ieee transactions on biomedical engineering, vol. 59, no. 6, june

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 59, NO. 6, JUNE 2012 1711

Image Reconstruction for HybridTrue-Color Micro-CT

Qiong Xu, Hengyong Yu∗, Senior Member, IEEE, James Bennett, Peng He, Rafidah Zainon, Robert Doesburg,Alex Opie, Mike Walsh, Haiou Shen, Anthony Butler, Phillip Butler, Xuanqin Mou, and Ge Wang∗, Fellow, IEEE

Abstract—X-ray micro-CT is an important imaging tool forbiomedical researchers. Our group has recently proposed a hybrid“true-color” micro-CT system to improve contrast resolution withlower system cost and radiation dose. The system incorporatesan energy-resolved photon-counting true-color detector into aconventional micro-CT configuration, and can be used for materialdecomposition. In this paper, we demonstrate an interior color-CTimage reconstruction algorithm developed for this hybrid true-color micro-CT system. A compressive sensing-based statisticalinterior tomography method is employed to reconstruct each chan-nel in the local spectral imaging chain, where the reconstructedglobal gray-scale image from the conventional imaging chainserved as the initial guess. Principal component analysis was used

Manuscript received January 2, 2012; revised February 25, 2012; acceptedMarch 20, 2012. Date of publication April 3, 2012; date of current versionMay 18, 2012. This work was supported in part by the National Institutesof Health/National Institute of Biomedical Imaging and Bioengineering underGrant EB011785 and a seed grant from the Wake Forest Institute for Regenera-tive Medicine. Asterisk indicates corresponding author.

Q. Xu is with the Institute of Image Processing and Pattern Recogni-tion, Xi’an Jiaotong University, Xi’an, Shaanxi 710049, China, and alsowith the Department of Radiology, Division of Radiologic Sciences, WakeForest University Health Sciences, Winston-Salem, NC 27157 USA (e-mail:[email protected]).

∗H. Yu is with the Department of Radiology, Division of Radiologic Sciencesand Biomedical Imaging Division, VT-WFU School of Biomedical Engineer-ing and Sciences, Wake Forest University Health Sciences, Winston-Salem, NC27157 USA (e-mail: [email protected]).

J. Bennett, P. He, and H. Shen are with Biomedical Imaging Division, VT-WFU School of Biomedical Engineering and Sciences, Virginia Tech, Blacks-burg, VA 24061 USA (e-mail: [email protected]; [email protected];[email protected]).

R. Zainon and R. Doesburg are with the Department of Physics and As-tronomy, University of Canterbury, Christchurch 8140, New Zealand (e-mail:[email protected]; [email protected]).

A. Opie is with the Department of Electrical and Computer Engineer-ing, University of Canterbury, Christchurch 8140, New Zealand (e-mail:[email protected]).

M. Walsh is with the Department of Radiology, University of Otago,Christchurch 8011, New Zealand (e-mail: [email protected]).

A. Butler is with the Department of Radiology, University ofOtago, Christchurch 8011, New Zealand, and also with the EuropeanOrganization for Nuclear Research (CERN), Geneva 23, Switzerland (e-mail:[email protected]).

P. Butler is with the Department of Physics and Astronomy, University ofCanterbury, Christchurch 8140, New Zealand, and also with the European Or-ganization for Nuclear Research (CERN), Geneva 23, Switzerland (e-mail:[email protected]).

X. Mou is with the Institute of Image Processing and Pattern Recognition,Xi’an Jiaotong University, Xi’an, Shaanxi 710049, China (e-mail: [email protected]).

∗G. Wang is with the Biomedical Imaging Division, VT-WFU School ofBiomedical Engineering and Sciences, Virginia Tech, Blacksburg, VA 24061,and also with Wake Forest University Health Sciences, Winston-Salem, NC27157 USA (e-mail: [email protected]).

Digital Object Identifier 10.1109/TBME.2012.2192119

to map the spectral reconstructions into the color space. The pro-posed algorithm was evaluated by numerical simulations, physicalphantom experiments, and animal studies. The results confirm themerits of the proposed algorithm, and demonstrate the feasibilityof the hybrid true-color micro-CT system. Additionally, a “colordiffusion” phenomenon was observed whereby high-quality true-color images are produced not only inside the region of interest,but also in neighboring regions. It appears harnessing that thisphenomenon could potentially reduce the color detector size for agiven ROI, further reducing system cost and radiation dose.

Index Terms—Color/spectral-CT, micro-CT, photon-counting, sta-tistical interior tomography.

I. INTRODUCTION

M ICROIMAGING technology for small-animal studieshas developed rapidly in the past decade due to its

critical importance for systems biology and medicine. Atpresent, biomedical research institutions and companies areextensively using microimaging tools, including x-ray micro-CT, micromagnetic resonance imaging (micro-MRI), microp-ositron emission tomography (micro-PET), microsingle photonemission computed tomography (micro-SPECT), microultra-sound (micro-US), micro-optical scanners, etc. X-ray micro-CT is generally considered superior among these microimagingtools; it outperforms micro-MRI in terms of imaging speed andcost effectiveness, captures much finer features than micro-PETand micro-SPECT, brings much less artifacts than micro-US,and allows significantly deeper sample penetration than opticalimaging.

