cbct_sdarticle

8/8/2019 cbct_sdarticle

http://slidepdf.com/reader/full/cbctsdarticle 1/9

doi:10.1016/j.ijrobp.2006.09.045

PHYSICS CONTRIBUTION

REDUCING METAL ARTIFACTS IN CONE-BEAM CT IMAGES BYPREPROCESSING PROJECTION DATA

YONGBIN ZHANG , M.S.,* L IFEI ZHANG , PH.D.,* X. RONALD ZHU, PH.D.,*ANDREW K. LEE, M.D., M.P.H., † MARK CHAMBERS , D.M.D., M.S., ‡

AND LEI DONG , PH.D.**Departments of Radiation Physics, † Radiation Oncology, and ‡ Dental Oncology, The University of Texas M. D. Anderson Cancer

Center, Houston, TX

Purpose: Computed tomography (CT) streak artifacts caused by metallic implants remain a challenge for theautomatic processing of image data. The impact of metal artifacts in the soft-tissue region is magnied incone-beam CT (CBCT), because the soft-tissue contrast is usually lower in CBCT images. The goal of this studywas to develop an effective ofine processing technique to minimize the effect.

Methods and Materials: The geometry calibration cue of the CBCT system was used to track the position of themetal object in projection views. The three-dimensional (3D) representation of the object can be established fromonly two user-selected viewing angles. The position of the shadowed region in other views can be tracked byprojecting the 3D coordinates of the object. Automatic image segmentation was used followed by a Laplaciandiffusion method to replace the pixels inside the metal object with the boundary pixels. The modied projectiondata were then used to reconstruct a new CBCT image. The procedure was tested in phantoms, prostate cancerpatients with implanted gold markers and metal prosthesis, and a head-and-neck patient with dental amalgamin the teeth.Results: Both phantom and patient studies demonstrated that the procedure was able to minimize the metalartifacts. Soft-tissue visibility was improved near or away from the metal object. The processing time was 1–2 sper projection.Conclusion: We have implemented an effective metal artifact-suppressing algorithm to improve the quality of CBCT images. © 2007 Elsevier Inc.

Metal artifacts, Cone-beam CT, Image-guided radiotherapy, Image reconstruction, Image-guided radiationtherapy.

INTRODUCTION

On-board cone-beam computed tomography (CBCT) hasbeen recently introduced into the clinic for image-guidedradiation therapy applications ( 1– 4). Artifacts caused bymetallic implants, such as ducial markers, dental llings,hip prostheses, and brachytherapy seeds, remain a challengefor the automatic processing of image data for image-guidedprocedures or accurate dose calculations. Metallic implantshave much higher attenuation coefcients than bones or softtissues. The X-rays were heavily attenuated after passingthrough metal objects, resulting in only very weak signalsreaching the detector. In this situation, if the X-ray detectorlacks a sufcient dynamic range in detecting the weak signal, there will be metal shadows in the raw projectiondata. These metal shadows will introduce streak artifacts,which can spread to nearby soft-tissue regions in the recon-structed CBCT images. Another possible source of the

streak artifact comes from the beam hardening effect, whichwas caused by the nonlinear attenuation of X-ray spectrum(5, 6). These streak artifacts will change the soft-tissuevisibility and may potentially affect the accuracy of soft-tissue target delineation. In CBCT, because a larger at-panel detector is used, the scatter radiation can degrade thequality of CBCT images. As a result, CBCT images usuallyhave a poor low-contrast resolution when compared withthe conventional computed tomography (CT) scanner usingnarrowly collimated X-rays. Therefore, the effect of streak metal artifacts is magnied in CBCT images because of thedifculty in detecting soft-tissue boundaries.

Over the years, researchers have proposed many algo-rithms to reduce the metal artifacts in CT images, with mostreports in the literature dealing with metal artifacts in con-ventional CT ( 7–20). Only until recently, research on CBCTartifact reduction caught enough attention owing to theincreasing popularity of CBCT in image-guided therapeutic

Reprint requests to: Lei Dong, Ph.D., Department of RadiationPhysics, Unit 94, The University of Texas M. D. Anderson CancerCenter, 1515 Holcombe Boulevard, Houston, TX 77030. Tel: (713)563-2544; Fax: (713) 563-2545; E-mail: [email protected] Acknowledgment —We would like to thank Michael Worley in the

Department of Scientic Publications for reviewing our manu-script.

Conict of interest: none.Received May 11, 2006, and in revised form Sept 23, 2006.

Accepted for publication Sept 24, 2006.

