there still remains a discrepancy between a rate of ∼5 × 106 s−1 achievable in a pixel and the CT rate of above 108 s−1 per mm2. The most straightforward approach here is sub-structuring of the pixels’ anodes to reduce the count rate per pixel. However, when the pixel size reaches the size of the charge cloud produced by a stopped photon and the subsequent diffusion toward the anode, charge sharing will strongly increase and spoil the spectral response. Furthermore, Compton scatter and K-escape crosstalk will also increase. Therefore, there is a physical lower limit for admissible sub-pixel sizes, determined by an unacceptable loss of spectral information due to the above-mentioned effects. Further steps have to be taken to reach CT count rate capability by photon-counting detectors. In a recent patent application it was suggested to use layers of detectors on top of each other to share the high count rates between more layers (Tkaczyk et al 2007). Since attenuation caused by humans is rather pronounced, the count rate drops quickly when going from the unattenuated beam into the patient. The ‘count rate problem’ is therefore confined to rays passing through the peripheral parts of the body. It was also suggested in Tkaczyk et al (2007) to design the detector such that the count rates are distributed unevenly between the layers. The result is that for the high rate in the unattenuated primary beam some layers remain active while other layers saturate. The idea now is to disable overloaded detector layers and read out only the layers that are still active. The influence of this data acquisition scheme on noise performance has still to be investigated. Finally, the experiments presented here were acquired in a fan-beam geometry with only one detector line and without the use of any anti-scatter grid. On the other hand, any Experimental feasibility of multi-energy photon-counting K-edge imaging in pre-clinical CT 4045 commercial CT system nowadays uses the cone-beam geometry instead. It is well known from conventional CT that scattered radiation can become a significant image quality issue when no counter measures are applied (Engel et al 2008b). For spectral CT this issue could potentially be even more problematic since energy loss through Compton scatter will introduce additional changes in the transmitted spectra. Simulations, however, showed that even for rather thick objects this issue can be solved by the use of 2D anti-scatter grids (Engel et al 2008a). A combination of all steps described above may allow photon-counting to enter the medical CT domain, which in combination with novel contrast agents will enable new K-edge imaging applications. Acknowledgments We acknowledge Naor Wainer and Ami Altman (Philips Healthcare, Haifa, Israel) for fruitful and lively discussions, Bjørn Sundal (GammaMedica-Ideas, Oslo, Norway) for ongoing technical detector support and Adam Webb (Hasylab/DESY, Hamburg, Germany) for technical support during the synchrotron measurements. Appendix A.1. Detector energy scale calibration For the energy scale calibration we measured the detector response at two known energies El , l = 1, 2 by threshold scans (see section 4.1). We then determined u(k,i) T ,MAX,El as the (k,i) d . Then for each pixel and threshold a linear positions of the maxima of (k,i) c uT duT and energy UT(k,i) was assumed. The individual gains g (k,i) and the relationship between u(k,i) T (k,i) of all thresholds described this relationship and could be determined from the two offsets o measurements: (k,i) u(k,i) T ,MAX,E2 − uT ,MAX,E1 (k,i) (k,i) (k,i) = g U + o , g = , u(k,i) T (A.1) E2 − E1 (k,i) o(k,i) = u(k,i) E1 . T ,MAX,E1 − g Equations (A.1) were used to calculate the correct voltage settings for all pixels, such that all thresholds were set to the same photon energy level. A.2. Response function model The phenomenological model for the response function was motivated by the analysis of the response function measurements (see section 4.1). The model function consists of two Gaussian peaks, one at the incident energy and one at an energy reduced by the average energy of K-escape photons. Furthermore, a constant background for energies below the incident energy is present in the model. This results in the following formula: 1 1 (U − E)2 exp − R(U, E) = c1 (E) √ 2 σ1 (E)2 2π σ1 (E) 1 1 (U − Ee − E)2 + c2 (E) √ + B(U, E) . (A.2) exp − 2 σ2 (E)2 2π σ2 (E) The background B(U, E) has a constant value on the U -scale of c3 for U < (E − 3σ1 ) and is then linearly ramped down to zero within a width of 6σ1 . Ee is the average energy of an 4046 J P Schlomka et al Table 3. Parameters used in equation (A.3) resulting from the fit of equation (A.2) to the measured data shown in figure 2. Parameter Value a1 a2 a3 a4 a5 a6 0.5 0.015 keV−1 0.042 keV−1 0.213 × 10−3 keV−2 1.61 keV 0.025 escape photon (For CdTe this is ∼25 keV). The constant c1 (E) is taken such that the number of photons of the incoming flux is preserved. The constants c2 (E) and c3 (E) as well as the width of the photo- and escape peak σ1 (E) and σ2 (E)depend on the incident energy E. We used the