A Segmentation-Based Method for Metal Artifact Reduction1

Technical Report

A Segmentation-Based Method for MetalArtifact Reduction1

Hengyong Yu, PhD, Kai Zeng, MS, Deepak K. Bharkhada, MS, Ge Wang, PhD, Mark T. Madsen, PhD, Osama Saba, PhDBruno Policeni, MD, Matthew A. Howard, MD, Wendy R. K. Smoker, MS, MD, FACR

Rationale and Objectives. We propose a novel segmentation-based interpolation method to reduce the metal artifactscaused by surgical aneurysm clips.

Materials and Methods. Our method consists of five steps: coarse image reconstruction, metallic object segmentation,forward-projection, projection interpolation, and final image reconstruction. The major innovations are 2-fold. First, astate-of-the-art mean-shift technique in the computer vision field is used to improve the accuracy of the metallic objectsegmentation. Second, a feedback strategy is developed in the interpolation step to adjust the interpolated value based onthe prior knowledge that the interpolated values should not be larger than the original ones. Physical phantom and realpatient datasets are studied to evaluate the efficacy of our method.

Results. Compared to the state-of-the-art segmentation-based method designed previously, our method reduces the metalartifacts by 20–40% in terms of the standard deviation and provides more information for the assessment of soft tissuesand osseous structures surrounding the surgical clips.

Conclusion. Mean shift technique and feedback strategy can help to improve the image quality in terms of reducing metalartifacts.

Key Words. Metal artifacts; mean shift filtering; image segmentation; projection interpolation; feedback.©
AUR, 2007
In x-ray computed tomography (CT), the attenuation coef-ficient of high-density objects, such as surgical clips,metal prostheses, or dental amalgams, is much higherthan that of soft tissues and osseous structures. Becausethe x-ray beam is highly attenuated by metals, an insuffi-cient number of photons reach the detector, producing

Acad Radiol 2007; 14:495–504

1 From the Biomedical Imaging Division, VT-WFU School of Biomedical En-gineering and Science, Virginia Technical, Blacksburg, VA 24061 (H.Y.,G.W.); Biomedical Imaging Division, VT-WFU School of Biomedical Engi-neering and Science, Wake Forest University, Winston-Salem, NC 27157(D.K.B.); SIEMENS Medical Solutions, 51 Valley Stream Parkway, Malvern,PA 19355 (O.S.); and the Departments of Biomedical Engineering (K.Z.),Radiology (M.T.M., B.P., W.R.K.S.), and Neurosurgery (M.A.H.), Universityof Iowa, Iowa City, IA 52242. Received Nov 25, 2006; accepted Dec 27,2006. This work is supported by NIH/NIBIB grants EB002667 andEB004287. Address correspondence to: H.Y. and W.R.K.S. e-mail:[email protected] and [email protected]

©
AUR, 2007doi:10.1016/j.acra.2006.12.015
corrupted projection data. Consequently, images recon-structed by the traditional filtered back-projection (FBP)method are marred by starburst artifacts, often referred toas metal artifacts. These artifacts significantly degrade CTimage quality and limit the usefulness of CT for manyclinical applications because tissues in the plane of themetal appliance are severely obscured. Hence, there is animportant need for methods that reduce metal artifacts.

The effects of metallic objects on x-ray scanning are2-fold: beam hardening, due to the poly-energetic x-rayspectrum, and a poor signal-to-noise ratio from photonstarvation. To suppress the metal artifacts, iterative recon-struction methods have been successfully applied thatavoid the corrupted data. For example, Wang et al. usedthe maximum expectation maximization (EM) formulaand algebraic reconstruction technique (ART) to itera-tively deblur metallic artifact (1, 2). A key step in their
algorithm is the introduction of a projection mask and the
495

YU ET AL Academic Radiology, Vol 14, No 4, April 2007

computation of a 3D spatially varying relaxation factorthat allows compensation for beam divergence and dataincompleteness. However, this approach is computation-ally expensive and not practical for clinical imaging. Con-ventional FBP methods (3) are computationally efficientbut produce image artifacts when complete and preciseprojection data are unavailable. Different linear and poly-nomial interpolation techniques have been developed forestimating the “missing” projection data (4–11). The ma-jor task of this method is to identify the corrupted seg-ments in the sinogram and interpolate these data fromnoncorrupted neighbor projections. Because the first stepof all the above methods requires segmenting of the metalparts from a coarse image reconstructed by FBP, segmen-tation is a key technique for metal artifact reduction.