Int. J. Radiation Oncology Biol. Phys., Vol. 67, No. 3, pp. 924–932, 2007Copyright © 2007 Elsevier Inc.

Printed in the USA. All rights reserved0360-3016/07/$–see front matter

924



procedures ( 1, 4, 21–26 ). Cone-beam CT is a volumetricimaging technique, and its detector collects multiple two-dimensional (2D) projection images at various viewing an-gles to construct the three-dimensional (3D) image mathe-matically.

The main goal of this study is to explore a simple yeteffective method to minimize the effect of metal artifacts in

CBCT images. We used a direct projection modicationmethod to minimize the metal image artifact by takingadvantage of the CBCT geometry cue. Our algorithm re-constructs the 3D voxels from as few as two user-annotatedprojections.

METHODS AND MATERIALS

In our approach, the user is required to select two projectionsand annotate the positions of the metal shadows at the beginning.A graphical user interface was designed that allows the user toselect the target position with a click of the mouse button. Forsmall metallic implants such as ducial markers, only the centroidof the metal shadows needs to be identied. For larger metalobjects with irregular shapes, a polygon dened by the corners of the object is used to represent the boundary of the object. Ouralgorithm consists of three essential steps: ( 1) 3D voxel trackingfrom two projections; ( 2) metal shadow segmentation; and ( 3)metal shadow replacement with boundary values.

3D voxel reconstruction and reprojectionWe started with two projections from two known viewing

angles separated by , assuming that the geometry calibration of the CBCT system was known. In particular, the location of theisocenter (the center of gantry rotation) should be known. In these

two viewing angles, the metal shadows caused by metallic objectsneed to be identied manually. After this is performed, the posi-tions of the metal shadows in all other views can be calculatedautomatically using the known relationship of the CBCT recon-struction geometry. Figure 1 illustrates an ideal geometry model of the CBCT system ( 27). The system consists of an X-ray source anda detector connected by a gantry arm that rotates around theisocenter O 0,0,0 . To reconstruct a 3D voxel X X , Y , Z , thecalibration parameters should be known as a priori . In our nota-tion, d and D denote source-to-axis distance (SAD) and source-to-detector distance (SDD), respectively; ud ,vd is the offset of thecone-beam projection onto the detector plane; and s x,s y representsthe scaling factor of projections on the detector. For ease of

notation, the isocenter is always dened as the origin of the worldcoordinate system (under current CBCT conguration). Mathemat-ically, it is convenient to use a homogeneous coordinate system torepresent the perspective projection of coordinate vectors. Thisrepresentation can be easily obtained by augmenting an additionalcoordinate equal to one to the original vectors of the Euclideancoordinate. Formally, let x̂ 1 x1, z1,1 and x̂ 2 x2, z2,1 be thecorresponding projections on the detector plane from two viewsseparated by , generated by the cone beam source passing throughan unknown 3D voxel X X , Y , Z , 1 in the projection space.Though it is straightforward to rotate the source S as shown in Fig.1, it is equivalent to x the source and rotate the object in theopposite direction. In this manner, the position of cone-beam

source S remains xed from any viewing angle. Given the cali-bration of the cone-beam system and rotation of the object, the 3

4 projection matrix P can be formulated by the multiplication of three matrices, representing geometry calibration, perspective projection, and rigid transformation, given as follows:

P M T (1)

where 3 3 matrix M encodes the CBCT geometry calibration,i.e. , scaling and offset of the detector of the CBCT system. It isgiven by

M

s x 0 ud

0 s y vd

0 0 1

(2)

is a 3 4 matrix describing the perspective transformation thatrelates cone beam source, the object, and its projection on thedetector, formulated by

D 0 0 0

0 0 d 0

0 1 0 0

(3)

and 4 4 matrix T describes the rigid transformation of theobject, given by

Fig. 1. Illustration of cone-beam geometry. The cone-beam sourceand detector rotate around the isocenter to collect attenuated X-raybeams. x̂ 1 and x̂ 2 are the perspective projections of three-dimen-sional voxel X from two viewing angles separated by .

925Reducing metal artifacts in cone-beam CT images ● Y. Z HANG et al.



T

cos sin 0 0

sin cos 0 d

0 0 1 0

0 0 0 1

(4)

where represents the angle of rotation around the axis throughthe isocenter. We should note here that we consider only therotation around the z axis. In practice, there may be out-of-planerotations such as the pitch of the couch, although they can beeliminated by a rigid afne registration and resampling processafter the reconstruction.