However, a major limitation of current micro-CT scan-ners is insufficient contrast resolution for soft tissue. Energy-integrating detectors are a major cause for this limitationbecause the detected x-ray signals are accumulated over theentire source spectrum, which makes it difficult to utilize theenergy-dependent property of x-ray attenuation [1]. In con-trast, recently developed energy-resolved photon-counting de-tectors can resolve individual photons and their associated en-ergies [2]. Photon-counting detectors can measure the x-rayattenuation under different spectral channels simultaneously us-ing a conventional polychromatic x-ray source. Consequently,they are capable of revealing elemental composition of dif-ferent materials. In addition, photon-counting detectors havean inherently higher signal-to-noise ratio (SNR) because elec-tronic noise below the counting threshold does not increment thecounter. These photon-counting detectors have been used to de-velop true-color CT systems that dramatically improve contrast

0018-9294/$31.00 © 2012 IEEE

1712 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 59, NO. 6, JUNE 2012

resolution [3]–[9]. Here, “true color” is relative to the well-known pseudocolor images mapped from gray-scale imagesin clinical applications. True-color CT is also referred to asspectral, multienergy, spectroscopic, energy-selective or energy-sensitive CT.

Given innumerable potential applications of spectral x-rayCT (e.g., tissue characterization, functional studies, cellular andmolecular imaging), the transition from gray scale to true coloris almost certain to occur sooner or later. However, there aretwo major challenges in this process: detector cost and radi-ation dose. First, the spectral detection technology is not yetmature for large area, high resolution, uniform performance,and long duration operation. The replacement of a gray-scalex-ray detector array with a true-color spectral detector will berather expensive in the near future. Second, radiation exposureis a public concern, and x-ray spectral detection would requiremuch higher radiation dosage if each spectral channel requiresthe same exposure as that for a corresponding energy-integratingdetector. These issues will require further research and refine-ment, but are far from terminal limitations.

While classic CT theory targets exact reconstruction ofthe sample’s entire cross section from complete projections,biomedical applications of CT and micro-CT often focus onsmaller, internal regions of interest (ROIs). Unfortunately, clas-sic CT theory cannot exactly reconstruct an internal ROI fromtruncated x-ray projections that only pass through the ROIbecause this “interior problem” does not have a unique solu-tion [10]. However, in recent years it has been proven that theinterior problem can be exactly solved if there is some addi-tional prior information. One method is based on the truncatedHilbert transformation (THT) with a known subregion insidethe ROI [11]–[16]. Another method is based on the compressivesensing (CS) theory, assuming a piecewise constant or poly-nomial ROI [17]–[19]. These “interior tomography” methodsprovide a way to exactly reconstruct an internal ROI with re-duced radiation dose and detector size.

We propose that the detector cost and radiation dose problemsof x-ray spectral micro-CT can be solved by applying interiortomography. Recently, we have designed a hybrid true-colormicro-CT system that incorporates spectral imaging and inte-rior tomography to improve micro-CT performance [20]. Ourdesign adds a narrow-beam true-color imaging chain in the cur-rent micro-CT system using a small energy-resolved photon-counting detector. This hybrid system has three advantages:first, it keeps the conventional gray-scale imaging chain, whichis necessary for a transition from gray-scale to true-color recon-struction and important for locating the ROI more accurately;second, it can provide higher contrast true-color ROI recon-struction with lower system cost and radiation dose; third, thegray-scale global reconstruction provides prior information fora better true-color ROI reconstruction.

The remainder of this paper is organized as follows. SectionII describes the hybrid true-color micro-CT system; Section IIIpresents the interior-color reconstruction algorithm; experimen-tal results will be shown in Section IV; finally, we will discussrelated problems and conclude this paper in Section V.

Fig. 1. 3-D rendering of the hybrid true-color micro-CT system. A wide-beam gray-scale imaging chain and a narrow-beam true-color imaging chainare combined on a rotating gantry.

II. HYBRID TRUE-COLOR MICRO-CT

Recently, our group has designed a hybrid true-color micro-CT system [20]. The system consists of a wide-beam gray-scale imaging chain perpendicular to a narrow-beam true-colorimaging chain on a rotating slip ring (see Fig. 1). The wide-beam imaging chain is designed to have a field of view (FOV)of radius 36 mm, which provides conventional projection datafor global gray-scale imaging. The narrow-beam imaging chainis designed to have an FOV of radius 5 mm, which provides aset of spectral projection data for true-color ROI reconstruction(see Fig. 2).

The scanner’s multispectral photon-counting x-ray detectorwas selected as Medipix3. The Medipix is a series of state-of-the-art photon-counting detectors for x-ray microimaging,and the Medipix3 is the latest version of Medipix series with ahigher energy resolution [21]–[25]. A single Medipix3 chip isemployed in our design, and has 256×256 matrix of 55×55 μmpixels. The readout logic of Medipix3 supports eight energythresholds for spectroscopic imaging. The gray-scale x-ray de-tector is a 120×120 mm sensitive area with a pixel size of50×50 μm. The two x-ray tubes are identical and designedto allow the K-edge imaging of gold nanoparticles (GNP) at80.7 KeV. The targets of this system are 50 μm spatial res-olution, 20 Hounsfield unit (HU) noise deviation, 8 spectralchannels, and up to 15 rotations per minute.