measured response functions (figure 2) and included also measurements using a 57Co radioactive source with 122 keV photon energy to derive the dependences by fitting the measured spectra to the model function. To reduce the number of free parameters, the widths of the photo peak and the escape peak were assumed to be equal (σ (E) = σ1 (E) = σ2 (E)). From the fit to the data one gets the following relationships valid for the energy range 0–150 keV (energy taken in keV) (see table 3): a1 · exp(−a2 E), for E Ee c2 (E) = 0, for E < Ee (A.3) c3 (E) = a3 − a4 E σ (E) = a5 + a6 E. References Alvarez R E and Macovski A 1976 Energy-selective reconstructions in x-ray computerised tomography Phys. Med. Biol. 21 733–44 Baumer C, Martens G, Menser B, Roessl E, Schlomka J P, Steadman R and Zeitler G 2008 Testing an energy-dispersive counting-mode detector with hard x-rays from a synchrotron source IEEE Trans. Nucl. Sci. 55 1785–90 Barnes G T, Sones R A, Tesic M M, Morgan D R and Sanders J N 1985 Detector for dual-energy digital radiography Radiology 156 537–40 Brody W R, Cassel D M, Sommer F G, Lehmann L A, Macovski A, Alvarez R E, Pelc N J, Riederer S J and Hall A L 1981 Dual-energy projection radiography: initial clinical experience Am. J. Roentgenol. 137 201–5 Carmi R, Naveh G and Altman A 2005 Material separation with dual-layer CT IEEE Nucl. Sci. Symp. Conf. Rec. 4 1876–78 Cowan G 1998 Statistical Data Analysis (Oxford: Oxford University Press) Eadie W T, Drijard D, James F E, Roos M and Sadoulet B 1971 Statistical Methods in Experimental Physics (Amsterdam: North-Holland) Engel K J, Baumer C, Wiegert J and Zeitler G 2008a Spectral analysis of scattered radiation in CT Proc. SPIE 6913 69131R-1 Engel K J, Herrmann C and Zeitler G 2008b X-ray scattering in single- and dual-source CT Med. Phys. 35 318–32 Frey E C, Taguchi K, Kapusta M, Xu J, Orskaug T, Ninive I, Wagenaar D, Patt B and Tsui B M W 2007a Microcomputed tomography with a photon-counting x-ray detector Proc. SPIE 6510 65101R Frey E C, Wang X, Du Y, Taguchi K, Xu J and Tsui B M W 2007b Investigation of the use of photon counting x-ray detectors with energy discrimination capability for material decomposition in micro-computed tomography Proc. SPIE 6510 65100A Hasylab 2007 http://hasylab.desy.de/facilities/doris iii/beamlines/x1 roemo ii/index eng.html Iwanczyk J S, Nygård E, Meirav O, Arenson J, Barber W C, Hartsough N E, Malakhov N and Wessel J C 2007 Photon counting energy dispersive detector arrays for x-ray imaging IEEE Nucl. Sci. Symp. Conf. Rec. 4 2741–8 Kek A C and Slaney M 2001 Principles of Computerized Tomographic Imaging (Philadelphia, PA: SIAM) Kelcz F, Joseph P M and Hilal S K 1979 Noise considerations in dual energy CT scanning Med. Phys. 6 418–25 Experimental feasibility of multi-energy photon-counting K-edge imaging in pre-clinical CT 4047 Klein O and Nishina Y 1929 Über die Streuung von Strahlung durch freie Elektronen nach der neuen relativistischen Quantendynamik von Dirac Zeitschrift für Physik 52 853 Knoll G F 2000 Radiation Detection and Measurement (Hoboken: Wiley) Kowase K and Ogawa K 2006 Photon counting x-ray CT system with a semiconductor detector IEEE Nucl. Sci. Symp. Conf. Record 5 3119–23 Lehmann L and Alvarez R 1986 Energy selective radiography: A review Digital radiography: Selected Topics ed J Kareiakes, S Thomas and C Orton (New York: Plenum) pp 145–88 Lehmann L A, Alvarez R E, Macovski A, Brody W R, Pelc N J, Riederer S J and Hall A L 1981 Generalized image combinations in dual kVp digital radiography Med. Phys. 8 659–67 Nelder J A and Mead R 1965 A simplex method for function minimization Comput. J 7 308–13 Niederlöhner D 2006 Untersuchungen zur Energiewichtung in der medizinischen Röntgenbildgebung mit dem Medipix2-Detektor PhD Thesis University of Erlangen Press W H, Teukolsky S A, Vetterling W T and Flannery B P 1992 Numerical Recipes in C, The Art of Scientific Computing 2nd edn (Cambridge: Cambridge University Press) Remy-Jardin M, Bahepar J, Lafitte J J, Dequiedt P, Ertzbischoff O, Bruzzi J, Delannoy-Deken V, Duhamel A and Remy J 2006 Multi-detector row CT angiography of pulmonary circulation with gadolinium-based contrast agents: prospective evaluation in 60 patients Radiology 238 1022–35 Roessl E and Proksa R 2007 K-edge imaging in x-ray computed tomography using multi-bin photon counting detectors Phys. Med. Biol. 52 4679–96 Sato G et al 2005 Development of a spectral model based on charge transport for the swift/BAT 32K CdZnTe detector array Nucl. Instrum. Methods Phys. Res. A 541 372–84 Sukovic P and Clinthorne N H 1999 Basis material decomposition using triple-energy x-ray computed tomography Proc. 16th IEEE Instrum. Meas. Technol. Conf. vol 3 pp 1615–8 Takahashi T et al 2005 Application of CdTe for the NeXT mission Nucl. Instrum. Methods Phys. Res. A 541 332–41 Tajima H et al 2004 Performance of a low noise front-end ASIC for Si/CdTe detectors in Compton gamma-ray telescope IEEE Trans. Nucl. Sci. 51 842–47 Tkaczyk J E, Short J D, Rose J W, Wu X and Basu S K 2007 Photon counting x-ray detector with overrange logic control US Patent US2007/0206721A1