- Open Access
Post-processing sets of tilted CT volumes as a method for metal artifact reduction
© Ballhausen et al.; licensee BioMed Central Ltd. 2014
- Received: 16 December 2013
- Accepted: 3 May 2014
- Published: 15 May 2014
We’re sorry, something doesn't seem to be working properly.
Please try refreshing the page. If that doesn't work, please contact support so we can address the problem.
Metal implants, surgical clips and other foreign bodies may cause ‘streaking’ or ‘star’ artifacts in computed tomography (CT) reconstructions, for example in the vicinity of dental restorations or hip implants. The deteriorated image quality complicates contouring and has an adverse effect on quantitative planning in external beam therapy.
The potential to reduce artifacts by acquisition of tilted CT reconstructions from different angles of the same object was investigated. While each of those reconstructions still contained artifacts, they were not necessarily in the same place in each CT. By combining such CTs with complementary information, a reconstructed volume with less or even without artifacts was obtained. The most straightforward way to combine the co-registered volumes was to calculate the mean or median per voxel. The method was tested with a calibration phantom featuring a titanium insert, and with a human skull featuring multiple dental restorations made from gold and steel. The performance of the method was compared to established metal artifact reduction (MAR) algorithms. Dose reduction was tested.
In a visual comparison, streaking artifacts were strongly reduced and details in the vicinity of metal foreign bodies became much more visible. In case of the calibration phantom, average bias in Hounsfield units was reduced by 94% and per-voxel-errors and noise were reduced by 83%. In case of the human skull, bias was reduced by 95% and noise was reduced by 94%. The performance of the method was visually superior and quantitatively compareable to established MAR algorithms. Dose reduction was viable.
A simple post-processing method for MAR was described which required one or more complementary scans but did not rely on any a priori information. The method was computationally inexpensive. Performance of the method was quantitatively comparable to established algorithms and visually superior in a direct comparison. Dose reduction was demonstrated, artifacts could be reduced without compromising total dose to the patient.
- Hounsfield Unit
- Metal Artifact
- Reconstructed Volume
- Dental Restoration
- Metal Artifact Reduction
Planning in external beam radiotherapy critically relies on the ability to precisely delineate target volumes. As soft tissue contrast is limited, the image quality of the planning CT is often a limiting factor in the exact determination of boundaries and exacerbates inter-observer uncertainties. Also, the simulation of a plan demands a faithful reconstruction of the attenuation coefficient per voxel. The planning of intensity modulated radiotherapy and in particular proton and heavy ion therapy crucially depends on an unbiased representation of Hounsfield units.
In this context, metal implants such as dental restorations or hip prostheses give rise to beam hardening, scattered radiation, projection noise, trans-axial non-linear partial volume effects, and photon starvation, all of which may contribute to ‘streaking’ or ‘star’ or other artifacts that deteriorate the reconstruction in the vicinity of the foreign body. A common workaround in clinical routine is to manually segment those artifacts slice by slice and to replace their voxel grey values by equivalent values for tissue, water, bone, air, respectively. If done by hand, this process is cumbersome, time consuming, and somewhat subjective.
Alternatively, artifacts can be suppressed during reconstruction. Metal artifact reduction (MAR) algorithms have been developed for this purpose[2–4]. These algorithms ideally operate on the raw sinogram data, but may use forward-projected ‘virtual sinograms’ instead. In general, some assumption on the missing information is necessary, and typically the metal-affected parts of the sinogram are replaced by interpolation. For example, metal is segmented in the image space of each uncorrected slice and the ‘metal map’ is forward-projected to sinogram space. Then, the affected areas are replaced by linear one-dimensional interpolation (, referred to as LI-MAR in this paper). The interpolation may be bi-directional within a slice and weighted ([6, 7], reffered to as BI-MAR in this paper) or even comprise interpolation across slices over the longitudinal scan range. The method can be extended by normalizing the sinograms ([9, 10], referred to as NMAR) and by combining the high frequencies of the original reconstruction (for sharper edge rendering) with the low frequencies of a MAR-processed image (, referred to as FSMAR). Alternatively, the in-painting may be based on a tissue-class model and an automatic segmentation of the initial reconstruction. This resembles the discrete in-painting by hand as described above, but the blending takes place in sinogram space. Adaptive filtering techniques have also been proposed, as well as the replacement of the filtered back-projection by algebraic solvers[14–17]. Hybrid approaches combining or iterating the above elements have also been suggested[18, 19]. Still, metal artifact reduction remains an open problem in clinical routine.
As a beneficial side effect, one would expect overall image quality to improve. First, noise would be reduced when summing over several exposures. And second, aliasing from finite slice thickness would be reduced because tilted CTs could be resampled to a mesh of finer resolution.
The general concept to reduce noise by averaging over independent data is one of the paradigms of quantitative science, and also in imaging related fields the idea to reduce both noise and artifacts by super-sampling is well established. Examples comprise but are not limited to resonance Raman spectroscopy[20, 21], optical coherence tomography and magnetic resonance imaging[23, 24]. In the narrower context of metal artifact reduction algorithms, hybrid algorithms such as and can be viewed as averaging over (semi-independent) data. Still, to our knowledge, the notion to reduce metal artifacts in computerized tomography by averaging over reconstructions independently acquired from different angles is new.
Facility-specific technical information
All CTs were acquired by a standard Toshiba Aquilion LB scanner, used routinely as the planning CT in our clinic. Standard settings for head and neck tomography were used. Resolution was (0.1 cm)3 per voxel. Volumes were reconstructed by the filtered back-projection algorithm of the integrated software suite of the CT scanner. Reconstructed volumes were stored as 16 bit Digital Imaging and Communications in Medicine (DICOM) on the server for post-processing.
In a second experiment about possible dose reduction, a single scan of the Gammex phantom was acquired at 400 mA/s, or 200 mA per cycle of 0.5 seconds, respectively. This scan was repeated eight times at 1/8th of dose, at 50 mA/s or 25 mA per cycle. For comparison, further eight scans at 50 mA/s or 25 mA per cycle were acquired while the phantom was rotated by 0°, ±22.5°, ±45°,±67.5°, and +90° about the vertical axis between scans in order to provide a set of independent scans for the MAR method.
Noise levels and severity of artifacts were compared between the single scan at full dose, the sum of the former eight scans at reduced dose, and the median of the latter eight scans after co-registration.
Three different measures for the impact of an artifact on the grey values of a given region of voxels are considered: “Bias”, “Error” and “Noise”. All three are conveniently measured in Hounsfield units. “Bias” is the overall increase or decrease of average grey value. “Error” is the average total deviation of each voxel from its true value. “Noise” is the standard deviation of voxel grey values, see. “Bias” and “error” require a knowledge of true grey values, “noise” is meaningful in case of a region of constant grey values. The three measures are quantitatively defined as follows:
Consider some sub-volume consisting of n voxels, i = 1, …, n. Assume that the true Hounsfield units in this sub-volume are known and constant c. Denote the grey value of voxel i of the reconstructed volume x i . “Bias” is defined as E Bias = x i - c. “Error” is defined as E Error = |x i - c|. “Noise” is defined as the square root of . In case of the Gammex phantom, a ring-shaped area around the titanium insert is considered with inner diameter 8 mm and outer diameter 16 mm. A stack of 7 mm of the innermost slices contains n = 4172 voxels in this area. This region is filled with ‘Solid Water’ and c = 11 HU is assumed, according to. Results for a generic assumption of c = 0 HU would be quantitatively similar.
In case of the human skull, a volume of 6.1 × 1.1 × 4.1 cm3 centred in the mouth cavity between the dental implants is considered. This region is entirely filled with air which has c = - 1000 HU by definition. Note that this is represented digitally by zero values, so E Bias and E Error coincide. Results were checked to be similar for other volumes outside the skull close to the dental implants.
Comparison to other algorithms
The performance of the method was compared to four established algorithms: linear interpolation of the metal-affected areas of the sinogram (, referred to as LI-MAR in this paper), a bi-directional interpolation of the sinogram looking for close-by unaffected pixels (both in terms of axis and angle) and weighting them by range ([6, 7], reffered to as BI-MAR in this paper), normalized MAR ([9, 10], referred to as NMAR), and frequency split MAR (, referred to as FSMAR). In all cases a virtual sinogram was generated from the original reconstruction and used as input to the algorithms. Straightforward, non-optimized implementations were programmed. The threshold for the computation of metal maps was 2000 HU. In case of FS-MAR, the low frequency image was generated by Gaussian blurring with a radius of 3 pixels, and a blurred version of the metal map with Gaussian radius of 30 pixels was used as blending weight. A computation in debug mode on consumer hardware (laptop, single processor, single-threaded execution on a single core running at 2.6 GHz) allowed for a rough comparison of computing times. The quantitative performance of all algorithms was measured in terms of bias, error and noise as above. The quality of the reconstruction was compared visually, too.
In the above experiment with the Gammex phantom, 197 HU noise were encountered at 50 mA/s or 25 mA per cycle. If current were eight times higher at 400 mA/s or 200 mA per cycle, noise levels were expected to decline to 197 HU / √8 ≈ 69 to 70 HU. In fact, in a second experiment with the same phantom, same experimental setting, 70 HU noise were recorded in a single scan at 400 mA/s or 200 mA per cycle (same region of interest in the reconstructed volume as before). Similarly, the scan was repeated eight times at 50 mA/s or 25 mA per cycle, and a volume was added up from the eight reconstructed volumes. In this average volume, noise of 78 HU was detected, of the same order of magnitude.
Finally, the eight scans were repeated at 50 mA/s or 25 mA per cycle, but this time the phantom was rotated between scans to acquire eight reconstructions from different angles as needed for the MAR method. In case of the co-registration based on the mean, noise level was 45 HU and in case of the co-registration based on the median, noise level was 60 HU. Noise levels are lower because of the non-cubic voxel dimensions which lead to effective anti-aliasing when rotated back and co-registered. At the same time, artifacts were reduced by 77.5% in terms of bias when compared to the single scan at 400 mA/s or 200 mA per cycle. Note that artifacts were not reduced at all when comparing the single 400 mA/s scan with the additive eight 50 mA/s scans, as expected. In conclusion, the MAR method did accomplish a significant reduction of artifacts, while retaining or even slightly improving noise levels, at constant overall dose.
Comparison to established algorithms
Computing times and memory demand
It should be noted that there was some ‘hidden’ computation time in case of the described method, as the reconstruction software of the CT scanner had to calculate each of the reconstructions, before they were even input to the algorithm. Also, the computing time is outweighed in practice by the time it takes to acquire the additional scans.
Memory demand is higher in case of the described method, if all available reconstructions are held in memory simultaneously. The latter is convenient for implementation, but can be avoided.
As proposed, ‘tilted’ CTs taken from different angles of the same object feature complementary information as far as metal artifacts are concerned. Consequently, the combination of such complementary CTs is less prone to artifacts. The method has been tested with two datasets in a clinical setting. In both cases, metal artifacts were reduced, constituting a proof of principle for this approach.
The reduction of bias and noise by up to 95% is significant and substantial. In all circumstances, the median suppressed artifacts more strongly than the mean, and Hounsfield units were more truthfully represented.
An advantage of the method is that it does not require access to proprietary data of the scanning equipment such as sinogram data or the computation of a virtual sinogram. The method works as a purely post-processing step which can be done at any facility independent of the equipment employed or the used methods for reconstruction. This facilitates introduction into the clinical routine a lot as it alleviates a number of safety concerns and does not involve the equipment manufacturer. Also, unlike most algorithms for metal artifact reduction, the method does not rely on any prior about the missing information.
As the method only relies on operations such as median and mean calculations of CTs from the usual and validated workflow, the safety of the method can be easily established and the post-processing step is quite transparent as far as the validity of the calculated Hounsfield units is concerned.
The drawback of the method is the necessity to obtain at least two independent scans from different angles. This could become problematic whenever time is a concern, organ movement is inevitable or patient compliance is limited. Care must be taken that the repositioning of the patient itself does not affect organ placement.
Computing time is less than in case of established MAR algorithms. However, the additional time to perform several scans will be much more of a concern in clinical routine.
The additional scans took minutes rather than seconds to capture, and there was some ‘hidden’ computation cost as the scanner had to reconstruct each of the volumes, before they were even input into the MAR method. It is understood that the described method has limitations due to this overhead effort.The method improves with the number of independent CTs and their relative angle. A substantial improvement by 75%, however, was observed for only three CTs taken at ±22.5° which should be clinically manageable. For geometrical reasons, a CT gantry with a large opening such as the Toshiba Aquilion LB helps in this respect.
Additional radiation exposure need not be a concern. As demonstrated, each of the n scans can be done at 1/n dose. The mean (and to some degree also the median) of the n CTs will feature a similar signal-to-noise ratio as if a single scan at full dose had been performed.
A simple post-processing method was described which reduced metal artifacts by up to 95%. While the method required one or more complementary scans, it did not rely on any a priori information. As a pure post-processing method it was independent of the actual image acquisition and did not require any particular instrumental setup or access to raw data. The method was computationally inexpensive, but the time necessary to perform additional scans may limit its practicability. Performance of the method was quantitatively comparable to established algorithms and visually superior in a direct comparison. Dose reduction was demonstrated, artifacts could be reduced without compromising total dose to the patient.
- De Man B, Nuyts J, Dupont P, Marchal G, Suetens P: Metal streak artifacts in X-ray computed tomography: a simulation study. Nuclear Science, IEEE Transactions on 1999,46(3):691-696. 10.1109/23.775600View ArticleGoogle Scholar
- Kalender WA, Hebel R, Ebersberger J: Reduction of CT artifacts caused by metallic implants. Radiology 1987, 164: 576-577.View ArticlePubMedGoogle Scholar
- Schulze R, Heil U, Gross D, Bruellmann DD, Dranischnikow E, Schwanecke U, Schoemer E: Artefacts in CBCT: a review. Dentomaxillofac Radiol 2011, 40: 265-273. 10.1259/dmfr/30642039PubMed CentralView ArticlePubMedGoogle Scholar
- Mouton A, Megherbi N, Van Slambrouck K, Nuyts J, Breckon TP: An experimental survey of metal artefact reduction in computed tomography. J Xray Sci Technol 2013, 21: 193-226.PubMedGoogle Scholar
- Abdoli M, Ay MR, Ahmadian A, Zaidi H: A virtual sinogram method to reduce dental metallic implant artefacts in computed tomography-based attenuation correction for PET. Nucl Med Commun 2010, 31: 22-31. 10.1097/MNM.0b013e32832fa241View ArticlePubMedGoogle Scholar
- Mahnken AH, Raupach R, Wildberger JE, Jung B, Heussen N, Flohr TG, Gunther RW, Schaller S: A new algorithm for metal artifact reduction in computed tomography: in vitro and in vivo evaluation after total hip replacement. Invest Radiol 2003, 38: 769-775. 10.1097/01.rli.0000086495.96457.54View ArticlePubMedGoogle Scholar
- Liu PT, Pavlicek WP, Peter MB, Spangehl MJ, Roberts CC, Paden RG: Metal artifact reduction image reconstruction algorithm for CT of implanted metal orthopedic devices: a work in progress. Skeletal Radiol 2009, 38: 797-802. 10.1007/s00256-008-0630-5View ArticlePubMedGoogle Scholar
- Yu L, Li H, Mueller J, Kofler JM, Liu X, Primak AN, Fletcher JG, Guimaraes LS, Macedo T, McCollough CH: Metal artifact reduction from reformatted projections for hip prostheses in multislice helical computed tomography: techniques and initial clinical results. Invest Radiol 2009, 44: 691-696. 10.1097/RLI.0b013e3181b0a2f9PubMed CentralView ArticlePubMedGoogle Scholar
- Meyer E, Raupach R, Lell M, Schmidt B, Kachelriess M: Normalized metal artifact reduction (NMAR) in computed tomography. Med Phys 2010, 37: 5482-5493. 10.1118/1.3484090View ArticlePubMedGoogle Scholar
- Lell MM, Meyer E, Kuefner MA, May MS, Raupach R, Uder M, Kachelriess M: Normalized metal artifact reduction in head and neck computed tomography. Invest Radiol 2012, 47: 415-421. 10.1097/RLI.0b013e3182532f17View ArticlePubMedGoogle Scholar
- Meyer E, Raupach R, Lell M, Schmidt B, Kachelriess M: Frequency split metal artifact reduction (FSMAR) in computed tomography. Med Phys 2012, 39: 1904-1916. 10.1118/1.3691902View ArticlePubMedGoogle Scholar
- Bal M, Spies L: Metal artifact reduction in CT using tissue-class modeling and adaptive prefiltering. Med Phys 2006, 33: 2852-2859. 10.1118/1.2218062View ArticlePubMedGoogle Scholar
- Kachelriess M, Watzke O, Kalender WA: Generalized multi-dimensional adaptive filtering for conventional and spiral single-slice, multi-slice, and cone-beam CT. Med Phys 2001, 28: 475-490. 10.1118/1.1358303View ArticlePubMedGoogle Scholar
- Wang G, Frei T, Vannier MW: Fast iterative algorithm for metal artifact reduction in X-ray CT. Acad Radiol 2000, 7: 607-614. 10.1016/S1076-6332(00)80576-0View ArticlePubMedGoogle Scholar
- Morsbach F, Wurnig M, Kunz DM, Krauss A, Schmidt B, Kollias SS, Alkadhi H: Metal artefact reduction from dental hardware in carotid CT angiography using iterative reconstructions. Eur Radiol 2013, 23: 2687-2694. 10.1007/s00330-013-2885-zView ArticlePubMedGoogle Scholar
- Morsbach F, Bickelhaupt S, Wanner GA, Krauss A, Schmidt B, Alkadhi H: Reduction of metal artifacts from hip prostheses on CT images of the pelvis: value of iterative reconstructions. Radiology 2013, 268: 237-244. 10.1148/radiol.13122089View ArticlePubMedGoogle Scholar
- Zhang Y, Yan H, Jia X, Yang J, Jiang SB, Mou X: A hybrid metal artifact reduction algorithm for x-ray CT. Med Phys 2013, 40: 041910. 10.1118/1.4794474View ArticlePubMedGoogle Scholar
- Watzke O, Kalender WA: A pragmatic approach to metal artifact reduction in CT: merging of metal artifact reduced images. Eur Radiol 2004, 14: 849-856. 10.1007/s00330-004-2263-yView ArticlePubMedGoogle Scholar
- Lemmens C, Faul D, Nuyts J: Suppression of metal artifacts in CT using a reconstruction procedure that combines MAP and projection completion. IEEE Trans Med Imaging 2009, 28: 250-260.View ArticlePubMedGoogle Scholar
- Garton SD, Hilton J, Oku H, Crouse BR, Rajagopalan KV, Johnson MK: Active site structures and catalytic mechanism of rhodobacter sphaeroides dimethyl sulfoxide reductase as revealed by Resonance Raman Spectroscopy. Journal of the American Chemical Society 1997, 119: 12906-12916. 10.1021/ja972109lView ArticleGoogle Scholar
- Arendsen AF, Hadden J, Card G, McAlpine AS, Bailey S, Zaitsev V, Duke EHM, Lindley PF, Kröckel M, Trautwein AX, Feiters MC, Charnock JM, Garner CD, Marritt SJ, Thomson AJ, Kooter IM, Johnson MK, van den Berg WAM, van Dongen WMAM: The "prismane" Protein Resolved: X-ray Structure at 1.7°A and Multiple Spectroscopy of two Novel 4Fe Clusters. Berlin: Springer; 1998.Google Scholar
- Pappuru RR, Briceno C, Ouyang Y, Walsh AC, Sadda SR: Clinical significance of B-scan averaging with SD-OCT. Ophthalmic Surg Lasers Imaging 2012, 43: 63-68. 10.3928/15428877-20110908-02View ArticlePubMedGoogle Scholar
- Flentje M, Zierhut D, Schraube P, Wannenmacher M: Integration of coronal magnetic resonance imaging (MRI) into radiation treatment planning of mediastinal tumors. Strahlenther Onkol 1993, 169: 351-357.PubMedGoogle Scholar
- Mirowitz SA, Lee JK, Brown JJ, Eilenberg SS, Heiken JP, Perman WH: Rapid acquisition spin-echo (RASE) MR imaging: a new technique for reduction of artifacts and acquisition time. Radiology 1990, 175: 131-135.View ArticlePubMedGoogle Scholar
- Baus WW: Wasseräquivalente Materialien und ihre Eigenschaften im Bestrahlungssystem. In 44 Jahrestagung der Deutschen Gesellschaft für Medizinische Physik. Edited by: Treuter H. Cologne; 2013:639-640.Google Scholar
- Hubbell JH, Seltzer SM: Tables of x-ray mass attenuation coefficients and mass energy absorption coefficients from 1 keV to 20 MeV for elements Z = 1 to 92 and 48 additional substances of dosimetric interest. National Institute of Standards and Technology (NIST) Standard Reference Database, vol. 126 1996. http://www.nist.gov/pml/data/xraycoef/Google Scholar
- Gammex 467 product specification [http://www.gammex.com] 
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.