Most researchers have adopted simple threshold meth-ods to segment the metallic objects and then forward-projected them into the sinogram domain. These projec-tions serve as a mask for interpolation and reconstruction(4–6, 10). The threshold-based methods may produceinaccurate segmentation of the metallic objects and,hence, the information from structures surrounding themetal may be lost. To improve the accuracy of this tech-nique, Yazdi et al. (11) proposed to automatically detectand reconnect the edges of the metal projection in theoriginal sinogram. The idea was to apply the interpolationscheme between the two corresponding projected edgesbelonging to the projection regions of the same object.While this method works well for metallic hip prostheses(11), it is not suitable for irregular or nonconvex objects.Recently, Bal and Spies (12) developed a novel methodto reduce artifacts by in-painting missing information intothe corrupted sinogram. The missing sinogram informa-tion was derived from a tissue-classes model extractedfrom the corrupted image, and a k-means clustering tech-nique was adopted for image segmentation. However, thismethod strongly depends on the prior tissue model.

Here, we describe a novel segmentation-based interpo-lation method to suppress artifacts from metallic aneu-rysm clips. The major contributions include the use of asophisticated mean-shift technique and a restriction on theinterpolation values. The mean-shift technique is thestate-of-the-art approach for feature space analysis inthe computer vision field (13, 14). Because the mean shifttechnique is not sensitive to image noise and metal arti-facts, it can provide high segmentation accuracy. Also,because the goal of the interpolation is to improve theprojection data affected by the metallic objects, prior in-
formation is used to ensure that the interpolated values
496

are consistent to the normal data, especially when multi-ple metallic objects are in the field of view. This is ac-complished with a feedback strategy in the interpolationstep to adjust the interpolated values automatically.

METHODS

Algorithm DescriptionCT scans were performed on a Siemens SOMATOM

Sensation 16 scanner using helical scanning geometry.After the dataset was acquired, we applied the single-slicerebinning method (15) to convert the multislice dataset toa stack of fan-beam sinograms, each associated with onehorizontal z-slice. Once the fan-beam sinograms weregenerated, metal artifact reduction and image reconstruc-tion were performed for each z-slice in the 2D fan-beamgeometry as illustrated in Figure 1. The fan-beam sino-gram can be represented as PO(�,�), where � is an anglethat indicates the position of the x-ray source, � is thefan-angle that refers to detector location, and the subscriptO represents the original dataset. For a full scan datasetfrom the Siemens SOMATOM Sensation 16 scanner, � isdiscretized as �i with i � 1,2,. . .1160, and � is dis-cretized as �j with j � 1,2,. . .672. Our improved segmen-tation-based method for metal artifact reduction has stepsthat are similar to state-of-the-art methods (10). As shownin Figure 2, our method can be summarized into fivesteps for each fan-beam projection dataset.

Step 1: Coarse image reconstructionFor a specified region of interest (ROI) in the field of

view (FOV), a coarse image IO(m,n) was reconstructed bythe conventional full-scan FBP algorithm, where m �

Figure 1. Fan-beam geometry of the Siemens SOMATOM Sen-sation 16 scanner associated with the rotational angular �

1,2,. . .M and n � 1,2,. . .N. Because the ROI is fixed, the

etho

Academic Radiology, Vol 14, No 4, April 2007 SEGMENTATION-BASED METHOD

higher the M and N, the smaller the pixel size. For theselection of M and N, our general rule was to make eachpixel represent a cubic region in the final 3D volume image.

