Metal Artifact Reduction in CT: Where Are We After Four Decades?

This article has been accepted for publication in a future issue of this journal, but has not been
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2016.2608621, IEEE Access
Metal Artifact Reduction in CT:

Where Are We After Four Decades?
Lars Gjesteby1, Bruno De Man2*, Yannan Jin2, Harald Paganetti3, Joost Verburg3, Drosoula
Giantsoudi3, and Ge Wang1*
1
Biomedical Imaging Center, Department of Biomedical Engineering, Rensselaer Polytechnic
Institute, Troy, New York
2
General Electric Global Research Center, Niskayuna, New York
3
Department of Radiation Oncology, Massachusetts General Hospital and Harvard Medical
School, Boston, Massachusetts
*Corresponding Authors: deman@ge.com and ge-wang@ieee.org
ABSTRACT
Methods to overcome metal artifacts in computed tomography (CT) images have been researched
and developed for nearly 40 years. When x-rays pass through a metal object, depending on its size
and density, different physical effects will negatively affect the measurements, most notably beam
hardening, scatter, noise, and the non-linear partial volume effect. These phenomena severely
degrade image quality and hinder the diagnostic power and treatment outcomes in many clinical
applications. In this review, we first introduce the fundamental causes of metal artifacts, categorize
metal object types, and present recent trends in the CT MAR literature. To improve image quality
and recover information about underlying structures, many methods and correction algorithms
have been proposed and tested. We comprehensively review and categorize these methods into six
different classes of metal artifact reduction (MAR): metal implant optimization, improvements to
the data acquisition process, data correction based on physics models, modifications to the
reconstruction algorithm (projection completion and iterative reconstruction), and image-based
post-processing. The primary goals of this review are to identify the strengths and limitations of
individual MAR methods and overall classes, and establish a relationship between types of metal
objects and the classes that most effectively overcome their artifacts. The main challenges for the
field of MAR continue to be cases with large, dense metal implants, as well as cases with multiple
metal objects in the field of view. Severe photon starvation is difficult to compensate for with only
software corrections. Hence, the future of MAR seems to be headed toward a combined approach
of improving the acquisition process with dual-energy CT, higher energy x-rays, or photon-
counting detectors, along with advanced reconstruction approaches. Additional outlooks are
addressed, including the need for a standardized evaluation system to compare MAR methods.
2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2016.2608621, IEEE Access
1. INTRODUCTION greatly attenuate low-energy x-rays. As the

Ever since the first computed tomography beam hardens, the physics of transmission
(CT) scan in the early 1970s[1], the effort to change. The remaining photons with higher
improve image quality for clinical average energy have subsequent attenuation
applications has been persistent. A long- dominated by Compton scattering, which
standing culprit in the degradation of image alters the path of x-rays and causes
quality is the metal object, such as a dental transmitted photons to hit the detector off the
filling, artificial hip, spine implant, or centerline of the incident beam [6]. Both
surgical clip. Metal objects in the field of beam hardening and scatter result in more
view will strongly attenuate x-rays or even photons reaching areas of the detector where
completely block their penetration, resulting they normally would not be, leading to an
in corrupt or missing projection data received underestimation of the attenuation
by the detector. When an image is coefficient. Hence, a metal object leads to
reconstructed using this incomplete data, it dark bands or streaks in the image along the
leads to unnatural changes in appearance, axis of greatest attenuation, typically
known as artifacts, which are often observed sandwiched by bright overshoots, while the
as bright or dark streaks in the image. Errors metal object itself will still show up white.
in CT number throughout the image affect Noise also contributes to metal artifacts in
diagnostic ability and hinder accurate CT. Photon flux measured by CT detectors
distinction of tissue types. This is particularly approximately follows a Poisson distribution,
detrimental in radiation therapy planning for so low photon counts will cause higher
cancer treatment, where inexact tumor relative statistical errors in associated regions
localization and characterization of of the detector [7]. These errors are random,
surrounding tissues can propagate to dose and therefore present as thin dark and bright
calculation errors that severely affect streaks in the image [4].
treatment success [2]. Another clinical area in The fourth significant cause of metal
which metal artifacts are a major impediment artifacts is the NLPV effect. This arises when
is orthopedics, due to the high image quality the edge of a metal object straddles certain
that is required very close to metal implants projection lines, causing a variation in the
[3]. To overcome artifacts in CT caused by attenuation coefficient perpendicular to the x-
metal objects, extensive research and ray propagation direction. Since the intensity
development efforts have been devoted to of this region measured by the detector is no
metal artifact reduction (MAR) over the past longer a linear function of attenuation,
four decades. inconsistencies arise in the projection data,
Metal artifacts are common in clinical resulting in density estimation errors in the
images and are caused by several image [8].
mechanisms, most prominently beam An additional artifact cause that stems
hardening, scatter, noise, and the non-linear from data sampling rather than data
partial volume (NLPV) effect [4], [5]. Beam inconsistency is aliasing. This occurs when
hardening occurs in polychromatic x-ray the Nyquist-Shannon sampling theorem is
beams when the average beam energy not satisfied and the measurement interval
increases (“hardens”) due to lower-energy between projections is too large.
photons being more easily absorbed by Theoretically, projection data should be
matter. This phenomenon is more sampled at half of the detector cell width or
pronounced in the presence of high atomic less to avoid aliasing [7]. Undersampling will
number materials, including metals, which lead to overlap in the measured frequency
spectra, and high-frequency components will that they more accurately represent the ideal
not be discernible, leading to streaks near line integrals. If the quality of the data is so
object edges. While some low-level aliasing poor that physics corrections are ineffective,
may always exist, it only presents as visible (4) complete replacement of projections can
image artifacts when associated with high- be performed, or an (5) iterative
density objects, especially metal objects. reconstruction can be used that down-weights
There are many existing MAR techniques or ignores the corrupt data. Lastly, if raw
that address the above-mentioned artifact projection data is not accessible, (6) MAR
causes in various ways. In this review, we can be implemented through post-processing
perform a thorough review and classification of the reconstructed images, such as filtering
of MAR methods from journal articles, or normalization. There are several hybrid
conference proceedings, and patents over the procedures that combine techniques and
past 40 years. Although not every single algorithms from multiple categories, but we
approach is included, we have selected nearly classify these in the section that best defines
all contributions of novel techniques that are the primary innovation of the method.
well-known in the field, or at least were at This review aims to provide a
their time of introduction, based on the comprehensive presentation and discussion
number of citations and the publication of existing MAR techniques and algorithms,
journal/conference. Our classification spanning from 1978 to the present. The
scheme defines six major classes of purpose of this endeavor is to define the
procedures: (1) The most obvious way of current limitations and offer insight into the
reducing metal artifacts in CT is to prevent future directions of the field for overcoming
them from occurring in the first place, which these challenges. An overview of trends in
can be achieved by using non-metal implants the literature (Figs. 1, 2, 3) gives a general
or removing metal from patients before idea of where the field is headed. The total
scanning. Unfortunately, this is not a realistic number of annual publications related to CT
option in most cases, so correction further MAR in 2015 has nearly tripled since 2010,
into the imaging process is needed. (2) The as shown in Fig. 1. An increase is certainly
next chance for minimizing artifacts is in the expected due to the “publish or perish”
data acquisition phase, where modifications mentality in academia, but nevertheless the
can be made to x-ray tube parameters, the upsurge is substantial. This growth in
detector, and the scan geometry. Multiple research indicates the large effort towards
energies can also be employed to obtain adequate MAR solutions for all applications.
information at different x-ray spectra. In particular, since 2010, more MAR papers
Although these adjustments increase the have been published related to hip and dental
intrinsic fidelity of the data, MAR with only implants than spine and cardiac applications,
scan acquisition adjustments still does not as illustrated in Fig. 2. Very recently, there
yield sufficient image quality in many has been an uptick in the amount of MAR
clinical applications, so (3) the raw data must papers related strictly to the topic of radiation
be corrected and/or the reconstruction therapy, which shows it is gaining
algorithm must be improved. These momentum as an area in need of better
correction- and reconstruction-based MAR techniques for improved image accuracy.
approaches are by far the most widely Roughly 20% of the CT MAR publications in
researched. Models of beam hardening, 2015 were concerned with radiation therapy
scatter, noise, and other physical causes of applications, which is second only to dental
artifacts can be applied to correct raw data so implants (24%).
monoenergetic extrapolation” by Bamberg et

al. (2011) is a dual-energy CT (DECT)
protocol that overcomes beam hardening
artifacts and generates monoenergetic
images.
Figure 1. Total annual publications on CT MAR

during the period of 1995 to 2015. Data obtained
from Web of ScienceTM using keywords “CT metal
artifact reduction.”
Figure 2. Annual publications on CT MAR for hip,
dental, radiation therapy, spine, and cardiac
A look at the most-cited paper in each of applications from 2010 to 2015. Data obtained from
four MAR method categories can reveal Web of ScienceTM using keywords “CT metal
information on the technical trends. The artifact reduction” AND the field, such as “hip.” It
should be understood that the same paper may be
most-cited Projection Completion publication is
counted among multiple categories if it addresses
“Reduction of CT artifacts caused by metallic multiple applications.
implants” by Kalender et al. (1987), which is
a benchmark linear interpolation algorithm to
which many new MAR methods are
compared. Among Iterative Reconstruction
algorithms, “Iterative deblurring for CT metal
artifact reduction” by Wang et al. (1996) has
the most citations, as a method that explicitly
discards data through the metal object and
estimates new projections with a
deterministically interpreted expectation
maximization formula. The most heavily
cited Physics-based Pre-processing MAR paper
is “Generalized multi-dimensional adaptive
filtering for conventional and spiral single- Figure 3. The most-cited paper in each of four MAR
slice, multi-slice, and cone-beam CT” by method categories (with total citation numbers
Kachelriess et al. (2001), which employs indicated), including Acquisition Improvement
adaptive filtering to suppress noise. Finally, (Bamberg et al.), Physics-based Pre-processing
(Kachelriess et al.), Projection Completion
the Acquisition Improvement category is (Kalender et al.), and Iterative Reconstruction
represented by a paper that managed to reach (Wang et al.). Data obtained from Web of S
100 citations within four years after cienceTM.
publication. “Metal artifact reduction by dual
energy computed tomography using
All four publications are discussed in the size and material of the implant, different
further detail in their respective category degrees of x-ray attenuation and physics
sections. Fig. 3 reflects the total number of effects will occur. For example, small objects
citations for these well-regarded papers. with low density, such as surgical clips, may
While surveying citation data, it was noted only cause minor beam hardening or scatter,
that iterative reconstruction algorithms that which can be solved by a physics-based pre-
account for a physical cause of artifacts have processing correction of the data. Larger
gained more recognition over the past five and/or denser implants, such as prosthetic
years, while interpolation algorithms may be hips or dental fillings, will likely require
falling out of favor for complex applications. modification of the reconstruction algorithm,
Additionally, there has been a rapid rise in either by a projection completion method or
popularity of DECT for MAR. statistical iterative reconstruction. Table 1
defines five categories of metal objects, their
2. METAL OBJECTS associated physical and image effects, and
A wide range of metal objects leads to a wide the most relevant MAR approaches.
range of CT image artifacts. Depending on
Table 1. Categorization of metal implants detailing the types of artifacts and effective techniques for image correction.
Adapted from Jin et al. [9] with permission.
Clinical MAR
Category Definition Examples Physics Effects Image Effects Approaches
Small metal Minor shading in

Surgical Minimal beam Physics-based
I object, low immediate vicinity
clip hardening/scatter correction
attenuation of metal object
Moderate
Medium metal Spinal Beam hardening, bright/dark bands Physics-based
object, cage, scatter, noise, but spanning along correction,
II
medium spinal no complete the directions of Improved
attenuation screw photon starvation highest acquisition
attenuation
Small Severe bright/dark Projection
Small metal
dental Complete photon streaks fanning completion,
III object, high
filling, starvation out from metal Iterative
attenuation
bullet object reconstruction
Medium/large Severe bright/dark Physics-based
Beam hardening,
metal object, streaks and thick correction +
scatter, noise,
portion of the Hip bright/dark bands projection
IV and some data
beam with prosthesis over a substantial completion or
with complete
high fraction of the iterative
photon starvation
attenuation field-of-view reconstruction
Large Thick bright/dark
Large metal and/or bands obliterating Iterative
V object, high multiple Photon starvation a substantial reconstruction +
attenuation dental portion of the prior knowledge
fillings image
3. METAL ARTIFACT REDUCTION that employ techniques in more than one of

METHODS the overall categories are discussed within
The classification scheme in this review the class that best defines their novelty. The
defines six major groups of MAR techniques: following is an in-depth description of each
Metal Implant Optimization, Acquisition of these groups and their subclasses, and a
Improvement, Physics-based Pre-processing, detailed presentation of the published
Projection Completion, Iterative methods that comprise them.
Reconstruction, and Image Post-processing.
The methods that primarily alter the 3.1 Metal Implant Optimization
reconstruction algorithm are discussed under The simplest way of reducing metal artifacts
Projection Completion and Iterative in a CT scan is to prevent the presence of
Reconstruction. A standard CT scan acquires metal in the field of view. For clinical
projection data of an object at different angles applications, this can be crudely achieved by
and reconstructs the measurements into an removing metal from the patient or opting for
image using filtered back-projection (FBP) or a non-metal implant in the first place. In
an iterative estimation. The sinogram is the practice, these options are often not realistic
collection of all line attenuation due to the invasive surgery required for
measurements acquired for an image. Since implant removal or inadequate metal
the number of individual measurements is alternatives. However, a study by Gray et al.
finite, a theoretically ideal image cannot be does show patient cases in which dental
obtained, but a suitable approximation can be fillings were extracted for radiation planning
generated using the available data. A metal of head and neck cancers [10]. The streak
object in the imaging region causes errors in artifacts were completely removed by this
the projection data mainly due to beam extraction, and the ability of physicians to
hardening, noise, scatter, or the NLPV effect. delineate anatomical regions of the head were
Reconstruction using corrupt or incomplete greatly improved.
data will result in an image with artifacts. To Beyond extraction, the majority of studies
eliminate these artifacts, many MAR have focused on assessing image artifacts
algorithms modify the process by which the produced by different types of metal to
image is reconstructed. A common approach identify the optimal implant with the least
is to directly correct or replace the corrupt attenuation. Weese et al. determined that for
projection data in the sinogram, such as by surgical clips used in hemostasis, those made
interpolation. Another method is to ignore or of titanium and stainless steel produced
statistically down-weight data affected by the significantly fewer artifacts than did clips
metal object and employ an iterative made of tantalum [11]. Further, plastic clips
procedure to arrive at an image using the caused no visible artifacts, so they would be
more reliable data. Each of these techniques ideal from an imaging standpoint, but are
can be improved by pre-processing the data limited by ineffective long-term utility.
with a model that accounts for a physical Ebraheim et al. later compared stainless steel
cause of the artifacts. Sometimes this and titanium in the context of screws and bars
correction step is enough by itself and for pelvic fixation [12]. Their results revealed
reconstruction modifications are not even that titanium implants substantially reduced
necessary. Prior knowledge is an additional image artifacts compared to stainless steel
consideration that can be applied to hardware while providing the same level of
projection completion and iterative fixation and quality of fracture healing. A
reconstruction methods. MAR procedures comparison between titanium and cobalt-
chrome artificial hips by Haramati et al. with medium-level characteristics. The

further confirmed that titanium produced image artifacts of vitallium remain a
superior image quality, especially in the area limitation.
directly surrounding the bone [13]. Fig. 4 In addition to optimizing metal types for
compares CT image artifacts from rods made reduced artifacts, an exploration into the
of stainless steel, titanium, and vitallium, complementary imaging contrast agents used
which is an alloy of cobalt, chromium, and in DECT was done by Lambert et al. [15].
molybdenum that is a candidate to replace Their work sought contrast enhancement
titanium. The characteristics of these materials that are more compatible for higher
materials relative to each other are displayed energy x-rays. Since one of the x-ray energy
in Table 2. The stiffness of titanium is low spectra used in DECT is typically higher than
compared to stainless steel, which can be a the x-ray spectrum of standard CT scans, the
positive or a negative depending on the photons can more easily penetrate metal and
application. Either way, its higher strength lead to fewer inconsistencies in the data.
and biocompatibility make titanium an Unfortunately, as beam energy increases, the
attractive implant option. The titanium rods signal from contrast agents such as iodine or
also produced the lowest degree of artifacts barium diminishes due to their low k-edges.
as seen in Fig. 4. Vitallium has a stiffness, Thus, contrast materials with higher k-edges
strength, and biocompatibility that ranks in are more effective with higher average
between those of stainless steel and titanium, energy beams.
which gives the material more versatility
Figure 4. Standard clinical CT images of a torso cadaver containing rods of 5.5 mm diameter stainless steel (left),
4.5 mm diameter titanium (middle), and 4.5 mm diameter vitallium (right). Reprinted from Knott et al. [15] with
permission from Elsevier.
Table 2. Mechanical characteristics of patient implant materials relative to each other. Adapted from Knott et al.
[15] with permission from Elsevier.
Characteristic Stainless Steel Titanium Vitallium
Stiffness High Low Medium
Strength Medium High Medium
Corrosion resistance Low High Medium
Biocompatibility Low High Medium
It was shown that tantalum, tungsten, and gantry angles. A head tilt routine was
bismuth preserved contrast signal much employed by Brown et al. for patients with
better than barium and iodine over the 40 to aneurysm clips so that the axial plane through
140 keV photon energy interval. Barium and the clip projected away from the cluster of
iodine signals declined 91-97%, while the major cerebral arteries to be evaluated [18].
higher k-edge group exhibited just a 21-32% This method showed that metal artifacts
reduction in CT number, supporting its could be minimized in the most important
potential for use in patients during DECT to regions of the image by altering the angle of
complement metal artifact reduction. acquisition. A study by Lewis et al. later
Recent exploration has identified explored gantry tilt angles of 5 to 15 degrees
magnesium screws as resorbable implants for imaging total knee prostheses [19]. Their
that can interface tissue with bone, and then results showed that gantry angles of 10 to 15
biodegrade [16]. This material has been degrees provided the most accurate mean
shown to maintain its shape and strength attenuation values around the implant. Lakits
while promoting bone growth in mice, with et al. then compared helical CT in the axial
mature bone forming around the implant after plane, as well as multiplanar reconstruction
three months. Future studies will look at of coronal and sagittal images, with
controlling the rate of degradation of the conventional CT acquisition in the axial and
material, which could shape the future coronal planes [20]. They determined that a
direction of patient-level metal artifact helical acquisition allowed significantly
reduction. Zimel et al. has also explored better localization of metal objects and
carbon fiber nails as an alternative to titanium reduced beam hardening artifacts, mainly due
nails for fixating large bones, and found that to the sagittal plane information, which is not
carbon fiber produced fewer CT image obtainable with conventional CT.
artifacts [17]. In the context of CT angiography for
Implant optimization within the patient is evaluating vessels around aneurysm clips
a basic starting point for MAR, but further after placement, Mamourian et al. found that
reduction measures are necessary to achieve axial imaging of titanium and cobalt-chrome
acceptable levels of image quality. implants led to diminished streak artifacts
when compared to helical scanning [21].
3.2 Acquisition Improvement However, for most applications, the
A second class of strategies is to alter the CT improvement in image quality with axial
scan acquisition parameters such as x-ray acquisition is not worth the loss of volumetric
tube voltage, tube current, scan plane, and coverage provided by helical CT. One unique
slice thickness. A more drastic approach is to method presented by Tolakanahalli et al. is to
employ dual-energy protocols that acquire obtain digital subtraction angiography-type
data using two x-ray spectra for material images using CT with contrast agents [22].
decomposition. Spectral CT with photon- Topographic images are obtained at different
counting is an even different method with the angles and the metal objects are subtracted
ability to resolve multiple energies from a out of the projection data to allow better
single spectrum. visualization close to the metal surface.
Other variations to the scan acquisition
3.2.1 Parameter Adjustment process have been employed to combat metal
Straightforward techniques to improve data artifacts. Link et al. extended the CT scale
acquisition are to simply scan around the from a maximum window of 4,000 to 40,000
metal object or obtain data from preferred Hounsfield units (HU) to decrease artifacts
from femur implants and improve diagnostic current variation did not have a significant
power [23]. This extension is beneficial effect.
because metal objects with high attenuation The x-ray energy can be increased even
coefficients have CT numbers from 8,000 to more dramatically with a linear accelerator.
20,000 HU, and their values are cut off on Schreiner et al. acquired images of hip
standard scanners, resulting in distortion implants using megavoltage CT with photon
[24]. It was also found that decreasing slice energies up to 4 mega-electron-volts (MV),
thickness from 5 mm to 2 mm increased which significantly reduced artifacts around
diagnostic scores when combined with the the metal surface [27]. The drawback of this
extended scale. Moon et al. further revealed technique is increased radiation dose, and so
that smaller slice thicknesses as thin as 0.75 in an effort to reduce the megavoltage
mm led to greatly reduced metal artifacts exposure, Wu et al. integrated select views
from stainless steel implants in the femur using MV photon energies into a standard
[25]. kilovoltage acquisition [28]. The high-
A more aggressive approach that has been voltage data replaces the metal-affected
achieved with ongoing engineering projections in the standard-voltage sinogram
advancements is to increase the voltage so that information lost due to photon
(kVp) and/or current (mA) of the x-ray tube starvation is restored. In the context of
so that either higher energy photons or a tomotherapy, Jeon et al. employed a similar
greater number of photons are generated to protocol and showed that images
penetrate the metal object and surrounding reconstructed from hybrid sinograms
tissue. These modifications inherently containing MV and kV x-ray data had higher
increase the quality of the raw data. The contrast-to-noise and signal-to-noise ratios
Moon et al. study also exhibited that higher than those from just kV or MV sinograms
tube voltage (140 vs 120 kVp) improved alone [29]. This, along with decreased metal
artifacts, but higher tube current (500 vs 300 streak artifacts, allowed for better distinction
vs. 100 mA) did not have a significant effect of relevant structures around implants.
on the image quality. Lee et al. assessed With each adjustment to the scanning
various protocols for imaging the spine after procedure that improves metal artifact
pedicle screw placement [26]. The tube reduction, there is a disadvantageous side
voltage was varied between 80, 100, and 120 effect. This usually comes in the form of
kVp, and the tube current was applied in the increased radiation dose to the patient or loss
range of 60 to 220 mA. As expected, higher of image contrast. Table 3 presents the effects
voltages led to fewer metal artifacts, but the and disadvantages of the acquisition
adjustments discussed in this section.
Table 3. Effects of adjustments to scan acquisition procedure for reducing metal artifacts. Adapted from Lewis
et al. [30] with permission from Springer.
Acquisition Adjustment Primary Advantage Primary Disadvantage
Reduce beam hardening and
Increase photon energy (kVp) Increased patient dose
noise
Increase tube current (mA) Reduce noise Increased patient dose
Lower pitch setting Reduce noise Increased patient dose
Decrease slice thickness Reduce partial volume artifact Increased noise
Extend CT scale Increased window width & level Decreased contrast resolution
3.2.2 Dual-energy CT conclusions when extrapolating single

To overcome beam hardening effects, dual- energy images from a dual-energy protocol
energy CT has been effective because it with 140 and 100 kVp scans [32]. For
allows for monoenergetic data sets to be assessing metal hardware in hip and upper leg
extracted from scans acquired at two separate bone injuries, the extrapolated energy range
polychromatic beam spectra. With of 100 to 130 keV produced the best
attenuation coefficients obtained at two diagnostic image quality. The median streak
energies, monoenergetic images can be intensities decreased from 413 HU at 88 keV
synthesized, often referred to as virtual to 153 HU at 120 keV, which resulted in
monochromatic spectral (VMS) images. In improved diagnostic quality of the images.
one of the most highly cited MAR papers Fig. 5 shows extrapolated VMS images with
over the past five years, Bamberg et al. artifacts decreasing as photon energy
acquired images of metal screws in patients increases. However, there are drawbacks at
at 100 and 140 kVp, and then extrapolated the higher energies, including higher radiation
energies of the standard 100 and 140 kVp dose to the patient. Additionally, Pessis et al.
spectra [31]. This method substantially showed that image contrast declined beyond
reduced artifacts by 48.6% at the optimal a certain energy level, such as 140 keV, and
119.5 keV energy setting, while increasing that lower energies around 80 keV were
diagnostic assessment of the metal surface better for detailing soft tissues [34]. A law of
and surrounding region. They also concluded diminishing returns in VMS reconstruction
that 105 keV was an optimal extrapolated was put forth by Lewis et al. in the context of
energy for visualizing screws. hip implant imaging [30]. They extracted 16
A later dual-energy CT study by Zhou et single photon energy images between 40 and
al. extracted higher photon energies, ranging 190 keV in steps of 10 keV. The results
from 70 to 190 keV, and determined that 130 demonstrated that streak artifacts could be
keV was optimal for reducing metal artifacts reduced with minimal loss of spatial
[33]. Meinel et al. reached similar resolution up to 150 keV. Above this
Figure 5. Metal artifacts decrease as photon energy increases in single energy extrapolated images from dual-energy
CT. Reprinted from Meinel et al. [32] with permission from Wolters Kluwer.
threshold, metal artifact reduction is no more seen in Fig. 6. The proposed method also
effective, and contrast-to-noise ratio repeatedly produced attenuation values of
decreases. This trend is consistent at tube teeth and implants that closely matched with
currents of 100, 200, and 400 mA. theoretical estimates.
A statistical framework to model noise in
dual-energy CT was proposed by Sukovic
and Clinthorne [35]. A penalized weighted
least squares (PWLS) objective function with
constraints in the density domain handles the
non-Poisson noise from amorphous silicon
(aSi:H) detectors, and is minimized using the
Gauss–Seidel algorithm. This method is
compared to the dual-energy FBP
counterpart, in terms of bias/standard
deviation, and demonstrates advantages in
high noise cases, specifically at the low flux
rate of 3 x 105 photons/second/detector
element (5mm).
3.2.3 Photon-Counting CT
More recently, photon-counting detectors
have been considered as a way to reduce
metal artifacts. Specifically, photon-counting
detectors in charge-summing mode are
capable of discriminating individual photon
energies for spectral reconstruction and are
more efficient at very low photon count Figure 6. Reconstruction results of a dental
phantom containing one tooth with a circular metal
levels. Rajendran et al. employed a filling. The novel spectral-driven iterative
Medipix3RX detector to acquire spectral CT reconstruction (SPIR) is compared with FBP and
images of titanium and magnesium scaffolds the penalized maximum likelihood iterative
at different energy ranges [36]. Data is reconstruction method without prior information
reconstructed using the simultaneous (IR). The right column is a zoomed view of the
implant in the left column. Reprinted from
algebraic reconstruction technique (SART) Nasirudin et al. [37].
that accounts for dead pixels in the detector
with oversampling compensation. Images
depicting the narrow high energy range of 50 3.3 Physics-based Pre-processing
to 80 keV yielded the best metal artifact Physics-based pre-processing techniques aim
reduction while preserving contrast. to model the physical cause of metal artifacts,
Nasirudin et al. also used a photon- such as noise, scatter, beam hardening, and
counting detector for spectral CT material the NLPV effect, and correct data in the metal
decomposition [37]. This information is trace to improve reconstruction. These
incorporated into a penalized maximum corrections are usually applied in the
likelihood reconstruction. Their results projection domain before reconstruction,
showed significant reduction in bright and although the same physics models can
dark streaks compared to FBP, while alternatively be incorporated in the forward
structure near metal edges was preserved, as
model of an iterative reconstruction technique by correcting for beam hardening

algorithm. first and then linearly interpolating corrupt
projections [44]. Verburg and Seco presented
3.3.1 Noise Suppression a novel beam hardening correction for low
Hsieh modeled local noise characteristics atomic number implants, such as titanium,
with an adaptive mean filter [38]. This that compares projections through the metal
reduces quantum noise artifacts and streaks with neighboring unaffected projections [45].
while attempting to preserve spatial For higher atomic number implants,
resolution. Rangayyan and Gordon also additional iterative reconstruction based on
combined adaptive filtering with algebraic total variation (TV) regularization was
reconstruction (ART) to suppress streak applied. Fig. 7 presents the proposed physics
artifacts [39]. A multidimensional adaptive correction technique on a titanium spinal
filtering approach was developed by implant compared with other common MAR
Kachelriess et al. to apply 3D nonlinear methods. Higher order beam hardening
filters in the raw data domain [5], [40]. artifacts were approached by Schüller et al.
Resolution tradeoffs are smaller as compared without segmentation of the image [46]. To
to one-dimensional smoothing approaches, as determine where beam hardening correction
demonstrated with patient data from single- should be applied, a histogram deformation
and multi-slice CT to reduce image noise or of the original image is obtained to produce a
patient dose. Image noise along most nonlinear CT value distribution.
attenuating paths is effectively reduced by For scatter reduction, Meyer et al.
30%–60% in non-cylindrical body regions, proposed an empirical scatter correction
such as through the shoulder, with resolution (ESC) algorithm that makes the assumption
loss below 5%. This method also helps that a linear combination of the uncorrected
reduce metal artifacts in the hip region. image with various basis images will be
An improved multi-dimensional adaptive scatter-free [47]. The coefficients for the
filtering approach was presented by Watzke, linear combination are calculated by
where the adaptive filtering strength was maximizing flatness criteria.
adjusted in the shadow of the metal object
[41]. Another noise reduction method for hip 3.4 Projection Completion
implant images using directional filtering Oftentimes, data in the metal trace are
was implemented by Prell et al. [42]. Boas completely corrupt or missing, so new
and Fleischmann developed two techniques, projection data must be synthesized to
metal deletion technique (MDT) and complete the sinogram. One way to achieve
selective ART, which both incorporate an this is by interpolating replacement values,
edge-preserving blur filter to reduce noise eitehr from neighboring projections or from a
and streaks [43]. Selective ART also includes mathematical model. We have grouped these
a model for noise and beam hardening as Interpolation methods. A second
effects. MDT involves forward projection projection completion approach that
after the filter, which guides the sinogram increases accuracy is to incorporate prior
completion in metal areas. knowledge to guide the estimation of the data
that replaces corrupt projections. This prior
3.3.2 Scatter and Beam Hardening image is reprojected (forward projected) to
Correction generate projection data for sinogram
Seitz and Rüegsegger improved an completion, and thus these types of
interpolation projection completion procedures are classified as Reprojection.
Figure 7. Titanium spine implant images show that the proposed beam hardening correction algorithm is more
effective than projection replacement techniques alone in reducing streaks and preserving anatomy in the image.
Reprinted from Verburg and Seco [45] by permission of IOP Publishing.
To improve projection completion projection is hollow, the density distribution

processes, some algorithms employ a is symmetric, or the projection does not
normalization step that compares raw change greatly in the interpolated section.
sinogram data to a prior image sinogram. Hinderling et al. applied an interpolation
These techniques are discussed in the technique based on the idea of Lewitt and
Normalization section of this class. Bates to reduce metal artifacts when imaging
Projection completion algorithms are the the bone-cement interface of hip implants
most commonly used and most widely [49]. They localize the edge of the metal
developed form of MAR. object by finding a sharp discontinuity in the
projection gradient, and then create a square
3.4.1 Interpolation wave pseudo-projection between edges of the
The first projection completion method was implant. A linear interpolation of the
published by Lewitt and Bates in 1978. Their measured data plus pseudo-projections is
technique calculates smooth continuations of used to replace the corrupt projection in the
hollow or truncated projections using either implant interval. Once reconstructed, the
polynomial interpolation or by synthesizing image artifacts are suppressed and the
data that satisfies consistency criteria [48]. distribution of cement and bone can be seen,
This criteria is determined by operations on though the homogeneity is not quite accurate.
the Fourier coefficients of the projections. Glover and Pelc implemented polynomial
The interpolation method is much more completion to address motion streak artifacts
straightforward and less computationally from surgical clips [50]. Using a “rubout”
intensive, but its success varies on a case-by- operation, the projections through the metal
case basis, with the highest effectiveness object are removed and replaced with data
occurring when a small fraction of the from polynomial interpolation of
neighboring projections. This allows streaks redetermine the transmission values of rays
to be eliminated from the image without passing through the metal object [54]. Their
affecting other areas. One side effect of the resulting images were superior to those
method is that nonlinear processing can obtained with simple linear interpolation.
introduce new artifacts. Mahnken et al. applied interpolation in a
Seitz and Rüegsegger developed a two-dimensional manner by replacing
consistent completion technique for attenuation values of metal-affected
quantitative CT bone densitometry based on sinogram data with the nearest non-corrupt
linear interpolation [51]. The computation point in the radial direction out of 16 points
time for such a protocol is long in complex [55]. If no reliable point is found, the value is
regions where the solutions to systems of weighted to zero. The original segmented
linear equations are not straightforward, metal image is adaptively mixed with the
although with today’s computing power this FBP of the corrected projection data to
may no longer be a challenge. achieve a final image that maintains the
Kalender et al. sought a simpler algorithm object’s shape. Moseley et al. devised a two-
using linear interpolation [52]. Unlike prior dimensional Taylor series polynomial
methods that segment metal in the sinogram interpolation MAR method that performed
domain, the boundaries of metal pelvic clips well on small fiducial markers in cone-beam
are manually segmented in original corrupt CT [56]. The markers can be completely
images, and this segmentation is forward masked and the values of image voxels in
projected to identify the corresponding which streaks were present can be reduced to
projection data, which are then replaced with 1% of the original magnitude after correction.
interpolated neighboring data. The success of Another polynomial interpolation
this algorithm depends mainly on the technique was employed by Wei et al. for
geometric complexity of the metal objects suppressing metal artifacts near bones [57].
and surrounding anatomy. It was shown to be Their “smoothing-plus-scaling” procedure
effective in pelvic regions, but failed when isolates bone structures and assigns average
used in the skull due to close areas of high CT values to bone pixels in the same region.
and low attenuation. The metal pixels are replaced with values
Bruyant et al. took a different approach to from polynomial interpolation of
reducing artifacts with interpolation by neighboring data. All corrected pixels are
increasing the number of projections that then added back to the original projection
could be used for image reconstruction [53]. data to reconstruct an image with artifacts
Their algorithm is called interpolation of suppressed.
projections by contouring (IPC), and Yazdi et al. developed an adaptive
involves plotting level lines in the sinogram interpolation algorithm that preserves metal
to connect pixels with the same intensity. object edges effectively by finding the
Interpolation is used to fill in values at all projection data associated with each edge and
pixels for each level line plot. The data were maintaining the distance between them [58].
resampled to multiply the number of Linear interpolation is performed between
projections by 2 or 3, which helped reduce these projections to replace the values in the
streak artifacts without increasing scan metal implant region. A median filter is
acquisition time. applied so that outlier projection values are
A more advanced pattern recognition removed in the implant region. The mean
scheme by Morin and Raeside adapted a attenuation values and standard deviation in
nearest-neighbor classification to the metal regions of the corrected image were
reported to match with that of the original prosthetic hip images. In addition, the
image before metal was inserted. technique more effectively removed residual
Veldkamp et al. adapted the same dark streaks.
interpolation technique in their protocol, but Liu et al. used a sinusoidal curve fitting of
the novelty of their approach was in the metal the metal-affected projection data instead of
segmentation by employing a Markov line integrals [64], [65]. Only the projections
random field to identify affected projections that matched the sinusoidal fit are included
[59]. They also incorporated linear for processing, which correspond to the metal
interpolation and Laplacian smoothing for object. These projections are amended by
sinogram completion. Yu et al. implemented subtracting out the attenuation coefficient of
an innovative algorithm that employs a the metal to minimize the intensity that will
mean-shift technique for highly accurate lead to streaks. The new sinogram is
segmentation of metal objects [60]. reconstructed with reduced artifacts. A more
Additionally, once the segmentation is advanced method for accurate structure
forward projected, a feedback-based linear representation and smooth sinogram filling is
interpolation scheme is used that ensures the to use interpolation based on Euler’s elastica
calculated values are not larger than the as implemented by Gu et al. [66]. This
original data they replace. The resulting technique defines a curvature for inpainting
images of aneurysm clips exhibited a 20-40% that treats missing data as an occlusion. The
reduction in artifacts as measured by standard resulting images have smoothly connected
deviation of surrounding tissue CT numbers. edges and junctions, though the computation
Projection completion efforts in the time is greater due to iterative steps.
wavelet domain have produced more Zhang et al. tried a Laplacian diffusion
accurate reconstructions in the image regions filter to interpolate metal regions using
close to metal objects. Zhao et al. presented boundary pixels [67]. The filter is calculated
the first algorithm that performs linear by minimizing the gradient field magnitude
interpolation in segmented metal areas, surrounding the metal. This method is not
transforms the sinogram to the wavelet effective for large or irregularly shaped
domain, and interpolates between wavelet objects. Joemai et al. also employed 2D
coefficients to obtain consistent data [61]. Laplacian interpolation for their correction
This method preserves edges and contrast scheme [68]. Kratz and Buzug applied
while keeping the computation time interpolation using non-equispaced Fast
comparable to FBP. Fourier Transforms (NFFT), which allows
A later technique in the wavelet domain higher dimension interpolations [69]. Prior
was developed by Mehranian et al. that takes knowledge of the metal object can be
advantage of prior wavelet coefficient included in this technique to improve its
information of CT sinograms and recovers accuracy.
missing data by solving a regularized inverse A novel concept of reformatted
problem [62]. It was shown to outperform projections was tested by Yu et al., which
linear interpolation and the NMAR algorithm combines projection data at the same view
based on prior image normalization [63] angle over the full longitudinal scan range in
(discussed in the Reprojection section), as helical CT [70]. Two-dimensional
illustrated in Fig. 8. Specifically, the interpolation is used to fill the metal
proposed technique achieved lower absolute segmented projection based on Delaunay
mean deviation values as compared to triangulation. The method is advantageous
NMAR in the select regions of interest on the for long hip implants because the entire metal
Figure 8. Wavelet-domain interpolation optimized by prior information is more effective than interpolation and
normalization in reducing residual artifacts. Reprinted from Mehranian et al. [62] by permission of IEEE.
region can be captured with reformatted treatment planning in which there are
projections. In a similar method, Li et al. also multiple dental fillings. Xu et al. aimed to
employed reformatted projections in helical improve segmentation with the Steger
CT and applied boundary mapping of metal method that detects curvilinear structures for
regions along with a dual-front active contour precise determination of small metal object
model to track the metal shadow in each edges [74]. Neighboring projections are used
projection [71]. This improved their accuracy to complete missing data.
and led to greater visibility of soft tissue in A TV inpainting method based on a partial
the corrected images. differential equation model was presented by
Tohnak et al. employed a sequential Xue et al. [75]. An active contour model is
substitution scheme for dental CT to swap also used for segmenting the metal projection
corrupt data through fillings with unaffected region. Together, the procedure yields a
adjacent projections [72]. Of course, the smooth projection completion for dual-
success of this technique is dependent on a energy CT that is reconstructed to produce
reliable segmentation of the metal objects. To high quality images for material
overcome this dependence, Yazdi et al. decomposition. The atomic number of water
proposed an opposite view replacement surrounded by metal objects was measured
method to replace corrupt data with values in within 2.3% accuracy with the MAR
the opposite direction of the affected correction, while measurements without
projection [73]. The segmentation is not as MAR yielded 12.8% error. Zhang et al.
crucial for this process because it does not devised a fractional-order TV inpainting
rely on adjacent projections. This is relevant technique that employs a fractional-order
mainly for dental CT or head and neck cancer
gradient parameter for regularization, which A tissue-class model is a segmentation of

handles wide data gaps well [76]. different materials that compose the original
MAR methods for cone-beam CT must image, including bone, soft tissue, air, and
deal with 3D reconstruction. Wang et al. metal, and is forward projected to be used in
calculate the 3D coordinates of metal objects the replacement of corrupt projection data.
and apply a bilinear interpolation to correct Olive et al. developed the first tissue-class
the projection data [77]. Toftegaard et al. model approach for MAR [83]. The initial
developed a method for gold radiotherapy CT image is segmented into material classes,
markers in cone-beam CT that replaces the such as bone, air, metal, and soft tissue, and
markers with linear interpolation, and then each region of the same class is assigned a
reinserts them by 3D trajectory estimation for constant value. This image model is forward
a final volumetric reconstruction [78]. projected to guide the replacement of corrupt
projection data. A difference sinogram is
3.4.2 Reprojection calculated between the class image and
Crawford et al. developed a reprojection- original image, and this data is then
based method that achieves fast computation downscaled and adaptively filtered before
time for reducing metal streak artifacts [79]. final FBP reconstruction. An earlier approach
Their process involves identifying metal was proposed by Watzke that involved
objects in an original reconstructed image segmenting and separately reprojecting bone
and filling those regions with the CT number regions to avoid introduction of new streak
of water. Then, the image is reprojected and artifacts introduced by the standard
those projections are subtracted from the interpolation methods [41].
original image projections. The corrected Bal and Spies also implemented an
sinogram is reconstructed into the final inpainting algorithm aided by an extracted
image. Naidu et al. patented a similar process tissue-class model [84]. First, an image is
based on this sequence of operations [80]. reconstructed and adaptively filtered. Then,
Tuy developed a method that employs beam the image is segmented using a clustering
hardening correction and then a refinement of algorithm. The missing information is
projection data by forward projecting the last estimated via reprojection of the segmented
reconstructed image [81]. The resulting image. In the inpainting stage, either the
images were enhanced by moderate streak original sinogram or reprojected sinogram
reduction, but new artifacts were introduced can be completed with data from the tissue-
around the edges of surgical clips. class model. The reprojected sinogram
Jeong and Ra employed a parallel method is particularly advantageous because
interpolation protocol that involves there is no need for original data. For an
reprojecting a metal-free image and merging optimal completion, the CT number of the
high-pass and low-pass filtered projection metal class is assigned to that of the
data [82]. The high-pass filtered data is surrounding material, and residual offsets at
interpolated in regions where multiple metal the edges are suppressed by correcting the
objects overlap, and this is merged with low- values of the model-derived data segments
pass filtered data in which the metal linearly. This method is compared with the
projections are replaced with linearly simple linear interpolation method and
interpolated values. The merged sinogram is evaluated with clinical images in which
reconstructed using FBP, and the resulting various metal artifacts exist, showing
images have reduced artifacts with minimal significantly improved image quality and
distortion to non-metal regions. organ contour detectability, which is
especially useful for radiation therapy interpolation [87]. The algorithm achieved an
planning. average correction of 1300 HU for CT
Lemmens et al. devised an interesting numbers of metal artifacts, and image noise
algorithm that includes a tissue-class model was reduced by 27%. Fig. 9 displays results
as well as a model for beam hardening and of the 3D interpolation algorithm with and
noise based on maximum a posteriori (MAP) without forward projection of the tissue-class
reconstruction [85]. Sole use of multimodal model.
prior image constraints will prohibit streak Philips Healthcare developed the O-MAR
artifacts, but soft tissue contrast will be method for orthopedic implants that is based
compromised at the same time. Hence, in this on the reprojections of a metal-segmented
method the constrained image is only used image and a tissue-class model derived from
for projection completion. More specifically, the original image with metal removed. All
the prior knowledge is available on tissue pixels in the tissue-class image are set
attenuation coefficients of human tissues. to a single value [88]. An error sinogram is
The typical modes are the attenuation generated from the difference between the
coefficients of air, fat tissue, soft tissue, and tissue-class sinogram and the original
bone. The intensity priors are defined as sinogram. This error sinogram is
Gaussian functions. By using such a prior in reconstructed to create a correction image.
combination with a Markov Gibbs smoothing Karimi et al. make use of a prior image
prior, an image can be produced free of streak derived from segmentation of metal and
artifacts. The algorithm was validated with tissue in the original image [89]. Metal
simulations, phantoms, and patient data, and regions are replaced with CT values of soft
favorably compared with other metal artifact tissue, and the prior image is reprojected to
reduction algorithms including the simple provide an estimation for the projection data
linear interpolation method and iterative that are used to substitute corrupt projections.
reconstruction with a polychromatic model. This method was tested on head images
Prell et al. made use of both interpolation containing dental fillings and other small
and tissue-class model-based reprojection implants, and was effective in preserving
methods to achieve more accurate image anatomical information. For luggage-
reconstructions [86], [87]. An initial three- screening applications of CT, Karimi et al.
dimensional interpolation algorithm is used developed a sinogram completion technique
to replace corrupt metal projections, and the that includes constraints for beam hardening,
sinogram is reconstructed for an intermediate noise, and scatter [90]. A prior-image is
artifact-reduced volume. Next, a tissue-class created by solving a constrained optimization
model is obtained from segmentation that problem, and the size of this problem is
distinguishes air, soft tissue, and bone in this reduced by using smaller artifact-only
volume, and is then reprojected to directly images. After the optimization of projections,
substitute missing attenuation values in the the small images are up-sampled to generate
projection data. A final reconstruction yields the prior-image, which is then reprojected
a clean, corrected image. A normalization of with metal regions segmented. The sinogram
the raw data relative to the forward projection replacement is completed according to the
data is also applied in one of the algorithms method of Naidu et al. [80].
to preserve edge information after
Figure 9. Forward projection of a tissue class model combined with 3D interpolation MAR improves image quality
better than either method alone. Arrows 1 and 2 illustrate the reduction of the secondary artifacts and better imaging
of structures close to the implants. Arrow 3 depicts a clear reduction of interpolation-based artifacts by using the
forward projection-based correction. Reprinted from Prell et al. [86] by permission of IOP Publishing.
A more accurate prior image for numbers [92]. The reprojection of this image
reprojection was sought by Wang et al. in is used to complete the corrupt sinogram
their fusion prior-based MAR scheme [91]. data. Heußer et al. proposed a prior-based
This algorithm first corrects the original algorithm that utilizes a planning scan of the
image projections by linear interpolation of same patient before metal implants were
the metal regions, and after reconstruction inserted or a scan of a similar patient [93].
applies an edge-preserving blur filter to This image must be coregistered with the
suppress artifacts introduced by interpolation measured image that contains artifacts, and
in this pre-corrected image. The CT numbers then reprojected to be directly used in
in the metal areas of the pre-corrected image sinogram completion. This can be a highly
replace the metal in the original uncorrected effective artifact correction method assuming
image, and the prior image is created by the that an accurate prior image is available.
fusion of these two images. This prior image For cases when an analogous artifact-free
is forward projected and used to guide the scan is not an option, Zhang et al.
sinogram replacement. The new sinogram implemented a prior image estimation
values are the result of subtracting the prior scheme by imposing TV minimization and a
sinogram from the original sinogram, uniformity constraint around the metal
performing linear interpolation in the metal objects for algebraic reconstruction of the
trace, and then adding the result to the prior original projection data [94]. The constrained
sinogram. prior image is then forward projected and
Li et al. utilized a prior image generated combined with smooth interpolations of
from the original reconstructed image with an original projection data to replace corrupt
edge-preserving filter applied and sinogram regions. This hybrid technique
replacement of metal with soft tissue CT shows promising results even for severe
metal artifacts. Similarly, Bannas et al. soft tissue. The image is reprojected and the
employed a prior image obtained using an original sinogram is divided by the prior
iterative compressed sensing reconstruction image sinogram. This normalization
technique [95]. The prior image constrains increases the homogeneity of the regions
the estimation of missing information to where linear interpolation will be applied,
improve image quality. which leads to more accurate interpolation
results. After interpolation, the projection
3.4.3 Normalization data are denormalized and the corrected data
Data normalization helps to improve the are reconstructed. The normalization and de-
projection completion method for metal normalization procedure fuses original and
artifact reduction. The idea is to use the x-ray completed data more smoothly. Image
path length through a comparable cross- quality measures compared NMAR against
section for data normalization so that the the standard linear interpolation and MAR
sinogram becomes comparatively flat and based on simple length normalization for CT
thus more straightforward for interpolation scans of hip prostheses, dental fillings, and
over corrupted metal traces. Müller and spine fixation. NMAR can greatly reduce
Buzug presented one way by first artifacts close to the metal surface even in
reconstructing an uncorrected image to find severe cases, but is dependent on a good prior
metal objects and traces, and then image with accurate segmentation. This
normalizing projections by dividing each method produces image quality gains and is
measured data point by the intersection computationally efficient compared to
length of the corresponding x-ray and the iterative methods.
metal object [96]. After interpolation of metal The same group improved their method
traces, data are denormalized for image with an adaptive NMAR (ANMAR)
reconstruction. algorithm that merges original and NMAR
Meyer et al. introduced a new type of projections using a weighted sum to avoid
projection completion technique, called data loss close to metal objects [97]. Fig. 10
normalized MAR (NMAR), that normalizes presents the comparison of this method to
original projection data to prior image NMAR and simple linear interpolation
projection data [63]. An artifact-free prior (MAR1). Lell et al. applied NMAR to
image is obtained by multithreshold clinical cases of head and neck CT to evaluate
segmentation of the original image after its performance and found that there were no
smoothing to define regions of bone, air, and algorithm-induced artifacts [98].
Figure 10. Adaptive NMAR achieves more natural images with comparable streak reduction to NMAR. Reprinted
from Meyer et al. [97] by permission of IEEE.
3.5 Iterative Reconstruction function is deterministic, meaning it is not

Iterative reconstruction starts with an directly interpreted in terms of statistics.
assumed image and compares its projections Corrupted data due to metal are explicitly
to the projection data actually measured by discarded. The EM procedure defines an
the CT scan acquisition. The goal is then to initial image estimate and reprojects it to
minimize the error between these sinograms obtain the estimated projections. The
by optimizing an objective function that measured projection data are divided by the
guides the reconstruction. Examples of this estimated projection data, and then
objective function include minimum least backprojected and multiplied by the current
squares error and maximum likelihood, image estimate (pixel by pixel) to generate an
which aims to find the distribution of linear improved image estimate. In simulations
attenuation coefficients from the projection with noise-free and additive noisy projection
data with the maximum probability. data from dental phantoms, both algorithms
produced superior image quality as compared
3.5.1 Corruption Avoidance to FBP after linearly fitting projection gaps.
One approach to iterative reconstruction is to Furthermore, the EM-type algorithm
completely ignore the subset of projection converges faster than the other algorithm in
data that is corrupted by metal objects. These terms of either the I-divergence or Euclidean
methods treat the MAR problem as the distance between ideal and reprojected data.
exterior problem, and use data outside the Also, for a given iteration number, the EM-
metal trace to arrive at a reconstruction result. type deblurring method produces better
The exterior problem gives a unique solution image clarity but stronger noise than the
outside the metal object, but is not that stable ART-type counterpart. These algorithms
close to the metal surface. were adapted for cone-beam CT with the
Medoff et al. defined the “bagel problem” addition of a 3D relaxation factor that
for MAR, which treats the object being accounts for beam inhomogeneity [101].
scanned like a bagel with the metal-affected The iterative nature of the EM-type
region as the hole in the center [99]. The algorithm makes it computationally slow
measurements of x-ray data passing through relative to MAR methods based on FBP, so
the hole are completely ignored and only the Wang et al. sought to speed up their method
area outside is reconstructed from unaffected by using row-action/ordered-subsets of the
line integrals. The hole is filled in with projection data [102]. Each iteration uses
constant attenuation values, optionally based only a subset of the data so that the image can
on the known density of the metal object, and be estimated more quickly. The computation
the corresponding line integrals are time is reduced roughly by a factor equal to
estimated. An iterative update of the image is the total number of data subsets used. Just
performed until steady state is reached. four iterations were able to achieve
Image quality is improved using this method, significantly improved image quality.
though information content at the metal An inverse iterative procedure was
edges is reduced. presented by August and Kanade in which a
Wang et al. formulated two iterative metal mask is generated to isolate projections
deblurring algorithms for metal artifact that need correction [103]. A penalized
reduction, using the expectation maximum likelihood method is applied only
maximization (EM) formulation and the to the metal mask, which significantly speeds
simultaneous algebraic reconstruction up processing time. The non-metal region is
technique (ART) [100]. The objective used as a constraint for ensuring corrected
data consistency, while smoothness is Like other statistical reconstruction

ensured by a gradient penalty term in the algorithms assuming Poisson data, low-count
optimization equation. A ten-fold reduction data are naturally deemphasized. Unlike
in processing time compared to full image earlier iterative algorithms, the
optimization was seen on images of hip polychromatic data model reflects the
implants, with comparable final results. A physics most realistically. The polychromatic
similar approach was developed data-driven iterative reconstruction algorithm
independently by De Man [104]. prevents beam hardening and metal artifacts,
with the spectrum of the x-ray tube modeled
3.5.2 Statistical Compensation as discrete energy bins. The energy-
Instead of explicitly omitting metal-affected dependent attenuation coefficients are taken
data, a statistical objective function can be into account by decomposing these
employed that down-weights data through coefficients into photoelectric and Compton
the metal object, but still includes them in the components, under the constraint of relative
iterative reconstruction. weight of these components based on prior
De Man et al. developed a maximum a material assumptions. Good results are
posteriori algorithm for metal artifact obtained in simulations and phantom
reduction based on the transmission experiments.
maximum likelihood algorithm (ML-TR) Hamelin et al. adapted polychromatic
that performs CT reconstruction by sinogram models that compensate for beam
optimizing the likelihood of measured hardening during iterative reconstruction
Poisson data [105], [106]. Statistically, ML- [109]. They also developed a model for
TR attributes less weight to low-count data, Gaussian noise that is appropriate for large
thereby reducing artifacts from directions of average photon counts. These models can be
low counts due to metal obstacles. This applied on a region of interest basis to
iterative algorithm uses a Markov random account for areas of the image that suffer
field smoothness prior with the Huber from different artifacts. A more complex
potential function. Iterations are performed to polychromatic beam model was employed by
produce intermediate images at doubled Van Slambrouck and Nuyts, but only in
resolution. The final image is down-sampled regions of the image close to metal [110]. A
to normal resolution. In comparison to the patch method is used to determine metal and
simple linear interpolation method and non-metal regions, and depending on the
typical iterative algorithms (ML-TR and ML- contents, a simple beam model (MLTRC) or
EM) ignoring corrupted data, this method complex model that can also increase
produces good results in numerical and resolution (IMPACT) is applied.
phantom tests, reducing streak artifacts and The maximum likelihood expectation
preserving linear details. maximization (MLEM) algorithm assumes a
A related polychromatic iterative Poisson noise distribution of the projection
reconstruction approach was proposed by data and reconstructs the image of linear
Menvielle, in which a hyperbolic attenuation values with the highest
regularization and conjugate gradient probability. This particular statistical model
optimization is used [107]. De Man et al. also is accurate for emission tomography, but not
incorporated a polychromatic data for transmission data in CT. The algorithm
acquisition model to account for beam weights transmission data incorrectly
hardening into an iterative maximum- because the noise on the x-ray path length
likelihood reconstruction algorithm [108]. (post-log) does not follow a Poisson
distribution. The MLEM procedure is still projections [114]. The algorithm outperforms
advantageous in down-weighting ART and EM methods for streak artifact
inconsistent projections to suppress metal reduction.
artifacts, but can be improved for modeling
transmission data. Several MAR methods 3.5.3 Knowledge Utilization
employ this technique for an iterative Iterative reconstruction methods involving
reconstruction after an interpolation scheme incomplete data can be improved with further
to minimize algorithm-induced artifacts. considerations, such as compressed sensing
Oehler and Buzug correct inconsistent and known component models. With an
projection data by directional interpolation in increased aging population, more and more
combination with entropy maximization to CT images contain implants, surgical tools,
obtain appropriate weightings [111]. The and other metallic objects. A significant
directional interpolation method is proposed opportunity for metal artifact reduction is to
to fill the data gap following the data flow in use these precisely known physical models.
the sinogram outside the metal traces, A component model defines the shape, size,
allowing interpolation from data in different and/or density information of the metal
projections. This directional method leads to objects in the image to help estimate the
superior results compared to the view-based missing projections. Kalvin and Williamson
interpolation schemes. A directional developed an iterative algorithm that uses a
interpolation is recommended with an scout image as a prior image for cases when
appropriate weighting to form a weighted unencoded projection data are not available
maximum likelihood framework for metal [115]. The scout image is acquired by
artifact reduction, which balances between keeping the x-ray tube stationary and moving
missing data and residual inconsistencies. A the patient table. Constraints on the corrected
regularization term in the log-likelihood image are imposed by scout projection data
function helps stabilize the model. for consistency and information about the
Kratz and Buzug also applied weighted objects in the image, such as size. Kalvin
MLEM after NFFT interpolation, which is patented a process based on this sequence
effective for higher interpolation dimensions [116].
[112]. This method can accommodate prior Chen et al. devised an algorithm to utilize
information about the metal object. NFFT sparsity of an underlying image and generate
helped reduced relative errors based on the a realistic prior background through
ground truth in the image domain by more compressed sensing reconstruction from
than 30% compared to linear and cubic sparse data, which requires only 20 view
interpolation. Aootaphao et al. introduced a acquisition angles [117].
modified convex algorithm to include a An extension of EM-type iterative
penalty term in the MLEM process, which deblurring by Snyder et al. incorporates the
overcomes noisy images that can result from shape and attenuation characteristics of the
MLEM alone on ill-conditioned problems known metal object to improve the iterative
[113]. This followed cubic interpolation of reconstruction for intracavitary applicators
metal regions with an offset that allows [118]. O’Sullivan and Benac developed
surrounding bone to be discerned. alternating minimization to monotonically
Zhang et al. use constrained optimization, decrease the objective function [119]. This
which minimizes an objective function that technique minimizes the I-divergence
preserves smoothness by restricting between measured and estimated model-
projection values based on original non-metal based data, which is a reformulation of the
maximum-likelihood. Murphy et al. aim to reconstructed, and do not rely on access to

estimate the pose of metal implants by using raw projection data. These techniques alone
a known component model of the shape and are often not as effective as raw data
composition [120]. This method employs a correction approaches since the damage
steepest descent gradient algorithm to already occurred, but combining post-
approximate the position and orientation of processing with reconstruction modifications
the objects. may have merits, as seen with some methods
Stayman et al. created a model-based discussed in this section.
penalized-likelihood approach that
incorporates a component specification 3.6.1 Direct
model, and performs an alternating Direct post-processing for streak reduction
maximization procedure for both the was attempted by Henrich with minimal
anatomy and the known object inside the effectiveness [122]. This includes a simple
patient [121]. This proposed method normalization of rows and columns to a gray
achieved high quality images in simulated value. Soltanian-Zadeh et al. identified streak
vertebral pedicle screw reconstructions. Fig. artifacts by thresholding a difference image
11 compares image quality of known- between original and low-pass filtered
component reconstruction (KCR) with FBP images and removed them successfully in
and penalized-likelihood for an image brain phantom images [123]. For images with
containing a pedicle screw. mild artifacts, Bal et al. employed a radial
adaptive filter that acts only on artifact
3.6 Image Post-processing regions of the reconstructed image while
A less popular MAR approach is to perform preserving spatial resolution elsewhere [124].
post-processing, which aims to correct in the The filter is derived from the local structure
image domain. Post-processing methods tensor at each pixel.
reduce artifacts after the image has been
Figure 11. KCR estimates the pose of the unilateral single-component screw by simultaneous registration and
reconstruction to yield superior image quality to FBP and penalized-likelihood reconstruction. Top: axial view.
Bottom: coronal view. Reprinted from Stayman et al. [121] by permission of IEEE.
3.6.2 Hybrid hardening and image noise. However, this

For more effective results, post-processing method often introduces new artifacts. On the
has been integrated with other categories of other hand, the MAF method greatly
MAR techniques. Watzke and Kalender suppresses image noise, but it does not
proposed to combine two artifact reduction effectively reduce artifacts. The synergy
methods via image-based weighted among the two methods is clearly shown with
superposition of two versions of an image patient data that leads to superior image
obtained using different MAR methods, the quality than what can be obtained from either
simple linear interpolation method and the method alone. The outcomes achieved with
multi-dimensional adaptive filtering (MAF) directional weighting are generally preferred
method [125]. The two techniques have their over those with the distance weighting. Fig.
strengths and weaknesses, since they were 12 illustrates the merging of linear
designed based on different considerations. interpolation and MAF images, and the
The linear interpolation method does not use image quality improvement with directional
corrupted data, thereby reducing beam weighting.
Figure 12. Images of two cobalt-chrome

hip implants with strong artifacts between
them. a) Original; b) MAF; c) Linear
interpolation (LI); d) LI-MAF merging
with distance weighting; e) LI-MAF
merging with directional weighting.
Merging with directional weighting
significantly improves image quality and
minimizes secondary artifacts. Reprinted
from Watzke and Kalender [125] with
permission from Springer.
Meyer et al. further introduced a processing correction may be sufficient, as

frequency split MAR protocol that combines there are several methods in these classes
high frequency information of the original geared to those implant categories.
image to preserve edge data with low Despite four decades of effort toward
frequency information of the image corrected reducing metal artifacts in CT (starting with
by NMAR [126]. A weighting scheme is used Lewitt and Bates in the late 1970s [48]), there
to optimize high frequency data containing are important applications in which metal
edge information without including excess objects remain troublesome. In light of Table
noise in the final reconstructed image. 4, we underline that larger, irregular metal
A baggage screening MAR technique was parts remain challenging, and hybrid iterative
presented by Mouton et al. in which reconstruction is the main approach to utilize
projection completion is combined with post- physical models and minimize metal
processing [127]. Metal projection data are artifacts.
replaced by spline interpolation of As a first step for many MAR algorithms,
neighboring projections and the corrected metal segmentation is performed on a first-
sinogram is reconstructed. The resulting pass reconstruction image. Due to severe
image is then processed to reduce artifacts beam-hardening or photon starvation, it is
introduced by interpolation. The pixel values rather difficult to have an accurate
are constrained so that they cannot exceed segmentation, especially close to the metal
their value in the original image to avoid surface. Huang et al. performed a thorough
bright streaks. More recently, Ballhausen et study of three commercially available MAR
al. combined CT volumes acquired at techniques on phantoms corresponding to
different tilt angles and computed the average different types of metal implants [129]. These
gray values from each volume for regions in methods are 1) the O-MAR algorithm by
space that had artifacts present in some Philips Healthcare, which is an iterative
volumes and not others [128]. This helped projection replacement technique based on a
strongly reduce streaks and improved visual tissue-class model, 2) the monochromatic
details near metal objects. gemstone spectral imaging (GSI) protocol by
GE, which employs dual-energy acquisition
4. DISCUSSION AND CONCLUSION at 80 and 140 kVp, without any MAR
Oftentimes, different MAR approaches have software and 3) the GSI protocol with
their unique niches. Table 4 summarizes the additional MAR software. For the case of
detailed assessment of MAR methods large hip implants, GSI with software
described in Section 3 by establishing their correction achieved the best artifact reduction
relationship to the metal objects whose results due to the decrease in beam hardening
artifacts they were intended to overcome. effects facilitated by dual-energy acquisition.
Projection completion algorithms make up More complex phantoms, such as dental
the majority of MAR techniques, and are fillings, presented a challenge for all three
primarily targeted to correct artifacts from MAR techniques, characterized by the
large and/or highly attenuating metal objects, production of additional artifacts in the image
such as prosthetic hips and dental fillings. without major reduction of the original
There are also several iterative reconstruction streaks. For titanium rod spinal implants, all
algorithms developed for reducing artifacts software algorithms induced streaks between
from these types of implants. For smaller and the metal objects and the imaging target,
less attenuating metals, improvements to the which was the lung. These results highlight
acquisition protocol or physics-based pre- the limitations of existing MAR methods.
Although the O-MAR algorithm is intended Radiologists assessing images of head and
for orthopedic implant applications, Kidoh et neck tumors in patients with fillings
al. tested it strictly for dental fillings and concluded that O-MAR improves the
determined that it does have utility in depiction of the oral cavity, but it causes
improving image quality, but it cannot unnatural texture in the images.
effectively overcome streak artifacts [130].
Table 4. Relationship between representative publications in each MAR method class and the categories of metal
objects for which they are intended to correct.
METAL OBJECT CATEGORY
I II III IV V
Bamberg[31],
Brown[18], Lewis[19],
Acquisition Lee[26], Nasirudin[37],
Mamourian[21], Jeon[29] Pessis[34],
Improvement Link[23], Wu[28]
Lakits[20] Schreiner[27]
Meinel[32]
Prell[42],
Verburg[45], Kachelriess[40],
Physics Pre-
Verburg[45] Meyer[47], Meyer[47],
Processing
Schuller[46] Van
Slambrouck[110]
Hinderling[49],
Zhao[61],
Mahnken[55],
Glover[50],
Joemai[68],
Kalender[52],
Meyer[97], Li[71], Mehranian[62],
Yu[60],
Olive[83], Mehranian[62], Tohnak[72],
Moseley[56],
Projection Philips O- Kratz[69], Veldkamp[59], Yazdi[73],
Toftegaard[78],
Completion MAR[88], Jeong[82] Yazdi[58], Lemmens[85],
Xu[74],
Prell[86] Bal[84], Meyer[63],
Prell[86],
Bannas[95], Zhang[94]
Prell[87],
Meyer[63],
Tuy[81]
Philips O-
MAR[88]
August[103],
Stayman[121], Wang[100],
Iterative Kalvin[115],
Snyder[118], De Man[105] Oehler[111]
Reconstruction Wang[102],
Murphy[120]
Oehler[111]
Watzke[125],
Image Post- Watzke[125],
Bal[124] Meyer[126],
Processing Meyer[126]
Ballhausen[128]
To optimize metal artifact reduction, models constrain the data to only realistic
important information could be extracted representations of the image. Compressed
from data segments compromised by metal sensing methods can be considered under
objects. It is emphasized that such data prior models in that they use known
segments could be improved using multi- properties such as sparsity, low-rank, and
energy scanning techniques, which can dictionaries to guide the reconstruction when
effectively facilitate metal artifact reduction. insufficient data is available from the
Other prior information in terms of expected acquisition process. Complementary data
image domain knowledge is useful as well. A from other imaging modalities has also been
most immediate objective of MAR is to shown to reduce CT metal artifacts in a
improve the fidelity of the acquired data so unique way. Tri-modality imaging systems
that segmentation quality can be optimized that perform CT, MRI, and positron emission
and advanced correction algorithms may not tomography (PET) in a single session can
be as necessary. Higher tube voltages acquire synergistic information for image
generate higher energy x-rays that can correction as demonstrated by Delso et al
penetrate dense metal objects and increase [134].
the inherent quality of the raw data. Multiple- A common observation in MAR papers is
energy protocols, high-energy sparse view that the results of a proposed technique are
acquisition, and photon-counting spectral CT often compared to results obtained using the
have emerged as promising techniques to simple linear interpolation algorithm.
overcome beam hardening and photon Unsurprisingly, nearly every new approach is
starvation artifacts that result from x-rays far superior to this benchmark. To improve
passing through metal objects. A high-kVp synergy in the MAR field and compare many
dual-energy protocol has been proposed by techniques to each other, there is a need for
Xi et al. in which sparse views at 160 kVp, or better benchmarks and metrics.
even 170 kVp, are acquired intermittently Reconstruction algorithms ought to be run on
during a standard 120 kVp scan [131]. The identical data from the same acquisition
data obtained at higher energy suffer from process. For techniques that alter the scan
less beam hardening and photon starvation procedures, identical phantoms should be
effects, thereby providing adequate available. Phantom templates could be shared
projections that can replace corrupt in a database, downloaded, and 3D printed so
projections through metal objects acquired at that there is common ground for comparison.
the standard energy. Simulation data of this These phantoms also cannot be too simple,
process has shown high image quality close otherwise the complex cases of metal objects
to the metal surface, and is a viable option for are not represented. Testing also needs to
improving radiation therapy planning. include task-based evaluation to assess
Model-based iterative reconstruction is clinical impact, such as radiation therapy
also a promising research direction for planning. An additional point is that many
reducing metal artifacts [132], [133]. This MAR algorithms require access to raw
general class of algorithms incorporates projection data, but these often cannot be
models of physics and materials that help implemented on commercial CT scanners
improve the consistency and relevance of x- due to restrictions.
ray data. System models account for the In recent years, public concerns over x-ray
polychromatic spectrum of the x-ray tube, radiation dose have led to great progress in
statistical noise models adjust for the size of low-dose CT. Sub-milli-Sievert scans are
the focal spot and detector shape, and prior becoming common. As a result, metal
artifacts could become more evident, which fashion to become a universal intelligent
will demand more attention to maintain image analyzer. This could be an exciting
diagnostic performance. For low-dose CT direction for big-data-driven intelligent metal
and spectral imaging, photon-counting artifact reduction.
detectors are emerging that have an inherent Forty years of research and development
benefit for metal artifact reduction, but have have produced effective methods for
yet to be fully explored. Among many reducing metal artifacts. Despite these
relevant clinical tasks, a primary example efforts, some metal objects still pose great
that heavily relies on successful metal artifact challenges for MAR, mainly large or
reduction is proton therapy planning. In this irregular prostheses and dense, highly
case, any residual metal artifacts can attenuating implants. There is still no
significantly compromise therapeutic universal solution for cases involving these
outcome margins. types of metal that demand high image
From a methodological perspective, metal accuracy, especially for proton therapy. To
artifact reduction and local/interior reach this goal, improved acquisition
tomography are closely related. Roughly protocols will need to be combined with
speaking, in the case of metal artifact reconstruction algorithms that account for the
reduction, portions of data inside a projection relevant physics effects and prior
profile are unavailable or seriously information. This approach fosters
compromised due to the presence of metal collaboration between multiple groups in
objects. Meanwhile, local/interior different disciplines.
tomography targets reconstruction of an In conclusion, multi-energy data
internal region of interest (ROI) from purely acquisition, advanced algorithms, extensive
truncated projection data segments delimited prior knowledge, standardized benchmark
by the ROI (interior tomography), or from datasets, as well as clinical evaluation and
these local data segments plus some validation would synergistically contribute to
additional measured and/or extrapolated data the most effective techniques for metal
along paths outside the ROI (local artifact reduction in CT.
tomography). Since interior tomography
solves the classical interior problem [135],
[136], in a good sense metal artifact reduction
methods define what we call exterior
tomography. Mathematically, exterior References
tomography allows a unique solution but
regularization is needed for stability. [1] G. N. Hounsfield, “Computerized transverse
axial scanning (tomography): Part 1.
As a side note, the recent news exposure Description of system,” Br. J. Radiol., vol.
on AlphaGo demonstrates an impressive 46, no. 552, pp. 1016–1022, 1973.
success of machine learning, and is inspiring [2] W. Kilby, J. Sage, and V. Rabett, “Tolerance
for metal artifact reduction [137]. It can be levels for quality assurance of electron
imagined that the above-described density values generated from CT in
radiotherapy treatment planning,” Phys. Med.
benchmark database can be expanded to Biol., vol. 47, no. 9, p. 1485, 2002.
serve as a source of big data for training so [3] D. D. Robertson, P. J. Weiss, E. K. Fishman,
that a machine learning algorithm can be D. Magid, and P. S. Walker, “Evaluation of
developed to recognize metal artifacts and CT techniques for reducing artifacts in the
infer underlying structures. Such an presence of metallic orthopedic implants,” J.
Comput. Assist. Tomogr., vol. 12, no. 2, pp.
algorithm can be optimized in a task-specific 236–241, 1988.
[4] B. De Man, J. Nuyts, P. Dupont, G. Marchal, imaging artifact after using three types of
and P. Suetens, “Metal streak artifacts in X- metal rods: stainless steel, titanium, and
ray computed tomography: a simulation vitallium,” Spine J., vol. 10, no. 9, pp. 789–
study,” IEEE Trans. Nucl. Sci., vol. 46, no. 3, 794, 2010.
pp. 691–696, 1999. [16] S. Galli, G. Szakacs, F. Lukac, M. Vlcek, R.
[5] M. Kachelriess, “Reduktion von Jimbo, Y. Naito, A. Wennerberg, J. Herzen,
Metallartefakten in der Roentgen-Computer- J. Hammel, and R. Willumeit,
Tomographie,” Friedrich-Alexander- “Osseointegration of resorbable magnesium
Universitat Erlangen-Nurnberg, 1998. screws–A SRμCT Study,” in 6th Symposium
[6] F. E. Boas and D. Fleischmann, “CT on Biodegradable Metals, 2014.
artifacts: Causes and reduction techniques,” [17] M. N. Zimel, S. Hwang, E. R. Riedel, and J.
Imaging Med., vol. 4, no. 2, pp. 229–240, H. Healey, “Carbon fiber intramedullary nails
2012. reduce artifact in postoperative advanced
[7] J. Hsieh, Computed tomography: principles, imaging,” Skeletal Radiol., vol. 44, no. 9, pp.
design, artifacts, and recent advances. 1317–1325, 2015.
Bellingham, WA: SPIE, 2009. [18] J. H. Brown, E. S. