Streak artifacts arising from metal implants such as dental fillings, surgical clips, coils, wires, and orthopedic hardware may obscure important diag

Transcription

1 Note: This copy is for your personal, non-commercial use only. To order presentation-ready copies for distribution to your colleagues or clients, contact us at ORIGINAL RESEARCH n TECHNICAL DEVELOPMENTS F. Edward Boas, MD, PhD Dominik Fleischmann, MD Evaluation of Two Iterative Techniques for Reducing Metal Artifacts in Computed Tomography1 Purpose: To evaluate two methods for reducing metal artifacts in computed tomography (CT) the metal deletion technique (MDT) and the selective algebraic reconstruction technique (SART) and compare these methods with filtered back projection (FBP) and linear interpolation (LI). Materials and Methods: Results: Conclusion: The institutional review board approved this retrospective HIPAA-compliant study; informed patient consent was waived. Simulated projection data were calculated for a phantom that contained water, soft tissue, bone, and iron. Clinical projection data were obtained retrospectively from 11 consecutively identified CT scans with metal streak artifacts, with a total of 178 sections containing metal. Each scan was reconstructed using FBP, LI, SART, and MDT. The simulated scans were evaluated quantitatively by calculating the average error in Hounsfield units for each pixel compared with the original phantom. Two radiologists who were blinded to the reconstruction algorithms used qualitatively evaluated the clinical scans, ranking the overall severity of artifacts for each algorithm. P values for comparisons of the image quality ranks were calculated from the binomial distribution. The simulations showed that MDT reduces artifacts due to photon starvation, beam hardening, and motion and does not introduce new streaks between metal and bone. MDT had the lowest average error (76% less than FBP, 42% less than LI, 17% less than SART). Blinded comparison of the clinical scans revealed that MDT had the best image quality 100% of the time (95% confidence interval: 72%, 100%). LI had the second best image quality, and SART and FBP had the worst image quality. On images from two CT scans, as compared with images generated by the scanner, MDT revealed information of potential clinical importance. For a wide range of scans, MDT yields reduced metal streak artifacts and better-quality images than does FBP, LI, or SART. q RSNA, From the Department of Radiology, Stanford University Medical Center, 300 Pasteur Dr, Room H-1307, Stanford, CA Received September 20, 2010; revision requested November 10; revision received November 24; accepted December 10; final version accepted December 20. Address correspondence to F.E.B. ( boas@stanford.edu ). Supplemental material: /suppl/doi: /radiol /-/dc1 q RSNA, radiology.rsna.org n Radiology: Volume 259: Number 3 June 2011

2 Streak artifacts arising from metal implants such as dental fillings, surgical clips, coils, wires, and orthopedic hardware may obscure important diagnostic information in computed tomography (CT). Metal streak artifacts occur because filtered back projection (FBP) assumes that each detector measurement is equally accurate ( 1 ). In reality, rays that pass through or near metal implants are highly attenuated and have a much larger error due to a combination of scatter, beam-hardening effects, noise from low photon counts (ie, photon starvation), edge effects, and patient motion ( 2 ). Scattered photons and dark current (background signal from the detector when no photons are detected) increase the detector signal, resulting in dark streaks between metal implants ( 3 ). Beamhardening effects are seen in CT images acquired with polychromatic x-ray sources. As the x-ray passes through metal, low-energy x-ray photons are absorbed first, and the remaining highenergy photons are not attenuated as easily. These extra photons also result in dark streaks between metal structures. Low photon counts have a statistical error that follows the Poisson distribution. This results in fine bright and dark streaks originating from the metal. Fine streaks seen along the longest dimension of images obtained in obese patients are Advances in Knowledge n Metal artifacts in CT images can be reduced by using MDT, which uses forward projection to replace detector measurements that involve metal. n MDT reduces metal artifacts caused by photon counting noise, beam hardening, and motion. n MDT avoids introducing new streaks between metal and bone, which are seen with other metal artifact reduction methods. n MDT yields significantly better image quality ( P =.0005) than filtered back projection and linear interpolation, and may affect the radiologic diagnosis. also due to Poisson noise. If no photons are detected, this can result in dark streaks similar to beam hardening, because zero must be arbitrarily replaced by a nonzero number of photons. Otherwise, no photons will correspond to an infinite x-ray attenuation coefficient. Edge effects produce fine bright and dark streaks that originate from sharp changes in x-ray attenuation. This can be caused by multiple mechanisms ( 2 ), including (a) patient motion, (b) collecting projections from too few angles (undersampling), (c) windmill artifacts generated by interpolating multidetector row helical scan data to generate two-dimensional fan projections, and (d) using x-ray beams of nonzero width, which results in an average attenuation that does not exactly correspond to the average attenuation expected by the reconstruction algorithms ( 4 ). Several strategies to reduce the severity of artifacts in CT images have been proposed. Photon counting noise can be reduced by increasing the tube current, but this increases the patient s exposure to radiation. A beam-hardening correction can be applied iteratively on the basis of the current reconstructed image ( 5 ). The algebraic reconstruction technique ( 1 ) can be modified to converge to a maximum likelihood, maximum entropy, or minimum norm solution ( 6 9 ). Noisy projection data can be replaced with smoothed ( 10 ) or interpolated ( ) data. Specifically, the linear interpolation (LI) technique ( 13,14 ) erases the metal by replacing rays that pass through metal with values linearly interpolated from rays that pass adjacent to the metal, then uses FBP to reconstruct the image. Current techniques for reducing metal artifacts in CT images have not achieved widespread clinical use, partly because they do not address all sources of streaking and their use can introduce Implication for Patient Care n MDT reduces metal streak arti- facts in CT images containing orthopedic devices, dental fillings, embolization coils, surgical clips, or moving pacer wires. new artifacts ( 7,15 ). The maximum likelihood method, for example, involves finding an image with the highest probability of generating the projection data, assuming that photon counts in each detector element follow a Poisson distribution. This ignores other sources of error, such as scatter, edge effects, and errors in the beam-hardening correction. Furthermore, many images are consistent with the projection data within the experimental error, but the maximum likelihood method does not specify which image to pick. Thus, the final reconstructed image is based on the initial image and the number of maximum likelihood iterations applied. Applying too many iterations results in overfitting and a noisier image ( 7 ). The LI method effectively enables one to erase the metal and thus reduce streaking due to beam hardening, photon starvation, and edge effects. However, it does not address Poisson noise for x-rays passing through soft tissue and bone. Furthermore, the LI process can introduce new artifacts for example, streaks between metal and bone ( 15 ). We attempted to develop techniques to address several sources of artifacts, including Poisson noise, beam hardening, and edge effects. The purpose of our study was to evaluate two of these methods for reducing the severity of metal artifacts in CT images the metal deletion Published online before print /radiol Radiology 2011; 259: Abbreviations: AMPR = adaptive multiple plane reconstruction FBP = fi ltered back projection LI = linear interpolation MDT = metal deletion technique SART = selective algebraic reconstruction technique Author contributions: Guarantors of integrity of entire study, F.E.B., D.F.; study concepts/study design or data acquisition or data analysis/ interpretation, F.E.B., D.F.; manuscript drafting or manuscript revision for important intellectual content, F.E.B., D.F.; manuscript fi nal version approval, F.E.B., D.F.; literature research, F.E.B..; experimental studies, F.E.B., D.F.; statistical analysis, F.E.B.; and manuscript editing, F.E.B., D.F. Potential confl icts of interest are listed at the end of this article. Radiology: Volume 259: Number 3 June 2011 n radiology.rsna.org 895

3 technique (MDT) and the selective algebraic reconstruction technique (SART) and compare these methods with FBP and LI. Materials and Methods Algorithms This retrospective HIPAA-compliant study was approved by our institutional review board, and informed patient consent was waived. SART involves the use of an algebraic reconstruction technique to try to match the projection data to within the experimental error. In each SART iteration, an edge-preserving blur filter is applied to guide convergence to a smoother image from the large set of images consistent with the projection data. Next, the image is modified by using the algebraic reconstruction technique to try to match the projection data to within the experimental error. The experimental error model includes Poisson noise and beam-hardening effects. We use all of the projection data to reconstruct the metal regions of the image, but we selectively use high-quality data (rays that do not pass through or near metal) to reconstruct the nonmetal regions of the image. This should reduce edge effects and other errors that are missing from our model. A total of 25 iterations are performed. With MDT, forward projection is performed iteratively to replace detector measurements that involve metal ( Fig 1 ). This produces a self-consistent set of projection data with the metal removed. First, each detector element is expanded until at least an estimated 30 photons are detected. Next, the initial image is constructed by using LI. Finally, FBP is iterated four times. At each iteration, rays that pass through metal (identified on the FBP image) are replaced with forward-projected values from the previous iteration. Several key modifications of the basic MDT framework described above are needed to achieve optimal image quality. First, to reduce noise and streaking, an edge-preserving blur filter is applied to the image before forward projection. Second, because the forward-projected values do not exactly match the original projection data due to beam hardening, attenuation outside the reconstructed region, and other factors, we add a linear function to the forward-projected values to eliminate discontinuities when they are spliced back into the experimental projection data. Finally, to further reduce streaking, rays passing near metal are replaced with a weighted average of the experimental projection data and the forward-projected data, allowing a smoother transition. Details of these methods are presented in Appendix E1 (online). Generation of Simulated Scan Data We simulated a cm elliptical phantom comprising a grid pattern of water and soft tissue and containing two cm iron implants and one cm cortical bone implant ( Fig 2 ). The properties of these materials were obtained from ICRU-44 (International Commission on Radiation Units and Measurements report 44) ( 16 ). Projection data for a 64 detector row CT scanner (Sensation 64; Siemens Medical Solutions, Erlangen, Germany) were simulated with a single detector row without table translation. Flying focal spot data ( 17 ) were not included. Exact projection data were calculated by assuming an ideal monochromatic 50-keV x-ray beam that was not susceptible to Poisson noise, hardening, scatter, or motion. We simulated Poisson noise by setting the x-ray tube output to result in photons per detector element when nothing was in the scanner. This corresponds to a tube current of approximately 500 ma and a tube voltage of 120 kv for the scanner that we used. (See Estimating Absolute Photon Counts in Appendix E1 [online].) At these tube settings in this phantom, rays passing through bone have a minimum expected photon count of 1.2, which will result in a large amount of Poisson noise, and rays passing through metal have an expected photon count of approximately zero, which means that they carry no information about nonmetal pixels. In simulations with Poisson noise, rays with zero photons detected were arbitrarily set to 0.1 photons, because the logarithm of zero is undefined. Beam hardening was simulated by assuming that 90% of the x-ray photons had an energy of 50 kev and that 10% of the photons had an energy of 100 kev. (A bichromatic energy spectrum is the simplest spectrum that will result in beam hardening.) Motion was simulated by rotating the center of the metal implants in a circle with a 0.6-mm radius. The reconstructions were performed with mm pixels, and a total image size of pixels. Exact simulated projection data as well as projection data corrupted by Poisson noise, beam hardening, motion, or all three factors were reconstructed by using FBP, LI, SART, and MDT. Evaluation of Simulated Scan Data We evaluated the simulated reconstructions quantitatively by calculating the average absolute error in Hounsfield units for each pixel compared with the Hounsfield units in the original phantom. Hounsfield units greater than 3000 HU were capped at 3000 HU before the average error was calculated. The central processing unit time for each reconstruction, performed on one core of an eight-core central processing unit ( Opteron 2384; AMD, Sunnyvale, Calif) was recorded. To determine how well the structures between the two metal implants could be seen in the projection data, we perturbed the Hounsfield units between the implants by 100 HU in either direction, along two parallel 0.4-mm-wide strips. We then calculated the maximum difference in the projection data after this perturbation. Clinical Scans in Study Subjects Clinical CT data sets containing metal implants were identified retrospectively by screening the CT scout images from all 234 scans performed on a 64 detector row outpatient CT scanner (Sensation 64) on eight random days between August 2007 and February The presence of metal streak artifact was then confirmed on the cross-sectional images. Forty-eight scans with metal artifact were identified. Due to technical limitations 896 radiology.rsna.org n Radiology: Volume 259: Number 3 June 2011

4 Figure 1 Figure 2 Figure 2: Phantom used for simulated scans contains two iron implants and one bone implant. Figure 1: Simplifi ed diagram of MDT. Forward projection is used iteratively to replace corrupted metal data, which are the cause of the streaks. End result is that nonmetal pixel values are determined using only nonmetal projection data. Metal pixels and projection data are shown in red. For simplicity, only a single projection is shown. Original projection data are plotted with thick lines, and revised projection data are plotted with thin lines. of our software, the data from scans ( n = 37) with a gantry tilt or a pitch higher than 1 were then excluded. Our final sample consisted of images from 11 scans performed in five men and six women (age range, years; average age, 64 years). The images from eight abdominopelvic scans, one chest scan, one neck scan, and one facial scan were included. Three scans contained embo lization coils, two contained hip replace ments, two contained dental fillings, two contained pacemakers, two contained cholecystectomy clips, and one scan contained splenectomy clips. The sizes of the metal implants, as measured on axial CT images, ranged from 0.8 mm to 50 mm. Images from each scan are shown in Figure E5 (online). Reconstruction of Clinical Scan Data The raw data from each scan, with the exclusion of the flying focal spot data, were reconstructed by using FBP, LI, SART, and MDT with a section thickness of 1.2 mm and intersection spacing of 5 mm. A total of 774 sections were reconstructed, 178 of which contained metal. The images generated by the scanner, which uses the flying focal spot data with adaptive multiple plane reconstruction (AMPR), also were evaluated. AMPR ( 17,18 ) was designed to reduce windmill artifacts. The MDT images were averaged so that their section thickness matched the section thick ness of the AMPR images from the scanner, which ranged from 1 to 5 mm. Thin sections were used when available. The omission of flying focal spot data in our reconstructions corresponded to a 48% average reduction (range, 0% 75%, depending on the number of flying focal spots, which can be one, two, or four) in the amount of data used for the reconstruction. Evaluation of Clinical Scan Data Data from the clinical scans were evaluated qualitatively by two independent radiologists (D.F. and a second radiologist not involved in the study) who were blinded to the reconstruction algorithms used. Both readers had 19 years of experience in CT image interpretation. Images obtained by using FBP, LI, SART, and MDT were displayed side by side in different random orders for each scan, and the two radiologists ranked the overall severity of the artifacts for each algorithm, with a rank of 1 indicating the least severe artifacts and a rank of 4 indicating the most severe artifacts. In a separate comparison, the two radiologists compared the MDT images with the AMPR images generated by the scanner and judged which algorithm yielded the least severe streak artifacts and which had the least noise for each scan. One radiologist (D.F.) evaluated the MDT images for new or altered findings, as compared with findings seen at the original image interpretation performed by using only the AMPR images from the scanner. One of the authors (F.E.B.) then performed a chart review to evaluate any discrepancies. Statistical Analysis The image quality ranks for the different artifact reduction algorithms and different observers were compared by using the binomial distribution to calculate P values. Specifically, the P value equals the cumulative binomial probability, where the number of trials equals the number of comparisons made, the number of successes equals the number of comparisons in which one algorithm (or observer) has a higher rank, and the probability of success is.5 (the null Radiology: Volume 259: Number 3 June 2011 n radiology.rsna.org 897