Step 2: Metallic object segmentationUsing the state-of-the-art computer vision mean-shift

technique (13, 14), we segmented the metallic objectsfrom the coarse image IO and obtained a characteristicimage IC, defined as

IC (m, n) ��1 if IO (m, n) belongs to metallic objects

0 otherwise,

(1)

where the subscript “C” denotes a characteristic function.IC functions as an index for the metallic objects in thespecified ROI. In the next subsection, we describe thisprocedure in detail.

Step 3: Forward-projectionIn the same fan-beam geometry, we forward-projected

the characteristic image IC into the sinogram domain andobtained PM (�i ,�j), which represents the intersectionallength of the corresponding x-ray path with the metallicobjects. We defined characteristic projection by

PC (�i, � j) ��1 PM (�i, � j) � 0

0 otherwise. (2)

PC functions as an index to specify which pixel in theoriginal sinogram had been corrupted by the high-inten-sity metallic objects.

Step 4: Projection interpolationThe corrupted projection data were deleted from the

original projection PO. From the feedback interpolationstrategy (see later subsection for detail), we obtained a

Figure 2. Flowchart of the proposed m

new interpolated projection PI, where the subscript I rep-

resents interpolation. A difference projection dataset PD

was obtained by

PD (�i, � j) � PO (�i, � j) � PI (�i, � j). (3)

Obviously, PD was completely formed by the metallicobjects.

Step 5: Final image reconstructionA background image, IB, and a metallic image, IM,

were reconstructed by the traditional FBP method fromthe interpolated projection, PI and difference projectionPD, respectively. The final image was composed by ascale scheme (16) as

IF (m, n) � IB (m, n) � � � IC∗ (m, n) � IM (m, n), (4)

where � is the scale factor, ranging from 0.05 to 0.5, andIC

� functions as a mask to protect the background imagefrom corruption by the metallic artifacts. To smooth theedge and preserve the structure of the metallic objects,IC

� is defined as the 2D convolution of the characteristicimage IC and a normalized Gaussian kernel.

Metallic Object SegmentationThe mean shift technique was first proposed in 1975

by Fukunaga and Hostetler (17) and largely forgotten un-til Cheng’s work rekindled interest in it in 1995 (13).Based on the mean shift technique, in 2002, Comaniciuand Meer proposed a nonparametric method for analysisof a complex multimodal feature space and to delineatearbitrarily shaped clusters within it (14). The techniquewas successfully applied for discontinuity, preservingsmoothing and image segmentation, and raised a hot topicin the computer vision field. Here, we use this techniqueto segment the metallic objects from the coarse image IO.

Let xk, k � 1,2,. . .K, be a point in one finite set X (the

d for clip artifact reduction.

“data” or “sample”) in the d-dimensional space �d. Let G

497


be a kernel and x ¡ (0, �) be a map W : x ¡ (0, �).The sample mean with kernel G at x is defined as (13),

MS (x) ��k�1

K

G(xk � x)W(xk)xk

�k�1

K

G(xk � x)W(xk). (5)

Let T � X be a finite subset. The evolution of T inthe form of iterations T ¢ Ms(T) with Ms(T) � {Ms(t);t� T} is called a mean shift procedure, and the differenceMs(t)�t is called mean shift. For each iteration, t ¢ Ms(t)is performed for all the t � T simultaneously. For eacht � T, there is a sequence, t,Ms (t),Ms (Ms (t)),&. . ., thatis called the trajectory of t. The weight function W(xk)can be either fixed throughout the process or re-evaluatedafter each iteration. The algorithm halts when it reachesfixed points Ms (T) � T.

For our application, the 2D gray-level image can beregarded as the 3D sampling points, that is, d �3. Thespace of the 2D coordinate lattice is known as the spatialdomain while the gray-level is called the range domain.The joint 3D spatial-range vector set of the coarse imageIO is used to form the finite set X and T. Correspond-ingly, the size of the set is the total pixel number of IO,that is K � M � N. As pointed out by Comaniciu andMeer (14), an Epanechnikov kernel or a truncated normalkernel always provides satisfactory performance. Hence,we selected the multivariate Epanechnikov kernel, whichis defined as the product of two radially symmetric ker-nels and the Euclidean metric allowing a single band-width parameter for each domain (14),

G(x) �C

hs2hr

g ��xs ⁄ hs�2�g ��xr ⁄ hr�

2�, (6)

with the common profile

g (s) ��1 s 1

0 otherwise. (7)

In Equation 6, xs is the spatial component, xr is therange component, hs and hr are the kernel bandwidth pa-rameters, and C is the corresponding normalized constant.The weighting function is fixed throughout all iterations.To construct the weighting W, we first computed the gra-
dient image IG of IO, which is defined as
498

IG (m, n) � �IO (m, n)

��IO (m, n)

�m �2

� ��IO (m, n)

�n �2

. (8)

Because the mean shift procedure preserves disconti-nuities, a small weight was assigned to a pixel in discon-tinuous regions, while a large weight was assigned to thatin flat regions. Hence, the weighting function W was con-structed as:

W(m, n) � 1 � IG (m, n) ⁄ max(IG), (9)

where max(IG) is the maximum of the gradient image IG

and 0 � � � 1 is a scale factor. As a result, all the nec-essary components have been determined for the meanshift technique.

Let xk and zk be the 3D input and output points in thejoint spatial-range domain with k � 1,2,. . .K. Based onthe work of Comaniciu and Meer (14), the mean shiftsmoothing filtering for each point can be performed as:

1. Initialize l � 1 and yk, l � xk ;2. Compute yk, l �1 � Ms(yk, l) according to Eq.(5) until

converging to the final result yk, L ;3. Assign the output zk � (xk

s,yk, Lr ).

After the coarse image has been smoothed by preserv-ing the discontinuity, the metallic object segmentation canbe implemented by the grouping method proposed in sub-section 4.2.1 in Comaniciu and Meer (14). Noting the factthat metallic objects have higher attenuation than humantissues, the simple threshold method can also be em-ployed after mean shift smoothing filtering. For briefness,we used the latter in our study. To compensate for theblurring of the projection introduced by the focal spotof the x-ray source, a morphological dilation was per-formed to the above segmentation result. Finally, thecharacteristic image Ic was obtained. Figure 3 presentssome typically interim images to illustrate the meanshift technique.

We make the following three comments on the imple-mentation of the mean shift filtering. First, the constant Cin Equation 6 can be omitted since it has been canceledin Equation 5. For the same reason, W can also be de-fined as W �m,n� � max �IG� � IG �m,n�. Second, many
techniques can be used to reduce the computational time


required for the mean shift technique (14). Third, the finalsegmentation is not sensitive to the selection of thethreshold because the mean shift filtering sharpens theedges of the metallic objects.

Feedback-Based Interpolation StrategyAfter the characteristic projection function PC has been

determined in the third step of the proposed method, wedelete the corrupted data and interpolate them via linearor polynomial interpolation in the fourth step. It is wellknown that a metallic object corrupts a strip-like region inthe projection domain. When there is only one metallicobject in the field of view, it is reasonable to interpolatethe missing corrupted data. However, when there are mul-tiple metallic objects inside the FOV, the reliability of theinterpolated value is lower, especially when it is near the

Figure 3. Illustration of metallic object segmThe original coarse image, (b) weighting funcfinal metallic object characteristic function.

overlap region of multiple strips. To improve the interpo-

lation accuracy, we propose a feedback strategy for theinterpolation based on prior information.

Assume that the original projection data are PO (�i,�j)and interpolated value is PI (�i,�j) . Because the interpola-tion is performed only for the region in which the projec-tions have been corrupted by the high-density metallicobjects, the difference projection PD (�i,�j) should be notbe smaller than zero. That is, the interpolated valuePI (�i,�j) should not be larger than PO (�i,�j). This factprovides the prior information in our feedback strategy.The feedback strategy can be described by the followingprocedure:

1. Perform the interpolation for the missing corrupteddataset indexed by PC ;

2. Compare the original and interpolated values and

tion using the mean shift technique. (a)(c) image after mean shift filtering and (d)

entation,

set PC (�i,�j) � 0 if PI (�i,�j) � PO (�i,�j);

499


3. Repeat (i) and (ii) until all the interpolated valuesare no larger than the original ones.

In our study, the linear interpolation was adopted sincethe strips of metallic clips in the projection domain arevery thin. For more complex applications, our feedbackscheme can be used directly without any modification.Figure 4 shows two representative interpolated profileswith and without the feedback strategy.

RESULTS

Algorithm ImplementationWe set up a platform to reconstruct and display images

from data collected along a helical scan. First, the sino-gram datasets were rebinned from spiral cone-beam ge-ometry into fan-beam geometry. Then, images werereconstructed using a conventional fan-beam FBP algo-rithm. Between these two steps, we inserted some addi-tional processing steps proposed in this paper to reducethe metal artifacts. All the reconstructed images wereconverted into the standard Hounsfield units (HU) by as-suming the attenuation coefficient of water was 0.019/mm. The proposed method was tested by both a physicalclip phantom and clinical patient datasets. The corre-sponding parameters were summarized in Table 1.

As a benchmark, the segmentation-based interpolationmethod proposed by Wei et al. (10) was also imple-mented. There are three major differences between Wei etal.’s method and ours. First, Wei et al. used an averagefiltering plus the threshold method to segment the metallicobjects, while we adopted the mean shift filtering plusthreshold method and morphological dilation. Second,

Figure 4. Two representative interpolated pbased interpolation strategy.

Wei et al. used the polynomial interpolation for the miss-

500

ing values, while we used linear interpolation along withthe feedback scheme. Third, Wei et al. generated the finalimage with the mask IC (see Equation 4), while we sub-stituted IC with IC

� to smooth the edge and preserve thestructure of the metallic objects. For comparison, both thethreshold for metallic segmentation and the scale factor �

in the last step were set to be the same values for bothmethods.

Clip Phantom ExperimentThe phantom was a water-filled plastic cylinder with

an inside diameter of 21.6 cm and a length of 18.6 cmwith a wall thickness of 3.2 mm. An 8-mm plastic rodwas located in the center of the phantom, and five differ-ent aneurysm clips were attached with rubber bands to therod. The axial separation between clips was at least 2 cm.The phantom was scanned with clinical brain CTA pa-rameters: 120 kVp and 496 mAs. Figure 5 shows a typi-

Table 1Parameter Selection for our Phantom and Patient Datasets

PhantomDataset

PatientDataset

Pixel size (mm2) 0.135 � 0.135 0.164 � 0.164Spatial domain bandwidth hs (mm) 0.33 0.82Range domain bandwidth hr (HU) 1000 1500Coefficient for weighting

function 0.5 0.5Threshold for metallic

segmentation (HU) 1500 3500Scale factor for final

reconstruction � 0.15 0.15

s with and without usimg the feedback-
rofile
cal phantom slice reconstructed by the proposed method

ows


and Wei et al.’s segmentation-based interpolation method(10). From the local magnified images, it is observed thatthe bright artifacts around the clip are better suppressedby our method compared to Wei et al.’s. This resultsfrom the more accurate metallic object segmentation, pro-

Figure 5. Representative images of the clip phantom recwere directly reconstructed from the fan-beam sinogram wrected using the conventional segmentation-based interpoing our proposed method. The middle row images are theare the characteristic functions for quantitative measuremtween (e) and (f) are indicated by arrows. The display wind

vided by the mean shift filtering, and the feedback strat-

egy that provides improved correction of the interpolatedvalues. Note the intensities of artifacts surrounding themetal vary significantly more than elsewhere, and the tis-sues are of major diagnostic interest, a standard deviation(SD) is utilized to quantify the improvement of the pro-

ructed using different methods. The left column imagesut any correction. The middle column images were cor-n method. The right column images were corrected us-l magnification of the top row. The bottom row imagesf the middle row images. The significant differences be-are [�100,300] HU.

onstitholatioloca

ent o

posed method,

501


D ��m, n

(IF (m, n) � T)2 � DC (m, n)

�m, n

DC (m, n)(10)

where DC is a characteristic function indicating thesurrounding region of the metallic objects and T is theideal average gray level (HU) around the metallic objects.Let IC

� denote the 2D convolution of the characteristicimage IC and a specific kernel function, DC can be ob-tained by subtracting IC from the characteristic functionof I. Because the phantom was filled with water, we setT � 0 HU and select DC as the last row in Figure 5. TheSDs from our clip phantom images in Figure 5 are listedin Table 2.

Patient StudyUnder the approval of the institutional review board

committee, the University of Iowa, we obtained the rawdata from a clinical study at our institution. The patientwas a 59-year-old woman with a known history of hyper-tension. She presented with a “worst headache of life”and right hemibody numbness. When she was first seen inthe emergency department, her Glasgow Coma Scalescore was 10 and the World Federation of NeurologicSurgeons score was 4. A head CT scan was obtainedshowing diffuse subarachnoid hemorrhage in the basalcisterns and Sylvian fissures, left greater than right. CTangiography demonstrated a left middle cerebral arteryaneurysm. She was taken to the operation room and theaneurysm was clipped. She had numerous head CT scansafter surgery for assessment of increased intracranial pres-sure to rule out rebleeding and hydrocephalus. Figure 6presents the reconstructed images of a representative CTslice. Comparing the local magnification of the imagereconstructed by our method and the state-of-the-artmethod (10), our method provides better quality withfewer artifacts around the metallic clip. Because the me-tallic clip was surrounded by soft tissues, we set T � 30

Table 2Standard Deviations Associated with Different CorrectionMethods

WithoutCorrection

Wei et al.Method

OurMethod

Clip phantom 681.80 181.12 142.66Patent head 1759.91 330.27 186.09

HU and selected the characteristic function as in the bot-

502

tom row in Figure 6 for quantitative measurement. Thequantified results are listed in Table 2, which shows thatour method reduces the artifacts by 43% of the SD com-pared to the state-of-the-art method (10). This enabled usto visualize soft tissue structures around the metallic clipsand improve clinical diagnostic accuracy.

DISCUSSION

In the proposed metal artifact reduction method, thereare several parameters that needed to be specified empiri-cally. General speaking, the larger the bandwidth parame-ters hr and hs, the sharper is the discontinuity of the me-tallic objects. The selection of hs should be larger than thesize of a pixel. On the other hand, metallic objects willbe erased if their sizes are in the same order as hs. Fourdifferent techniques can be considered to optimize thebandwidth parameters (14). The local adaptive solutionscan also be used to solve the difficulties generated by thenarrow peaks and the tails of the underlying density (18).Because the mean shift procedure smoothes the wholeimage while preserving the discontinuity, the selection ofa threshold to segment metallic objects is more flexiblethan the conventional methods (4–6, 10). The generalrule is that the threshold should be smaller than the me-tallic gray level and larger than the maximum gray levelof other human tissues.

As an initial study, our method was implemented inMatLab on a regular PC (2.8 GHz CPU). The total com-putational time was about 10 minutes to process one512�512 slice. The forward-projection step accounted formost of this time. Currently, a ray-driven method wasadopted for the forward-projection of the characteristicfunction IC into the projection domain. In the future, weplan to use a voxel-driven method for the forward-projec-tion and only project the metallic object parts of the char-acteristic function IC. Furthermore, we may transplant thecodes into C��. Hence, the total time will be reduced inan order of magnitude, making the final time less thanone minute. If it is necessary, parallel-computing techniquescan be used for real-time or near real-time performance.

CONCLUSION

We have proposed a clinically feasible approach forsurgical aneurysm clip artifact reduction, which reduced
metal artifacts by 20–40% in terms of the standard devia-

play


tions with the soft tissues and osseous structures sur-rounding the metallic clips. Compared to the conventionalsegmentation-based methods, our method has two distin-guishing features: mean shift filtering for metallic objectsegmentation and feedback-based interpolation for projec-

Figure 6. Representative images of the patient head recwere directly reconstructed from the fan-beam sinogram wrected using the conventional segmentation-based interpoing our proposed method. The middle row images are theimages are the characteristic functions for quantitative meences between (e) and (f) are indicated by circles. The dis

tion data consistency. The method may be also applied in

other applications such as dealing with metallic prosthesesand dental amalgams. It is hypothesized that the imagingperformance in the 3D case will be similar to that in the2D case. Further efforts will be made to improve effec-tive metal artifact reduction techniques and bring them

ucted using different methods. The left column imageut any correction. The middle column images were cor-n method. The right column images were corrected us-l magnification of the top row images. The bottom rowement of the middle row images. The significant differ-windows are [�100,300] HU.

onstritholatiolocaasur

into clinical arenas.

503


REFERENCES

1. Wang G, Snyder DL, Osullivan JA, Vannier MW. Iterative deblurring for CTmetal artifact reduction. IEEE Trans Med Imaging 1996; 15:657–664.

2. Wang G, Vannier MW, Cheng PC. Iterative X-ray cone-beam tomogra-phy for metal artifact reduction and local region reconstruction. Mi-crosc Microanaly 1999; 5:58–65.

3. Kak AC, Slaney M. Principles of Computerized Tomographic Imaging.New York: IEEE Press; 1988.

4. Mahnken AH, Raupach R, Wildberger JE, et al. A new algorithm formetal artifact reduction in computed tomography - In vitro and in vivoevaluation after total hip replacement. Investigative Radiology, 2003.38(12): p. 769–775.

5. Kalender WA, Hebel R, Ebersberger J. Reduction of CT artifactscaused by metallic implants. Radiology 1987; 164:576–577.

6. Watzke O, Kalender WA. A pragmatic approach to metal artifact reduc-tion in CT: merging of metal artifact reduced images. Eur Radiol 2004;14:849–856.

7. Kudo H, Saito T. Sinogram recovery with the method of convex projec-tions for limited-data reconstruction in computed-tomography. J OpticSoc Am A Optics Image Sci Vision 1991; 8:1148–1160.

8. Zhao SY, Bae KT, Whiting B, Wang G. A wavelet method for metal arti-fact reduction with multiple metallic objects in the field of view. J X-raySci Technol 2002; 10:67–76.

9. Zhao SY, Robertson DG, Bae KT, Whiting B, Wang G. X-ray CT metalartifact reduction using wavelets: An application for imaging total hip
prostheses. IEEE Trans Med Imaging 2000; 19:1238–1247.
504

10. Wei JK, Chen LG, Sandison GA, Liang Y, Xu LX. X-ray CT high-densityartefact suppression in the presence of bones. Phys Med Biol 2004;49:5407–5418.

11. Yazdi M, Gingras L, Beaulieu L. An adaptive approach to metal artifactreduction in helical computed tomography for radiation therapy treat-ment planning: Experimental and clinical studies. Int J Radiat OncolBiol Phys 2005; 62:1224–1231.

12. Bal M, Spies L. Metal artifact reduction in CT using tissue-classmodeling and adaptive prefiltering. Med Phys 2006; 33:2852–2859.

13. Cheng YZ. Mean shift, mode seeking, and clustering. IEEE Trans Pat-tern Anal Machine Intellig 1995; 17:790–799.

14. Comaniciu D, Meer P. Mean shift: A robust approach toward featurespace analysis. IEEE Trans Pattern Anal Machine Intellig 2002; 24:603–619.

15. Noo F, Defrise M, Clackdoyle R. Single-slice rebinning method for heli-cal cone-beam CT. Phys Med Biol 1999; 44:561–570.

16. Chen LG, Liang Y, Sandison GA, Rydbery J. Novel method for reduc-ing high-attenuation object artifacts in CT reconstructions. Proc SPIE2002; 4684:841–850.

17. Fukunaga K, Hostetler LD. The estimation of the gradient of a densityfunction, with applications in pattern recognition. EEE Trans InforTheory 1975; 21:32–40.

18. Comaniciu D, Ramesh V, Meer P. The variable bandwidth mean shiftand data-driven scale selection. 8th IEEE International Conference on
Computer Vision, 2002, pp 438–445.

Documents

A Segmentation-Based Method for Metal Artifact Reduction1