III. METHODOLOGY

A single global polychromatic (energy-integrating) projec-tion dataset and several local spectral projection datasets canbe obtained with our proposed hybrid imaging system. Conven-tional analytical or iterative reconstruction methods can be usedto reconstruct the global projection data, while interior tomog-raphy methods can be used to reconstruct an interior ROI imageon each channel from the local spectral projection data. Aninterior-color image can be rendered from this set of ROI recon-structions via image analysis methods; however, the gray-scaleglobal reconstruction can be used to improve the performance

XU et al.: IMAGE RECONSTRUCTION FOR HYBRID TRUE-COLOR MICRO-CT 1713

Fig. 2. 2-D sketch of the hybrid true-color micro-CT system. A wide-beamgray-scale imaging chain and a narrow-beam true-color imaging chain are com-bined on a rotating gantry.

of the interior-color reconstruction. We will next describe thisreconstruction method in detail.

A. CS-Based Interior Tomography

We used CS-based interior tomography to reconstruct eachspectral channel (energy threshold) for two reasons. First, CS-based interior tomography does not need a known subregionwithin the ROI, which is required by landmark-based methodsand usually difficult to be exactly obtained, especially for de-termining the exact attenuations of a subregion on a differentspectral channel. Second, CS-based interior tomography betteraccommodates the statistical property of the projections andperforms better with noisy data, which is important for multi-energy data due to the lower counts and higher noise in eachspectral channel. Recently, it has been demonstrated that the CS-statistical interior tomography (CS-SIT) performs very well inlow count scenarios [26]. In this paper, we employ an enhancedCS-SIT algorithm for interior reconstruction on each spectralchannel, aided by the gray-scale global reconstruction.

The detected photons of each spectral channel s(s =1, . . . , S, S is the number of spectral channels) can be approxi-mately modeled as a Poisson distribution

ysi ∼ Poisson {bs

i exp (−lsi ) + rsi } , i = 1, . . . , I (1)

where ysi is the measured data of the spectral channel s along the

ith x-ray path, bsi is the blank scan factor, lsi =

∑Jj=1 aijμ

si =

[Aμs ]i is the integral of the x-ray linear attenuation coefficients,

A = {aij} is the system matrix, μs = (μs1 , . . . , μ

sJ )T is a distri-

bution of linear attenuation coefficients, rsi accounts for read-out

noise, and I and J are the number of projections and pixels.According to the CS-based interior tomography theory, a

piecewise constant ROI can be exactly reconstructed from thelocally truncated projections by minimizing the image total vari-ation (TV). Combining the Poisson property of projection dataand the TV regularization in the maximization of a posterior(MAP) framework, an ROI image on each spectral channel scan be exactly reconstructed via minimizing the following ob-jective function:

Φ (μs ) =I∑

i=1

wsi

2

([Aμs ]i − lsi

)2+ βsTV (μs ) , s = 1, . . . , S

(2)where ws

i = (ysi − rs

i )2 /ys

i is the statistical weight for each x-ray path, lsi = ln (bs

i /(ysi − rs

i )) is the estimated line integral,βs is the regularization parameter to balance the data fidelityand TV terms, and TV (μs) is the operator of computing TV ofthe reconstructed image.

B. Principle Component Analysis

After we obtain the reconstructed images {μs}Ss=1 of all

spectral channels, the next important task is to extract featuresfrom these reconstructions to render a color image μC . Principlecomponent analysis (PCA) is a simple and effective method toidentify independent information from highly correlated data[27]. It has been applied in spectral CT and performs well [6],[28], which is why we adopted this method in our study.

The sample’s attenuation coefficients are embedded in eachspectral channel from the spectral image set {μs}S

s=1 . Identicalmaterials should have the same pattern of attenuation changefrom one spectral channel to another (i.e., different energythresholds). We aim to discriminate different materials by iden-tifying these patterns through the PCA method, which will bebriefly described.

First, we introduce a matrix M = {Msj} ∈ RS×J , Msj =

μsj − (1/S)

∑Ss=1 μs

j , and the covariance of the images set{μs}S

s=1 in spectral domain is C = (1/(S − 1))MMT (the su-perscript “T ” is the transpose operator). Applying singular valuedecomposition to C and arranging the eigenvalues in decreasingorder, the corresponding eigenvectors {vs}S

s=1 form a new ba-sis, which are called principal components. The matrix M canbe transformed by this basis to form the principal componentimage (PC image) set {Ps}S

s=1 , PTs = vT

s M. We can use thefirst N PC images {Ps}N

s=1 to discriminate between variousmaterials.

Furthermore, we have a gray-scale global image μG with agood spatial resolution, which can be used for color mapping.Thus, we combined the PC images with the gray-scale globalimage to produce the final high-resolution color image μC .

As for color mapping, we can either adopt the RGB or HSVmapping. The red, green, blue components or the hue, satu-ration, value components of the final color image will be lin-ear/nonlinear mapped from the PC images and gray-scale globalimage or the linear/nonlinear combinations of these images. Our


Fig. 3. Image reconstruction workflow for the hybrid true-color micro-CTsystem.

team is working to optimize a general map scheme for bettervisualization based on computer vision theory.

C. Overall Workflow

The overall workflow of the hybrid imaging system recon-struction algorithm is shown in Fig. 3.