5 Figure 3 Figure 3: Simulated CT images using either exact projection data or projection data corrupted by Poisson noise, beam hardening, motion, or all three phenomena. Scan data were reconstructed using FBP, LI, SART, and MDT. MDT produced the least streak artifacts and prevented introduction of new streaks between metal and bone. Average error in Hounsfi eld units relative to original image is shown below each image. hypothesis) ( 19 ). The ranks assigned by the two reviewers for a given scan were averaged before the different reconstruction techniques were compared. Bonferroni correction was applied to determine the cutoff for significance. Confidence intervals for proportions were calculated by using the method of Clopper and Pearson ( 20 ). Results Simulated Scan Data FBP, LI, SART, and MDT reconstructions of exact simulated projection data and projection data corrupted by Poisson noise, beam hardening, motion, or all three factors are shown in Figure 3. Overall, MDT produced the smallest error in Hounsfield units for each of the three sources of metal artifact. When Poisson noise, beam hardening, and motion were all included, the average error in Hounsfield units for MDT compared with the phantom was 90 HU; the average error for FBP, 375 HU; the average error for LI, 154 HU; and the average error for SART, 108 HU. The time to reconstruct each image was 1.5 minutes with FBP, 3.1 minutes with LI, 28 minutes with MDT, and 99 minutes with SART. The long reconstruction times were due to the fact that the reconstructions were performed on a generalpurpose central processing unit without any hardware acceleration. MDT is approximately 19 times slower than FBP, but it needs to be applied only to sections with metal. The FBP reconstruction of exact projection data (ie, data with no noise, beam hardening, or motion) contained streaks that were due to the finite number of detector measurements. Use of LI and SART introduced new streaks around the bone implant on the reconstructions of exact data ( Fig 3 ). For exact data, MDT had a slightly larger average error than did SART due to diffusely increased noise. Visually, however, MDT appeared to produce a better result for exact data with almost no streaking around the bone implant, emphasizing 898 radiology.rsna.org n Radiology: Volume 259: Number 3 June 2011

6 the need to also examine the images qualitatively rather than rely solely on quantitative measures. For exact data, the line of soft-tissue between the two metal implants was seen with FBP but not with the other reconstruction methods. This was due to the fact that this line of soft tissue can be clearly seen only by looking through the metal, and the metal projection data are discarded with LI, SART, and MDT. However, when Poisson noise and the discrete nature of photons are taken into account, the line of soft-tissue is not clearly seen with any of the reconstruction techniques, and, thus, no loss of information results from discarding the metal projection data. This is because fractional photons cannot be detected, so exact measurements are not possible. For example, an exact detector measurement of 0.01 photons or photons almost always corresponds to an actual detector measurement of zero photons in both cases. In this phantom, the metal blocks all photons, so large perturbations can be made along the line of soft tissue between the metal structures with no change in the actual number of photons detected. Perturbing the Hounsfield units between the two metal implants by HU results in a maximum difference in projection data of 0.05 photons. In other words, information about this line of soft tissue is not present in the projection data. Clinical Scan Findings Representative images from each reconstruction technique showed that LI and SART introduced new streaks between metal and high-attenuating material that did not meet the cutoff for metal (such as bone and contrast material). These new streaks were almost completely eliminated with MDT ( Fig 4, Fig E5 [online]). Both reviewers independently agreed that MDT rendered the best image quality in 100% (95% confidence interval: 72%, 100%) ( n = 11) of the patients ( Table, Table E4 [online]). MDT image quality was significantly better than LI image quality ( P =.0005), which was significantly better than FBP image quality ( P =.001). There was no significant difference in image quality between SART and FBP. Image Quality Ranks for Each CT Image Reconstruction Method Reconstruction Method There were no significant differences in image quality rank between the two reviewers ( P =.06 to P =.75). When we directly compared the MDT images with the AMPR images generated by the scanner, the noise level was higher on the MDT images ( P =.01), as expected given the lack of flying focal spot data, which corresponds to a reduction in the theoretical radiation dose. However, des pite the theoretical reduction in radia tion dose and the simpler helical inter polation method used, the MDT images still were judged to have decreased streak artifacts compared with the AMPR images in all (100%) 11 scans (95% confidence interval: 72%, 100%; P =.0005) (Table E5 [online]). In two cases, the MDT images incidentally revealed potentially important new findings ( Figs 5 and 6 ). In the first case, an 85-year-old man with stage T3 N0 M0 rectal cancer was originally found after low anterior resection and chemotherapy to have rectal wall thickening but no definitive evidence of cancer recurrence. However, MDT revealed perirectal lymphadenopathy that was obscured by artifacts from his bilateral hip replacements on the original AMPR images. In retrospect, a hint of nodularity might have been visible through the streaks, but this was not addressed prospectively. The MDT reconstruction was not performed until after the patient died (8 months after the scan), however, and recurrent rectal cancer, which was not suspected at the time of death, was confirmed at autopsy. In the second case, a 57-year-old man Image Quality Rank * Reviewer 1 Reviewer 2 Average had a pacemaker lead for which concerns about perforation were reported in the original interpretation; however, the device was found to be nonperforated after MDT reconstruction. At clinical follow-up 2 months later, the patient had no signs of lead perforation (chest pain, failure to pace, or friction rub). Discussion P Value MDT LI FBP SART NA * Mean ranks 6 standard errors of the mean. Average of mean ranks for reviewers 1 and 2. P values for comparison with the next best technique (in the next row). NA = not applicable. Signifi cant difference ( P,.017 [.05/3]). Our results from both simulated and clinical scans show that use of MDT, as compared with other reconstruction techniques, significantly reduces metal artifacts and prevents the introduction of new streaks between metal and bone (seen with LI and SART). Results of our simulation experiments confirmed that MDT reduces artifacts caused by multiple different mechanisms including edge effects, Poisson noise, beam hardening, and patient motion. Poisson noise is reduced by expanding the detector elements in regions with low photon counts. Dark streaks (caused by beam hardening, scatter, or complete attenuation of photons by metal), edge effects, and patient motion are addressed by deleting the metal. Projection data that include metal are then replaced by values derived from the rest of the projection data. MDT is conceptually related to the recently developed forward-projection metal artifact reduction technique ( 15 ), with several key differences. For example, we use an edge-preserving blur filter to reduce noise before the forward Radiology: Volume 259: Number 3 June 2011 n radiology.rsna.org 899

7 Figure 4 Figure 4: Representative axial sections from clinical CT scans reconstructed with FBP, LI, SART, and MDT. Top row shows embolization coils, second row shows dental fi llings, third row shows embolization coils, and bottom row shows cholecystectomy clips. At blinded comparison, MDT was judged to yield the best overall image quality. Note streaks between metal and other high-attenuating material (such as bone or contrast material) seen on LI and SART images but not on MDT images. projection instead of segmenting the image into air, soft tissue, and bone. This facilitates a greater reduction in noise and streaking and allows us to use the actual Hounsfield units rather than assume that all soft tissue, for example, has the same Hounsfield units. Furthermore, we iterate the method to improve results. Finally, we reduce photon counting error by expanding the detector elements in regions with low photon counts. From a theoretical viewpoint, MDT seems very similar to SART. Both techniques attempt to address all major sources of metal streak artifact: Poisson noise or photon starvation, beam hardening, and motion or edge effects. With both techniques, rays that pass through or near metal are ignored when nonmetal portions of the image are reconstructed. The performance of these methods was similarly good for simulated data. However, for the clinical scans, SART was tied with another method for the worst image quality. This discrepancy highlights the importance of testing reconstruction algorithms on a wide range of both simulated and clinical scans. Several techniques that work well with specific types of metal implants, such as hip replacements ( 12,21,22 ), have been developed. In contrast, our initial clinical results show that use of MDT reduces artifacts in a wide range of patient scans with single or multiple pieces of metal of various shapes and sizes (0.8 mm to 50 mm). Images depicting hip replacements, which are large and stationary, and pacer wires, which are small and mobile, were constructed by using the same technique. No manual steps or parameter adjustments were 900 radiology.rsna.org n Radiology: Volume 259: Number 3 June 2011

8 Figure 5 Figure 5: Axial CT images obtained in an 85-year-old man after resection and chemotherapy for rectal cancer. (a) FBP image shows minimal irregularity in the perirectal space, which is obscured by metal artifacts from bilateral hip implants. Patient was thought to be cancer free at time of death. (b) Corresponding MDT image shows perirectal lymphadenopathy (arrow). Recurrent rectal cancer was found at autopsy. Figure 6 Figure 6: Axial CT images in a 57-year-old man with a pacer wire. (a) FBP image shows streak artifacts around the pacer wire, which was reported to be concerning for perforation through the right ventricular wall. (b) Corresponding MDT image shows reduced artifacts and correct position of pacer wire within the right ventricle. of MDT, the reduction in artifacts revealed new findings that were potentially clinically important in two of the 11 patients. In these two cases, the MDT images, as compared with the AMPR images (from the scanner) used for the original interpretations, showed denecessary. MDT was judged to be superior to the other methods in all 11 cases. Although our sample size was small, differences in image quality ranks were significant. Although this study was not designed to assess the diagnostic accuracy creased streaking and new findings despite the use of only 50% of the projections. We anticipate that using MDT in conjunction with the flying focal spot data and AMPR would further improve image quality. The main limitations of MDT stem from the fact that metal projection data are discarded, resulting in a potential loss of spatial resolution near the metal implant. (See the Image Spatial Resolution section in Appendix E1 [online]). The actual degree of spatial resolution loss depends on how much information the discarded data contained. In some cases, if the metal projection data contain no information about soft-tissue, then no soft-tissue information will be lost when the metal data are discarded. This occurs if the metal implant is large and attenuates all photons, or if it is smaller than the degree of motion of the implant during the scan which will corrupt the metal data. However, in other cases, the metal data do contain information about soft tissue and bone, so discarding these data results in a loss of spatial resolution around the metal implant. Due to the difficulty in determining whether spatial resolution loss has occurred, we recommend always reviewing the MDT images in conjunction with the traditional FBP images. To limit the loss of spatial resolution, future research should attempt to determine which metal data Radiology: Volume 259: Number 3 June 2011 n radiology.rsna.org 901

9 cannot be safely discarded. Applying a fixed photon cutoff will not resolve this problem, as small surgical clips affected by respiratory motion will cause substantial streaking despite having a relatively high photon count. MDT has several other limitations. First, beam hardening due to bone or other nonmetallic high-attenuating materials such as contrast medium is not addressed. Second, step-off artifacts can be introduced at metal edges. Third, unlike with FBP or LI, with MDT, limited field-of-view reconstructions are not possible. Fourth, we only tested MDT with a single CT scanner model and have not tested it on scans performed with a helical pitch higher than 1. Finally, MDT is 19 times slower than FBP, and we have not yet implemented hardware acceleration. In summary, MDT reduces metal artifacts in both simulated and clinical scans and has the potential to improve diagnostic accuracy. Future work is needed to compare MDT with several recently introduced methods and to determine the diagnostic accuracy of this technique for specific applications. Acknowledgments: The authors thank Christoph Panknin, PhD, of Siemens Medical Solutions for help reading the raw data generated by the scanner; Chris Beaulieu, MD, PhD, for reviewing the quality of the images produced by each reconstruction technique; Sandy Napel, PhD, for providing valuable feedback regarding the manuscript; and Sam Mazin, PhD, for helpful discussions on metal artifact reduction algorithms. Disclosures of Potential Conflicts of Interest: F.E.B. No potential conflicts of interest to disclose. D.F. Financial activities related to the present article: none to disclose. Financial activities not related to the present article: received grants or grants are pending from GE Medical Systems; received grants or grants are pending from Siemens Medical Solutions; received payment for lectures from the Bracco Group. Other relationships: none to disclose. References 1. Kak AC, Slaney M. Principles of computerized tomographic imaging. New York, NY : IEEE Press, De Man B, Nuyts J, Dupont P, Marchal G, Suetens P. Metal streak artifacts in x-ray computed tomography: a simulation study. IEEE Trans Nucl Sci 1999 ; 46 ( 3 ): Joseph PM, Spital RD. The effects of scatter in x-ray computed tomography. Med Phys 1982 ; 9 ( 4 ): Joseph PM, Spital RD. The exponential edgegradient effect in x-ray computed tomography. Phys Med Biol 1981 ; 26 ( 3 ): Hsieh J, Molthen RC, Dawson CA, Johnson RH. An iterative approach to the beam hardening correction in cone beam CT. Med Phys 2000 ; 27 ( 1 ): De Man B, Nuyts J, Dupont P, Marchal G, Suetens P. Reduction of metal streak artifacts in x-ray computed tomography using a transmission maximum a posteriori algorithm. IEEE Trans Nucl Sci 2000 ; 47 ( 3 ): Vandenberghe S, D Asseler Y, Van de Walle R, et al. Iterative reconstruction algorithms in nuclear medicine. Comput Med Imaging Graph 2001 ; 25 ( 2 ): Verhoeven D. Limited-data computed tomography algorithms for the physical sciences. Appl Opt 1993 ; 32 ( 20 ): De Man B, Nuyts J, Dupont P, Marchal G, Suetens P. An iterative maximum-likelihood polychromatic algorithm for CT. IEEE Trans Med Imaging 2001 ; 20 ( 10 ): Hsieh J. Adaptive streak artifact reduction in computed tomography resulting from excessive x-ray photon noise. Med Phys 1998 ; 25 ( 11 ): Glover GH, Pelc NJ. An algorithm for the reduction of metal clip artifacts in CT reconstructions. Med Phys 1981 ; 8 ( 6 ): Mahnken AH, Raupach R, Wildberger JE, et al. A new algorithm for metal artifact reduction in computed tomography: in vitro and in vivo evaluation after total hip replacement. Invest Radiol 2003 ; 38 ( 12 ): Kalender WA, Hebel R, Ebersberger J. Reduction of CT artifacts caused by metallic implants. Radiology 1987 ; 164 ( 2 ): Rinkel J, Dillon WP, Funk T, Gould R, Prevrhal S. Computed tomographic metal artifact reduction for the detection and quantitation of small features near large metallic implants: a comparison of published methods. J Comput Assist Tomogr 2008 ; 32 ( 4 ): Prell D, Kyriakou Y, Beister M, Kalender WA. A novel forward projection-based metal artifact reduction method for flat-detector computed tomography. Phys Med Biol 2009 ; 54 ( 21 ): Hubbell J, Seltzer S. Tables of x-ray mass attenuation coefficients and mass energyabsorption coefficients. National Institute of Standards and Technology Web site. http: //physics.nist.gov/physrefdata/xraymass Coef/cover.html. Accessed February 21, Flohr TG, Stierstorfer K, Ulzheimer S, Bruder H, Primak AN, McCollough CH. Image reconstruction and image quality evaluation for a 64-slice CT scanner with z-flying focal spot. Med Phys 2005 ; 32 ( 8 ): Schaller S, Stierstorfer K, Bruder H, Kachelrieß M, Flohr T. Novel approximate approach for high-quality image reconstruction in helical cone beam CT at arbitrary pitch. Proc SPIE 2001 ; 4322 : Devore JL. Probability and statistics for engineering and the sciences. Pacific Grove, Calif : Brooks/Cole, Clopper CJ, Pearson ES. The use of confidence or fiducial limits illustrated in the case of the binomial. Biometrika 1934 ; 26 ( 4 ): Prell D, Kyriakou Y, Kachelrie M, Kalender WA. Reducing metal artifacts in computed tomography caused by hip endoprostheses using a physics-based approach. Invest Radiol 2010 ; 45 ( 11 ): Robertson DD, Yuan J, Wang G, Vannier MW. Total hip prosthesis metal-artifact suppression using iterative deblurring reconstruction. J Comput Assist Tomogr 1997 ; 21 ( 2 ): radiology.rsna.org n Radiology: Volume 259: Number 3 June 2011