Once the projection matrix is given, the perspective projectionequations that relate the voxel X and its projections from twoviewing angles can be obtained. Without loss of generality, we canspecify the rotation angle of the projection x̂ 1 as 0, and that of projection x̂ 2 as . Therefore, a total of four projection equationscan be formed with three unknown parameters, given as:

x̂ 1 M T0 X

x̂ 2 M T X (5)

where “ ” means that the left and right are equal up to a scalingfactor; the nal vectors of 2D projections should be transformedback to the Euclidean coordinate system by dividing all the ele-ments by the third one. Equation 5 has mathematically veried thatat least two projections from different viewing angles are requiredto reconstruct the 3D voxel of the object. This is an overcon-strained linear system; a closed-form solution does not exist. Alinear least-square method was used to nd the optimal reconstruc-tion of 3D voxel X . Once X is obtained, the location of metalshadows in any view angle can be obtained by reprojection of X using the equation:

x̂ M T X * (6)

We would like to express a word of caution regarding thestability of the least-square solution. While, in theory, any twoprojections with angle difference 0 can be used to reconstruct3D voxels, a too-small can result in a singularity problem,leading to poor numerical stability in the solution. Therefore, it isalways a good practice for the user to keep in mind that shouldbe large enough, and preferably at orthogonal angles so that thevalues x̂ 1 and x̂ 2 will not be too close. Two examples from ourexperimental results are displayed in Fig. 2. The rst example (thetop row) shows the reconstruction and reprojection of three inten-tionally inserted metal screws in a phantom. In this case, the userrst annotates the centroids of two projections of the screws, andthe positions of those of all the other viewing angles are trackedusing Eq. 6. Here, the centroid alone will be sufcient to localizethe metal shadows in the projections, because its size is reasonablysmall and the shape is regular. The bottom row is another examplewith large, irregularly shaped dental llings. In this case, a polygondened by four corners on the projected region boundary is usedfor reconstruction and reprojection.

Metal shadow segmentationAfter the location of the metallic object is identied in the 2D

projection, segmentation of the object can be performed locally,instead of using the entire 2D projection image. The size of theimage patch should be determined to be large enough to containthe region of the metal shadow, while small enough to discriminatebetween the metal shadow and its surrounding pixels. This usuallywill result in a bimodal-like image histogram to represent thepixels in the foreground (metal object) and the background (sur-rounding materials). Therefore, a simple thresholding method fol-lowed by a binary postprocessing would sufce in our applicationbecause of the high attenuation of the foreground (metal) object. Inour implementation, the algorithm proposed by Otsu ( 28) was used

to nd the optimal threshold from the histogram of the imagepatch. In this algorithm, the optimal threshold in the histogram wascalculated by minimizing the intra-class intensity variance, whilemaximizing between-class intensity variance. After the threshold-

Fig. 2. Shape representation of the gaps in the projection. Left: the screw represented by its centroid. Right: a large areaof dental llings dened by its four boundary points.

926 I. J. Radiation Oncology ● Biology ● Physics Volume 67, Number 3, 2007



ing was performed, a morphologic dilation process was applied tothe binary image to remove small “holes” and smooth zigzagboundaries. Figure 3 illustrates this procedure. Finally, a binarymask was created to represent the shape of the metal object in eachprojection for image interpolation in the next step. The binarymask was dened to have the value “1” inside the region of themetal shadow and “0” elsewhere. This is similar to the conceptused by other authors ( 29–31 ).

A variational method to ll in the metal shadowsAfter the binary masks of the metal shadows on the projections

are obtained, the next step is to interpolate those metal shadowsusing the surrounding pixels. Though many techniques have beenused in previous studies, e.g. , linear interpolation ( 7, 13), polyno-mial (10), and Taylor approximation ( 6), we chose to use adiffusion method to achieve the goal of replacing the pixels in themetal region with the pixel intensities from the surrounding neigh-borhood. A variational cost function can be dened in this math-ematical framework. By minimizing the variational problem, adiffusion lter was derived to interpolate the region using bound-ary pixels. Mathematically, let the function I(x, y) denote the 2Dimage dened on a rectangle R of 2, and let be the subset of R where the region of the metal shadow is located. To modify theintensity of I over , our goal is to nd a piece-wise differentiablefunction G(x, y) dened on R, which interpolates the intensity of I over and satises G R ⁄ I R ⁄ . In our implementation, wehave chosen to minimize the magnitude of the image gradientinside the region of the metal shadow; the variational cost functionis given by

G * argmin G* G( x, y) 2d

subject to G( ) I ( ).(7)

where stands for the boundary of and . x y is the gradient operator. The above function is minimized

when G* satises the Euler-Lagrange equation:

G 0

G I .(8)

where stands for the Laplacian operator. Equation 8 is thewell-known Laplacian equation with Dirichlet boundary condi-

tions. Physically, the process we have used here is the simulationof a membrane model, which leads to an isotropic diffusion lter.One of the major advantages of the variational principle is that itallows us to dene the cost function related to a physical modeland to solve it within a general Euler-Lagrange framework. WhenLaplacian diffusion is not appropriate for more complex structures,a more sophisticated cost function needs to be dened. Priorknowledge can be incorporated in the diffusion framework. Forexample, if an extra force is introduced in Eq. 7, the diffusionequation of Eq. 8 becomes the Poisson equation, which was usedin Perez et al. (32) for various image editing tasks.

It is important to note that continuous representation is used forease of notation. In practice, an image is dened on a 2D discrete

grid; therefore, Eq. 8 should be rst discretized accordingly andthen solved numerically. Finally, the corrected projection datawere sent back to the CBCT reconstruction engine to rebuild the3D CBCT images. An illustration of the entire process is shown inFig. 3 using a phantom with three embedded metal screws.

Data acquisitionThe proposed approach was tested using both phantom data and

clinical patient data. A rst-generation CBCT system was used toacquire the projection data and perform the 3D reconstruction(Trilogy, Varian Oncology Systems, Palo Alto, CA). All scanswere acquired using the kilovoltage X-ray mode at 125 KVp. Thespatial resolution of the projection images was 1,024 768 pixels.

The SAD and SDD of the CBCT system were set to 1,000 mm and1,500 mm, respectively, giving the scaling factor of 1.5. The axial

Fig. 3. Modifying raw projection data using segmentation and Laplacian diffusion. First, the position of the gaps in eachof the projections is tracked by using a reprojection by only two calibrated views. Second, image segmentation is

performed over a small image patch dened by the position of gaps using a histogram analysis method. Third, aLaplacian diffusion lter derived from variational minimization is used to interpolate the region of gaps dened by thebinary mask obtained by segmentation.




resolution of the reconstructed CT image is 512 512 pixels witha slice thickness at 2.5 mm.

Phantom studiesTo quantify the improvement in the proposed algorithm, we

designed an experiment to acquire two sets of CT images of thesame phantom: one with and one without a gold ducial marker inthe center of the water-equivalent solid phantom. The latter CTimage was used as the ground-truth reference CT (we denote it asthe “uncorrupted” CT) for comparison. The gold marker was acylinder, 1.2 mm diameter by 3 mm long, which was commonlyused for prostate implants (Northwest Medical Physics Equipment,Seattle, WA). There were 661 raw projections. The algorithm wasapplied on the raw projection data acquired for the phantom withthe embedded gold marker. The modied project data were sentback to the CBCT system to reconstruct a new CT image (the“corrected” CT). The original “uncorrected”, the “corrected”, andthe “uncorrupted” CT images without the gold marker are shownin Fig. 4.

To quantify the improvement in CT image quality for the nearbyregions adjacent to the implanted gold marker, we calculated thehistogram and the root mean square (RMS) to characterize therecovery of the CT numbers after correction for the three CTimages in Fig. 4. The RMS is dened as follows:

RMS i, j( I 1(i, j)) ( I 2(i, j))2

N (9)

where I 1 and I 2 stand for the two images to be compared, and N is thenumber of voxels in the region of interest. Four regions were selected,where region 1 included the entire marker, region 3 and region 4corresponded to two nearby areas that were highly affected by thestreak artifacts, and region 2 was an area that was least affected.

RESULTS

Phantom studiesThe histograms of region 3 are displayed in Fig. 4 . The

histogram of the corrected image is less spread out than thatof the uncorrected one, indicating an improved uniformityin the region, similar to the uncorrupted CT image in thisphantom experiment. The RMS values between the uncor-rected images and the uncorrupted images were 194.43,201.05, 194.20, and 194.90 for region 1, 2, 3, and 4,respectively. After correction, the RMS values were greatlyreduced to 10.25, 6.50, 7.55, and 6.24, respectively for thesame regions. These results indicate a signicant recoveryof the original pixel values using the proposed correctionalgorithm.

Patient studiesA clinical case is shown in Fig. 5 for a prostate cancer

patient with three gold markers inserted for daily setup. Inthis case, the streak artifacts were almost completely re-moved while preserving the soft-tissue contrast in thenearby region. It is also interesting to observe that the subtlering artifact caused by imperfect CBCT detector calibrationin this version of the CBCT software was also restoredfaithfully near the center of the prostate in Fig. 5, whichfurther demonstrates the effectiveness of our metal artifactremoval algorithm. The size of the patient (who weighedmore than 320 pounds) caused inadequate exposure for thereconstructed CBCT image, which made the metal artifacteven more distinct.

Metal prosthesis is another example of the source of severe metal artifacts in kilovoltage-based CT images,

Fig. 4. Comparative results for a phantom with a larger gold marker inserted. Left: Computed tomographic reconstruc-tion with gold marker. Middle: Computed tomographic reconstruction after metal artifact correction. Right: Originalcomputed tomographic image without metallic implants.




which can cause major difculties in the visualization of anatomy and dose calculations. In Fig. 6, we demonstratedthe application of our metal artifact removal algorithm forthis prostate cancer patient who had one (left side) hipprosthesis. Although the overall performance is not as idealas the example in Fig. 5 for small metal objects, we can stillsee the improvement in soft-tissue visibility, as indicated inthe regions pointed by the arrows. The streak artifacts in thesoft tissue were reduced, although not completely removed.In this example, the patient had a prior brachytherapy pro-

cedure. We only corrected the shadow caused by the hipprosthesis. The large number of metal seeds in the prostate(not shown in the slice in Fig. 6) made our algorithm veryineffective to identify individual seeds even for only twoprojection views.

We also tested a head-and-neck cancer patient with largeand irregular dental llings. Unlike the point-based trackingin the case of ducial markers, the irregular shape of thedental lling needs to be tracked in each projection viewwith multiple points marked at the corner of each metal

shadow. This is illustrated in Fig. 2 (right). Figure 7 showsthe results of both corrected and uncorrected CT images ina CT slice near the dental llings. The soft-tissue contrastnear the teeth was improved by using the metal artifactremoval algorithm, although a few new minor streaks wereintroduced elsewhere as a result of the imperfect segmen-tation and interpolation of pixel values in the projectionspace. The overall visibility of the soft tissue was improved,especially in the oral cavity.

DISCUSSION

Artifact-free CT images are important for radiation ther-apy for many reasons. With better quality images, physi-cians will have an easier time and will be more consistent indelineating target structures. In an internal nine-physicianinterobserver variation study, we demonstrated that a largercontouring variation occurred among physicians in CT im-ages with the presence of dental-lling image artifacts ( 33).In addition, inaccurate Hounseld numbers can introduce

Fig. 5. Comparative results for a prostate cancer patient’s cone-beam computed tomography images. Top panels:Original computed tomographic images of three slices showing the implanted gold markers. Bottom panels: Recon-structed computed tomographic images at the same slice location after applying our algorithm.

Fig. 6. Comparative results for a prostate patient with one hip prosthesis. Left: Uncorrected cone-beam computedtomography (CBCT) images. Right: Reconstructed CBCT image after applying our algorithm. The soft-tissue contrastnear the left prosthesis was improved, as indicated by the arrow. Overall, the streak artifacts in the corrected CBCT(right) were less severe than in the original CBCT (left).




incorrect assessment of electron density, thus resulting inerroneous dose calculations. Some authors ( 34, 35) haveconducted dosimetric studies on how inaccurate CT num-

bers affect dose calculations. More recently, Bazalova et al.(36) quantitatively studied the impact of streaking metalartifacts in CT images on Monte Carlo dose calculation, andcompared that of using CT images after correction. AnAmerican Association of Physicists in Medicine task groupreport stated that CT streak artifacts can be locally signi-cant. Dose normalization in the artifact-affected regionshould be avoided ( 37). Therefore, it is important to payattention to the CT image artifacts when performing accu-rate dose calculations. Even for artifact-corrected CT im-ages using our algorithm, there were still residual artifacts(such as the ones shown in Fig. 7). For image-guided

radiotherapy, the quality of the CBCT image is also impor-tant. The reliability of the soft-tissue-based, computer-as-sisted alignment software can be reduced in the presence of metal artifacts if the alignment algorithm relies on the imageintensity information to calculate setup shifts ( 38–40).

Current literature on CT metal artifact reduction can bedivided into two categories: using the iterative method andusing the projection modication method. The key idea of the iterative method lies in the reconstruction of the CTimage using only those noncorrupted projections while dis-carding those projections affected by the metal objects. Theiterative deblurring algorithm was rst proposed by Wang et

al. (17) for reconstructing CT images in the presence of metallic implants. This method was further extended by

various researchers ( 11, 14, 16, 18, 20, 41 ). Compared withthe traditional ltered back-projection algorithm, iterativemethods have been shown to be more robust in dealing with

incomplete projections caused by metallic implants. How-ever, being an “expectation-maximization” type algorithm,the iterative method can achieve only a suboptimal solution,and its convergence is sensitive to the initialization condi-tions. Moreover, it is computationally expensive for clinicalapplications, although there were fast implementations(16, 18).

The projection modication method has been increas-ingly favored in recent years because of its simplicity. In theprojection modication approach, the metal shadows in theraw projection data caused by the X-ray passing through themetallic implants are rst segmented and then replaced

using some estimated values. Kalender et al. (8) and Klotzet al. (9) manually identied the missing projections andreplaced them by interpolation using noncorrupted neighborprojections. Rajgopal et al. (13) used a linear predictionmethod to replace the missing projections, and Lewitt andBates (10) employed a polynomial interpolation techniqueto bridge the missing projections. A multiresolution waveletanalysis of projections was also proposed for missing regiondetection and interpolation ( 20). Chen et al. (7) used a linearinterpolation followed by an adaptive scaling and lteringprocess on the CT reconstructions to mitigate the metallicartifacts. Instead of performing interpolation, Olive et al.

(12) proposed to use model data derived from the originaltomogram to ll in the metal shadows based on an image

Fig. 7. Comparative results for a head-and-neck patient’s cone-beam computed tomography (CBCT) image with largeareas of irregular dental llings. Top panel: Original uncorrected data. Bottom panel: Modied projection view and thereconstructed CBCT images. Left column: One of the projection views without (top) and with (bottom) the correctionmethod described in this work. The arrows indicate locations of the dental llings to be corrected. Middle and rightcolumns: Reconstruction CBCT images in two slice positions with and without correcting for metal artifacts using thetracked and interpolated projection data. The soft-tissue visibility in the oral cavity was improved, as indicated by thearrow.




segmentation process. Adaptive ltering was also used inthis approach for postprocessing of the reconstructed image.More recently, Yazdia et al. (19) proposed an approach formetallic artifact reduction in conventional CT for clinicaluse in radiation therapy treatment planning. In their ap-proach, the 3D positions of the metal implants were rstdetected in the reconstructed CT, and then they were re-

projected to localize the metal shadow region in each of theraw projection data. An adaptive interpolation scheme wasproposed to preserve the structure of adjacent projections. InCBCT applications, Moseley et al. (6) used a segmentationand then reprojection method to reduce artifacts. In thismethod, the metal shadow regions in each CBCT projectionwere segmented rst by detecting metal implants in theoriginal 3D CT data set and then reprojecting them back into the projection space. The second-order Taylor serieswere employed to interpolate the missing regions.

Though the projection modication method does not re-quire expensive computation, a major challenge lies in the

accurate detection of metal shadows in the raw projectiondata. Many have suggested the manual identication of allthe metal shadows in all of the projections ( 8, 9). Unfortu-nately, hundreds of projections are generally used to recon-struct a CBCT image. Fully manual segmentation of metalshadows such as in Kalender et al. (8) and Klotz et al. (9)was shown to be impractical for clinical applications. Themethod used in Yazdia et al. (19) and Moseley et al. (6)attempted to directly delineate the metallic implants fromthe original CT images (containing the metal image arti-facts), which may not be easy because the exact location andshape of the true metal object may not be known.

In our approach, we used a direct projection modicationmethod. Users are only required to select at two views anddelineate the metal shadow on these two projection images.Then the rest of work is fully automatic. The metal shadowsare automatically tracked and subsequently segmented in allprojection views. This is an ofine correction algorithm; thetime to track and correct one projection view was approxi-mately 1 to 2 s using an Intel Pentium processor at 2.6 GHz.

Our results clearly demonstrated that the proposed algo-

rithm can suppress metal artifacts and restore the CT num-bers in nearby soft tissues. However, under certain circum-stances when large or irregularly shaped metal objectspresent, the algorithm may not perform well. This is mainlybecause the algorithm assumes that the pixel values insidethe metal shadow region can be derived from the boundarypixel values using a diffusion process. This may not be

valid. As a result, there will be imperfect corrections or newartifacts. This can be seen in Fig. 7: some new, minor streak artifacts were introduced although the overall image arti-facts were reduced. For the same reason, the performance of the metal-artifact reduction algorithm was not perfect in theprosthesis example ( Fig. 6) because of the large size of theprosthesis metal rod. The assumption of a smooth diffusingprocess across a large metal object may not be valid. TheLaplacian diffusion lter is derived from minimizing themagnitude of the gradient eld in the region of the metalshadow.

Finally, the main goal of our research was to improve the

soft-tissue contrast near the metal artifact. This wasachieved by developing an algorithm with the goal to re-move the metal object, instead of restoring the true repre-sentation of the metal object in the CT image. For applica-tions that require the knowledge of implant marker locationsfor target alignment (such as the implanted ducials), thespatial location of the ducial markers is known to oursystem (after solving Eq. 5), which can be directly used forimage-guided setup. Alternatively, we could also re-insertthe ducial image back into the CT image. A cleaner imagewould facilitate automatic image processing of the CBCTimages, allowing for fully automatic soft-tissue image reg-

istration for patient setup.In conclusion, we have designed a simple and effectivealgorithm to suppress or minimize CBCT image artifactscaused by metal implants. Preliminary results in phantomand patient cases demonstrated its effectiveness in removingthe heavy streak artifacts to the surrounding soft tissue. Theimage quality near the metal implant region was improvedsignicantly. The proposed algorithm is computationallyefcient and requires minimal user intervention.

REFERENCES

1. Jaffray DA, Siewerdsen JH, Wong JW, et al. Flat-panel cone-beam computed tomography for image-guided radiation ther-apy. Int J Radiat Oncol Biol Phys 2002;53:1337–1349.

2. McBain CA, Henry AM, Sykes J, et al. X-ray volumetricimaging in image-guided radiotherapy: The new standard inon-treatment imaging. Int J Radiat Oncol Biol Phys 2006;64:625–634.

3. Oelfke U, Tucking T, Nill S, et al. Linac-integrated kV-conebeam CT: Technical features and rst applications. Med Do-sim 2006;31:62–70.

4. Oldham M, Letourneau D, Watt L, et al. Cone-beam CTguided radiation therapy: A model for on-line application. Radiother Oncol 2005;75:271–278.

5. Joseph PM, Ruth C. A method for simultaneous correction of

spectrum hardening artifacts in CT images containing bothbone and iodine. Med Phys 1997;24:1629–1634.

6. Moseley DJ, Siewerdsen JH, Jaffray DA. High-contrast objectlocalization and removal in cone-beam CT. Progress Biomed Optics Imaging—Proc SPIE 2005;5745:40–50.

7. Chen LM, Liang Y, Sandison A , et al. A novel method forreducing high-attenuation object artifacts in CT reconstruc-tions. Proceedings of SPIE – The International Society for Optical Engineering 2002;4684:841–850.

8. Kalender WA, Hebel R, Ebersberger J. Reduction of CTartifacts caused by metallic implants. Radiology 1987;164:576–577.

9. Klotz E, Kalender WA, Sokiranski R, et al. Algorithms forreduction of CT artifacts caused by metallic implants. In:Dwyer SJ, Jost RG, editors. Proceedings of SPIE–The International Society for Optical Engineering . 1990;

1234:642–650.10. Lewitt RM, Bates RHT. Image reconstruction from projec-




tions: III. Projection completion methods. Optik 1978;50:189–204.

11. Nuyts J, De Man B, Dupont P, et al. Iterative reconstructionfor helical CT: A simulation study. Phys Med Biol 1998;43:729–737.

12. Olive CS, Eck K, Spies L, et al. Segmentation aided adaptiveltering for metal artifact reduction in radio-therapeutic CTimages. In: Fitzpatrick JM, Sonka M, editors. Proceedings of SPIE—The International Society for Optical Engineering,Vol. 5370 III. 2004. p. 1991–2002.

13. Rajgopal K, Srinivasa. N, Ramakrishnan KR. Image recon-struction from incomplete projection data: A linear predictionapproach. In: Leondes CT, editor. Medical imaging systems,techniques and applications: Modalities. New York: Gordonand Breach; 1997.

14. Robertson DD, Yuan J, Wang G, et al. Total hip prosthesismetal-artifact suppression using iterative debluring recon-struction. J Comput Assist Tomogr 1997;21:293–298.

15. Roeske JC, Lund C, Pelizzari CA, et al. Reduction of com-puted tomography metal artifacts due to the Fletcher-Suitapplicator in gynecology patients receiving intracavitarybrachytherapy. Brachytherapy 2003;2:207–214.

16. Toft P. A very fast implementation of 2D iterative reconstruc-

tion algorithms. IEEE Proc Med Imaging 1996;3:1742–1746.17. Wang G, Snyder DL, O’Sullivan JA, et al. Iterative deblurringfor CT metal artifact reduction. IEEE Trans Med Imaging1996;15:657–664.

18. Wang G, Frei T, Vannier MW. Fast iterative algorithm formetal artifact reduction in X-ray CT. Acad Radiol 2000;7:607–614.

19. Yazdia M, Gingras L, Beaulieu L. An adaptive approach tometal artifact reduction in helical computed tomography forradiation therapy treatment planning: Experimental and clin-ical studies. Int J Radiat Oncol Biol Phys 2005;62:1224–1231.

20. Zhao S, Robertson DD, Wang G, et al. X-ray CT metal artifactreduction using wavelets: An application for imaging total hip

prostheses.IEEE Trans Med Imaging

2000;19:1238–1247.21. van Herk M, Jaffray D, Betgen A, et al. First clinical experi-ence with cone-beam CT guided radiation therapy: Evaluationof dose and geometric accuracy [Abstract]. Int J Radiat Oncol Biol Phys 2004;60:S196.

22. van Herk M, Jaffray D, Sonke JJ, et al. Kilovoltage cone beamCT guided radiotherapy. Radiother Oncol 2005;76:S3.

23. Hawkins MA, Brock K, Eccles C, et al. Breath hold conebeam CT for image guided stereotactic radiotherapy for livercancer. Radiother Oncol 2006;78:S80.

24. Jaffray D. Volumetric image-guidance for therapy: Perfor-mance, quality and opportunity. Med Phys 2005;32:2067.

25. Daly M, Siewerdsen J, Moseley D, et al. Cone-beam CT forimage-guided head and neck surgery: Assessment of doseand image quality using a C-arm prototype. Med Phys2006;33:2229.

26. Yeung I, Moseley DJ, Crook J, et al. Cone beam CT basedintraoperative dose assessment in TRUS guided prostate seedimplant—geometric verication. Med Phys 2004;31:1746.

27. Feldkamp LA, Davis LC, Kress JW. Practical cone-beamalgorithm. J Optical Soc America A–Optics Image Sci Vision1984;1:612–619.

28. Otsu N. Threshold selection method from gray-level histo-grams. IEEE Trans Systems Man Cybernetics 1979;9:62–66.

29. Sijbers J, Postnovz A. Reduction of ring artefacts in highresolution micro-CT reconstructions. Phys Med Biol 2004;49:N247–N253.

30. Nuyts J, Stroobants S. Reduction of attenuation correctionartifacts in PET-CT. IEEE Nuclear Science Symposium Con-ference Record, Vol. 4. 2005. p. 23–29.

31. Oiao F, Yue Y, Pan T, et al. Segmentation of contrast en-hanced CT images for attenuation correction of PET/CT data.IEEE Nuclear Science Symposium Conference Record, Vol.5. 2004. p. 2686–2689.

32. Perez P, Gangnet M, Blake A. Poisson image editing. ACM Transactions on Graphics 2003;22:313–318.

33. O’Daniel JC, Rosenthal DI, Barker JL, et al. Inter-observercontouring variations of head-and-neck anatomy. Int J Radiat Oncol Biol Phys 2005;63:S370.

34. Williams G, Tobler M, Gaffney D, et al. Dose calculationerrors due to inaccurate representation of heterogeneity cor-rection obtained from computerized tomography. Med Dosim2002;27:275–278.

35. Keall PJ, Chock LB, Jeraj R, et al. Image reconstruction andthe effect on dose calculation for hip prostheses. Med Dosim2003;28:113–117.

36. Bazalova M, Palefsky S, Beaulieu L, et al. Correction of streaking artifacts in CT images and its inuence on MonteCarlo dose calculations. Med Phys 2006;33:2158–2159.

37. Papanikolaou N, Battista JJ, Boyer AL, et al. Tissue inhomo-geneity corrections for megavoltage photon beams: Report of Task Group no. 65 of the Radiation Therapy Committee of theAmerican Association of Physicists in Medicine. New York:American Institute of Physics for the AAPM; 2004.

38. Court LE, Dong L. Automatic registration of the prostate forcomputed-tomography-guided radiotherapy. Med Phys 2003;30:2750–2757.

39. Smitsmans MHP, de Bois J, Sonke JJ, et al. Automatic pros-tate localization on cone-beam CT scans for high precisionimage-guided radiotherapy. Int J Radiat Oncol Biol Phys2005;63:975–984.

40. Smitsmans MHP, Wolthaus JWH, Artignan X, et al. Auto-matic localization of the prostate for on-line or off-line image-guided radiotherapy. Int J Radiat Oncol Biol Phys 2004;60:623–635.

41. Wang G, Vannier MW, Cheng PC. Iterative X-ray cone-beamtomography for metal artifact reduction and local region re-construction. Microsc Microanal 1999;5:58–65.


cbct_sdarticle

Documents