Global projection data are denoted as {yGi }IG

i=1 and IG is thenumber of global projections. In this case, the statistical itera-tive reconstruction method with a TV regularization constraint(SIR-TV) is used to reconstruct the gray-scale global image μG

due to better performance for low count data. Due to the addi-tional spectral imaging chain, the gray-scale chain should havereduced flux, and therefore reduced counts, to keep a reasonableoverall dose for the combined imaging chains. Besides, SIR-TVand CS-SIT are identical except for the input data type. Thereconstruction of μG by SIR-TV is to minimize the followingobjective function:

Φ(μG

)=

IG∑

i=1

wGi

2

([AμG

]i− lGi

)2+ βGTV

(μG

)(3)

where wGi is the statistical weight for each projection of this

global data, and lGi and βG are the estimated line integral andthe regularization parameter of this case, respectively. Equation(3) can also be effectively and efficiently solved by the softthreshold filtering-based alternating minimization algorithm as(2). This global result μG will be used to regularize the interiorcolor reconstruction.

On each spectral channel, the image μs is reconstructed bythe CS-SIT with μG as the initial image. Then a PCA is per-formed on this set of spectral results {μs}S

s=1 . The first NPC-images {Ps}N

s=1 are selected and combined with the globalreconstruction μG to render a color image μC .

IV. EXPERIMENTAL RESULTS

To evaluate the proposed reconstruction method and demon-strate the feasibility of the hybrid true-color micro-CT system,we first performed a numerical simulation experiment. Further-more, the reconstruction method is validated using the datasetsfrom spectral micro-CT scans of physical phantoms, in vitrohuman samples, and euthanized GNP-injected mice.

Fig. 4. Transverse plane of the cylindrical phantom with seven differentmaterials.

TABLE ICYLINDRICAL PHANTOM GEOMETRY AND COMPOSITION

A. Numerical Simulation

In this simulation, a cylindrical phantom was designed toevaluate the spatial and contrast resolution of the proposed hy-brid system and reconstruction methodology; it contains severalcylindrical inclusions of various radii and seven materials withdifferent energy attenuation properties. The phantom parametersin transverse plane are shown in Fig. 4 and Table I. The concen-tration percentages of base materials are calculated by weight.The linear attenuation coefficients of the phantom materials areshown in Fig. 5.

Both the wide-beam gray-scale imaging chain and narrow-beam color-scale imaging chain were simulated with a fan-beam geometry and equidistant detector. The virtual detectorswere centered at the system origin and perpendicular to the linesfrom the system origin to its corresponding x-ray source. Thedistance from each source to the system origin was 115 mm. Thewide-beam and narrow-beam imaging chain detectors have 540and 256 detector elements, respectively, with an element widthof 0.04 mm. Six hunderd equiangular projections were collectedover 360◦. The x-ray tube voltage was assumed as 120 kVp; its


Fig. 5. Linear attenuation coefficients with respect to incident photon energyof the phantom materials in the numerical simulation.

Fig. 6. Source photon emission spectra used for numerical simulation, whichexpresses the normalized ratios of photons with different energies. Four specificenergies are denoted along the x-axis, which are the k-edges of iodine (33 keV),barium (37.4 keV), gadolinium (50.2 keV), and gold (80.7 keV).

normalized emission spectrum is shown in Fig. 6. Taking intoaccount the k-edges of materials in this phantom, which areiodine (33 keV), barium (37.4 keV), gadolinium (50.2 keV),and gold (80.7 keV), five spectral channels were selected in thecolor-scale imaging chain: ≤32 keV, 33–37 keV, 38–42 keV,51–56 keV, and ≥81 keV. In order to simulate various noisyscenarios, both imaging chains were simulated with emittedphoton counts along each x-ray path as 105 , 5×104 , 2×104 , and104 . The emitted photon counts of the true-color imaging chainwere calculated by weighting the tube spectrum distribution;thus, each spectral channel contained only about 10% of thetotal photons and resulted in more noise in the projection data.

The image reconstruction process was implemented follow-ing the workflow illustrated in Fig. 3. An ordered-subsets tech-nique [29] and a fast iterative shrinkage threshold algorithm(FISTA) [30] were employed to speed up the iterative process.The reconstructed images are 600×600 pixels covering a region

of 10×10 mm; the iteration count was fixed at 20. For compar-ison, images were also reconstructed by the conventional FBP(filtered-backprojection) method, which first smoothly extrapo-lates the truncated local projections to zeros and then performsglobal reconstruction in each spectral channel. A constant shiftwas imposed to each FBP reconstruction to make the central re-gion match the original phantom image. Finally, the same PCAand color mapping scheme were used to render color imagesfrom the images reconstructed by the FBP method.

The reconstructed results are shown in Fig. 7. The gray-scale images are reconstructed from global projectionsvia the SIR-TV method. In these gray-scale results, thematerials “12.4%Ca+87.6%Water,” “1.2%Iodine+98.8%Water,” “1.4%Barium+98.6%Water,” “1.5%Gadolinium+98.5%Water,” and “1.6%Gold+98.4%Water” have similargray scales, which cannot be discriminated between eachother. However, in the true-color interior reconstructions by theproposed method, different materials are mapped into differentcolors and can be easily distinguished. Therefore, we can seethat spectral scanning can provide color reconstructions withmuch higher contrast. At the same time, it can be seen that usingthe proposed method the true-color interior ROIs have the samenoise and spatial resolution as the gray-scale reconstructions.When the emitted photon count decreases from 105 to 104 ,both the gray-scale and true-color results have increased noiseand decreased spatial resolution. In the cases of 2×104 and 104

photons, the smallest inclusion is contaminated by the severenoise and cannot be recognized. It is worth noting that althoughthe spectral scanning is only performed in the ROI, high-qualitytrue-color reconstruction can be produced not only inside theROI but also in neighboring regions. As for the true-colorimages based on the FBP reconstructions, there is much noiseand the materials, such as “1.2%Iodine+98.8%Water” and“1.4%Barium+98.6%Water,” cannot be discriminated effec-tively, which is due to the inaccuracy of interior tomographyby the FBP method in each spectral channel.

B. Phantom Experiment

To further evaluate the proposed reconstruction algorithmand demonstrate the feasibility of the hybrid true-color micro-CT system, we designed a physical phantom. It is a Perspexcylinder of 10 mm diameter with six cylindrical inclusions of2.5 mm diameter filled with solutions of 2.0M calcium chlo-ride, 0.4M ferric nitrate, 0.02M iodinated water, sunflower oil,water, and air. This phantom was scanned by the state-of-the-art spectral micro-CT scanner—Medipix All Resolution System(MARS)—with a Medipix MXR Si detector [31]. The distancesfrom the source to the system origin and the source to the de-tector are 141 and 203 mm, respectively. Two hundred and fiftyequiangular projections were collected over 360◦ to form a fullscan. We used the Medipix3 detector chip for this experiment;it has 256×256 elements of size 55×55 μm. The detector chipwas horizontally translated for each projection to cover a widerFOV, producing a 438×256 projection for each view. The cen-tral slice was selected for image reconstruction with typicalfan-beam geometry of 250 views and 438 equidistant detector


Fig. 7. Reconstructed images in numerical simulation. The first to third rows are the global gray-scale reconstructions by SIR-TV, the true-color interiorreconstructions by the proposed method, and the true-color results based on FBP reconstructions (with appropriate extrapolation and constant shift), respectively.From first to fourth columns are the results from the projections assuming 105 , 5×104 , 2×104 , and 104 photons, respectively.

cells. The x-ray source was selected as the Source-Ray SB-80-1K (Source-Ray Inc., New York), which has a minimum focalspot of 50 μm. The x-ray tube voltage and current was set to50 kVp and 500 mA, respectively. Each spectral channel detectsphotons with energy greater than the selected threshold. Six en-ergy thresholds were selected as approximately 9.8, 15.1, 20.4,25.6, 30.9, and 36.2 keV. The acquired datasets were processedby flat-field correction and denoising processing. The datasetwith the lowest energy threshold is assumed as the projection ofgray-scale global imaging chain; other datasets are truncated tosimulate the spectral ROI scanning. The ROI was located in thephantom center with a radius of 2.75 mm.

The reconstructed images are 500×500 pixels covering a re-gion of 11.46×11.46 mm (see Fig. 8). The reconstruction resultsare similar to those of the previous phantom simulation exper-iment. Fig. 8 shows that the proposed algorithm performs verywell for true-color image reconstruction in the ROI: the recon-structed materials have strong contrast between each other andcan be easily discriminated in the color space. Additionally, itis shown that a good color reconstruction can also be obtainedin the neighboring region of the ROI.

C. In Vitro Study

Another study to validate our proposed reconstruction tech-nique was an in vitro human tissue scanned on the MARS micro-CT. An extracted human atheroma sample was placed in a plastictube (15.8 mm diameter) and scanned using a quad-Medipix 3detector (silicon layer, four detectors in 2×2 arrangement). Thedistance from the source to the system origin was 129.2 mm,and the distance from the source to the detector was 197.6 mm.Two hundred and fifty equiangular projections were collectedover 360◦ to form a full scan. The central slice was picked forimage reconstruction in typical fan-beam geometry. The x-raysource was Source-Ray SB-80-1K and the x-ray tube was set to50 kVp and 500 mA with 1.8-mm aluminum filter. Four spectralchannels were selected as ≥10, ≥16, ≥22, and ≥28 keV. Thedatasets were processed identically to the previous phantom ex-periment with the ROI (3.95 mm radius) located as shown inFig. 9.

The reconstructed images are 516×516 pixels covering aregion of 18.56×18.56 mm. The gray-scale global image is re-constructed by the SIR-TV method. From Fig. 9, it can be seenthat there are some plaque calcifications in this human atheroma


Fig. 8. Interior color reconstruction results of a physical phantom. The right one is the gray-scale image reconstructed from global projections using the statisticaliterative reconstruction method (SIR-TV); the middle and right one are the images of true-color interior reconstruction (CS-SIT) using two different color mapschemes.

Fig. 9. Results from the spectral scan of excised human atheroma sample; theleft one is the global gray-scale reconstruction, and the right one is the true-colorinterior reconstruction.

sample. The plaque calcifications in the ROI can be easily dis-criminated in the color space. Besides, the plaque calcificationsoutside of the ROI can also be discriminated in this color-interiorreconstruction.

D. GNP Mice Study

To evaluate the feasibility of the designed hybrid true-colormicro-CT system for GNP-based applications, two euthanizedmice were scanned on the MARS micro-CT with a MedipixMXR CdTe layer detector. The x-ray source was selected as theSource-Ray SB-120-400 (Source-Ray Inc., New York), whichhas a minimum focal spot of 75 μm. One mouse was injectedwith 0.2 mL of 15 nm Aurovist II GNP (Nanoparticles; Yaphank,NY) into the tail vein and was alive for ∼3 h between injectionand euthanasia. The second mouse was injected with 0.2 mLof 15 nm Aurovist II GNP (Nanoparticles; Yaphank, NY) in adirect cardiac puncture injection and immediate euthanasia. Arepresentative sample of the scanned mice is shown in Fig. 10.The distances from the source to the system origin and the sourceto the detector are 158 and 255 mm, respectively. Three hundredand seventy-one equiangular projections were collected over360o to form a full scan. Thirteen energy bins were collectedwith the source tube operated with 120 kVp and 175 mA. Thedetector chip was moved horizontally with overlapped pixelsto cover a wider FOV of 34.89 mm diameter and correct forproduction defects in the detector sensor layer. Since there was

Fig. 10. Photo of a BALB/C mouse used in this study after GNP injection.

Fig. 11. Sinogram ring artifacts reduction; the left one is the sinogram beforepreprocessing, and the right one is the sinogram after the ring artifacts removing.

significant noise in the sinogram, neighboring detector bins weremerged to form a new sinogram of size 512×371. Next, weused the method in [32] to reduce ring artifact (see Fig. 11).As mentioned, the dataset with the lowest energy threshold wasassumed to be the gray-scale projection from the global imagingchain. Other datasets were truncated to simulate spectral ROIscanning; the ROIs are indicated in Fig. 12 with a radius of3.4 mm.

The reconstructed images are 512×512 pixels covering a re-gion of 18.41×18.41 mm. The reconstructed results are shownin Fig. 12. The gray-scale global images for these two mice werereconstructed by the SIR-TV method. These image slices are ofthe mouse upper thorax and include the front limbs, thoracic ver-tebra, and scapula. The true-color reconstructions demonstratethat the GNP was present within the vascular structures of the


Fig. 12. Results from GNP mice studies. The first and second rows are theglobal gray-scale and true-color reconstruction results; and the third row is thecolor interior reconstructions. The left and right columns are the central slicesfrom the mouse with tail vein injection of GNP and with direct cardiac punctureinjection of GNP, respectively.

upper thorax. It can be seen that the color interior tomographycan nearly reproduce the same color result in ROI as with globalreconstruction. The calcium in the bones and injected GNP havevarying energy-specific attenuation properties compared to thebackground soft tissues which is represented as color variationin the PCA images.

V. DISCUSSION AND CONCLUSION

X-ray attenuation is both material and energy dependent.Conventional energy-integrating detectors measure x-ray atten-uation along the path integrated in the whole energy spectrum.The resultant images from these detectors can be viewed asthe weighted average of the material’s x-ray attenuation coef-ficients across the entire source spectrum, and do not captureenergy-dependent characteristics. In many situations, the x-rayattenuation coefficients of different materials are similar in theconventional gray-scale CT, and it is hard to discriminate be-tween these materials. As shown in our numerical simulations,“12.4%Ca+87.6%Water,” “1.2%Iodine+98.8%Water,” “1.4%Barium+98.6%Water,” “1.5%Gadolinium+98.5%Water,” and“1.6%Gold+98.4%Water” have similar gray scales. This phe-

nomenon contributes to lower contrast for traditional gray-scaleCT.

However, the photon-counting detectors are able to resolve in-cident photon energy and thus can measure the x-ray attenuationcoefficients along the x-ray path in each spectral channel simul-taneously. The energy-dependent property of x-ray attenuationis embedded in the spectral dataset, which can be used to ex-tract additional information about the materials. One importantfeature is the material’s absorption edge (k-edge/L-edge), whichproduces a sudden increase in attenuation coefficient as the x-rayphoton energy matches that of the material’s valence shell elec-trons. This kind of sudden change would be reflected in spectraldata, while it would be concealed in energy-integrated scanning.Consequently, the contrast resolution will be improved signifi-cantly by using the photon-counting technique, which has beenconfirmed by our results in Section IV. However, the true-colorCT technique is limited by the expense of photon-counting de-tector chips and higher radiation dose to provide each spectralchannel with sufficient counts.

Interior tomography exactly reconstructs an interior ROI onlyfrom the data passing through the ROI, and can be applied toboth decrease the detector size and reduce radiation dose. Con-sidering that many imaging applications are focused on only asmall ROI, it will be useful to incorporate the interior tomog-raphy and true-color techniques with modern CT systems; thisis the motivation for our work reported in this paper. Our pro-posed hybrid true-color micro-CT system adds a narrow-beamspectral imaging chain into the conventional micro-CT systemand aims to provide high contrast resolution in the ROI with alow system cost and radiation dose. From the results in SectionIV, it can be seen that this system is feasible and can providea high-quality color reconstruction. In the conventional gray-scale imaging chain, the noise and spatial resolution are relatedto the SNR of projection data, and, from the simulation ex-periment in Section IV, we can see that as the photon numberis decreased from 105 to 104 , the noise is increased and thespatial resolution is degraded. For the spectral data, usually thecounts in each spectral channel are much lower than the energy-integrating projection datasets which will lead to higher noiseand lower spatial resolution in the reconstruction. Fortunately,the gray-scale imaging chain can be used not only to provideglobal gray-scale image, but also to enhance the interior tomog-raphy in each spectral channel, and improve spatial resolutionand suppress image noise. It can be seen from the simulationexperiment that the interior color reconstructions can maintainthe same spatial resolution and noise level as the gray-scalereconstruction.

Given all the results in Section IV, one can see that a high-quality true-color image is typically available in the neighboringregion of the ROI, even as the spectral imaging chain only coversthe ROI itself. We propose to name this phenomenon as “colordiffusion,” which appears to benefit from stabilizing the interiortomography of each spectral channel aided by the global gray-scale reconstruction. Using this color diffusion phenomenon,we can potentially use a smaller true-color detector for a givenROI and thus reduce system cost. For example, if we wanted toreconstruct an ROI of radius 20.0 mm, we may choose a smaller


true-color detector which only covers a 14.0 mm radius field ofview.

At present, the Medipix photon-counting detector is the state-of-the-art spectral imaging technology and a very exciting re-search topic in the CT field. We believe that this technique willtransform CT imaging from gray scale to true color by provid-ing the spectral information for the imaging object. However,this technique is still under development and has some unsolvedproblems, such as the charge sharing which negatively affectsthe spectral data accuracy and hinders the absorption edge de-tection [33]. More efforts are needed to develop and improvephoton-counting technology.

In conclusion, we proposed an image reconstruction methodfor a hybrid true-color micro-CT system, which has an addi-tional spectral imaging chain based on a conventional micro-CT configuration. The spectral imaging chain incorporates anenergy-resolving photon-counting detector and an interior to-mography reconstruction technique. With the proposed recon-struction method, this system can provide a low noise, highspatial and contrast resolution true-color images with low sys-tem cost and radiation dose. Furthermore, the observed colordiffusion can help to reduce the color detector size for a givenROI to further reduce system cost and radiation dose.

REFERENCES

[1] R. E. Alvarez and A. Macovski, “Energy-selective reconstructions in x-ray computerised tomography,” Phys. Med. Biol., vol. 21, pp. 733–744,1976.

[2] J. Jak◦ubek, “Semiconductor pixel detectors and their applications in life

sciences,” J. Instrum., vol. 4, Article ID P03013, pp. 1–21, 2009.[3] E. Roessl and R. Proksa, “K-edge imaging in x-ray computed tomography

using multi-bin photon counting detectors,” Phys. Med. Biol., vol. 52,pp. 4679–4696, 2007.

[4] S. Feuerlein, E. Roessl, R. Proksa, G. Martens, O. Klass, M. Jeltsch,V. Rasche, H. J. Brambs, M. H. K. Hoffmann, and J. P. Schlomka, “Multi-energy Photon-counting K-edge imaging: Potential for improved luminaldepiction in vascular imaging,” Radiology, vol. 249, pp. 1010–1016,2008.

[5] J. Schlomka, E. Roessl, R. Dorscheid, S. Dill, G. Martens, T. Istel,C. Baumer, C. Herrmann, R. Steadman, and G. Zeitler, “Experimental fea-sibility of multi-energy photon-counting K-edge imaging in pre-clinicalcomputed tomography,” Phys. Med. Biol., vol. 53, pp. 4031–4047, 2008.

[6] N. Anderson, A. Butler, N. J. A. Scott, N. Cook, J. Butzer, N. Schleich,M. Firsching, R. Grasset, N. de Ruiter, and M. Campbell, “Spectroscopic(multi-energy) CT distinguishes iodine and barium contrast material inMICE,” Eur. Radiol., vol. 20, pp. 2126–2134, 2010.

[7] D. P. Cormode, E. Roessl, A. Thran, T. Skajaa, R. E. Gordon, J. P.Schlomka, V. Fuster, E. A. Fisher, W. J. M. Mulder, and R. Proksa,“Atherosclerotic plaque composition: Analysis with multicolor CT andtargeted gold nanoparticles,” Radiology, vol. 256, pp. 774–782, 2010.

[8] X. Wang, D. Meier, S. Mikkelsen, G. Maehlum, D. Wagenaar, B. Tsui,B. Patt, and E. Frey, “MicroCT with energy-resolved photon-countingdetectors,” Phys. Med. Biol., vol. 56, pp. 2791–2816, 2011.

[9] X. Wang, D. Meier, K. Taguchi, D. J. Wagenaar, B. E. Patt, and E. C. Frey,“Material separation in x-ray CT with energy resolved photon-countingdetectors,” Med. Phys., vol. 38, pp. 1534–1546, 2011.

[10] F. Natterer, The Mathematics of Computerized Tomography. Philadel-phia, PA, USA: Society for Industrial Mathematics, 2001.

[11] Y. B. Ye, H. Y. Yu, Y. C. Wei, and G. Wang, “A general local reconstructionapproach based on a truncated Hilbert transform,” Int. J. Biomed. Imag.,vol. 2007, Article ID 63634, pp. 1–8, 2007.

[12] Y. B. Ye, H. Y. Yu, and G. Wang, “Exact interior reconstruction with cone-beam CT,” Int. J. Biomed. Imag., vol. 2007, Article ID:10693, pp. 1–5,2007.

[13] Y. B. Ye, H. Y. Yu, and G. Wang, “Exact interior reconstruction fromtruncated limited-angle projection data,” Int. J. Biomed. Imag., vol. 2008,Article ID 427989, pp. 1–6, 2008.

[14] H. Y. Yu, Y. Ye, and G. Wang, “Interior reconstruction using the trun-cated Hilbert transform via singular value decomposition,” J. X-Ray Sci.Technol., vol. 16, pp. 243–251, 2008.

[15] H. Kudo, M. Courdurier, F. Noo, and M. Defrise, “Tiny a priori knowledgesolves the interior problem in CT,” Phys. Med. Biol., vol. 53, pp. 2207–2231, 2008.

[16] M. Courdurier, F. Noo, M. Defrise, and H. Kudo, “Solving the interiorproblem of computed tomography using a priori knowledge,” InverseProblems, vol. 24, Article ID 065001, pp. 1–27, 2008.

[17] H. Y. Yu and G. Wang, “Compressed sensing based interior tomography,”Phys. Med. Biol., vol. 54, pp. 2791–2805, 2009.

[18] H. Y. Yu, J. S. Yang, M. Jiang, and G. Wang, “Supplemental analysison compressed sensing based interior tomography,” Phys. Med. Biol.,vol. 54, pp. N425–N432, 2009.

[19] J. S. Yang, H. Y. Yu, M. Jiang, and G. Wang, “High-order total variationminimization for interior tomography,” Inverse Problems, vol. 26, ArticleID 035013, pp. 1–29, 2010.

[20] G. Wang, “Computed tomography based imaging systems,” US PatentApplication, No. 12916458, Oct. 29, 2010.

[21] C. Bert, D. Niederlohner, J. Giersch, K. F. Pfeiffer, and G. Anton, “Com-puted tomography using the Medipix1 chip,” Nuclear Instrum. MethodsPhys. Res. Section A: Accelerators, Spectrometers, Detectors AssociatedEquipment, vol. 509, pp. 240–250, 2003.

[22] X. Llopart, M. Campbell, R. Dinapoli, D. San Segundo, and E. Pernigotti,“Medipix2: A 64-k pixel readout chip with 55-μm square elements work-ing in single photon counting mode,” IEEE Trans. Nuclear Sci., vol. 49,no. 5, pp. 2279–2283, 2002.

[23] M. Firsching, P. Takoukam Talla, T. Michel, and G. Anton, “Materialresolving X-ray imaging using spectrum reconstruction with Medipix2,”Nuclear Instrum. Methods Phys. Res. Section A: Accelerators, Spectrom-eters, Detectors Associated Equipment, vol. 591, pp. 19–23, 2008.

[24] M. Firsching, A. P. Butler, N. Scott, N. G. Anderson, T. Michel, and G.Anton, “Contrast agent recognition in small animal CT using the Medipix2detector,” Nuclear Instrum. Methods Phys. Res. Section A: Accelerators,Spectrometers, Detectors Associated Equipment, vol. 607, pp. 179–182,2009.

[25] R. Ballabriga, M. Campbell, E. Heijne, X. Llopart, and L. Tlustos, “TheMedipix3 prototype, a pixel readout chip working in single photon count-ing mode with improved spectrometric performance,” IEEE Trans. Nu-clear Sci., vol. 54, no. 5, pp. 1824–1829, Oct. 2007.

[26] Q. Xu, X. Q. Mou, G. Wang, J. Sieren, E. A. Hoffman, and H. Y. Yu,“Statistical interior tomography,” IEEE Trans. Med. Imag., vol. 30, no. 5,pp. 1116–1128, May 2011.

[27] I. Jolliffe, Principal Component Analysis (Springer Series in Statistics),2nd ed. New York: Springer, 2002.

[28] J. Butzer, A. Butler, P. Butler, P. Bones, N. Cook, and L. Tlustos, “Medipiximaging—Evaluation of datasets with PCA,” Image and Vision Comput-ing New Zealand, IVCNZ 2008, pp. 1–6, 2008.

[29] C. Kamphuis and F. Beekman, “Accelerated iterative transmission CTreconstruction using an ordered subsets convex algorithm,” IEEE Trans.Med. Imag., vol. 17, no. 6, pp. 1101–1105, Dec. 1998.

[30] A. Beck and M. Teboulle, “A fast iterative shrinkage-thresholding algo-rithm for linear inverse problems,” SIAM J. Imag. Sci., vol. 2, pp. 183–202,2009.

[31] J. Butzer, N. Anderson, R. Tipples, N. Cook, T. Melzer, and P. Butler,“Bio-medical X-ray imaging with spectroscopic pixel detectors,” NuclearInstruments and Methods in Physics Research A, vol. 591, no. 1, pp. 141–146, 2008.

[32] C. Raven, “Numerical removal of ring artifacts in microtomography,” Rev.Sci. Instrum., vol. 69, pp. 2978–2980, 1998.

[33] E. Kalemci and J. Matteson, “Investigation of charge sharing among elec-trode strips for a CdZnTe detector,” Nuclear Instrum. Methods Phys. Res.Section A: Accelerators, Spectrometers, Detectors Associated Equipment,vol. 478, pp. 527–537, 2002.

Authors’ photographs and biographies not available at the time of publication.

ieee transactions on biomedical engineering, vol. 59, no. 6, june

Documents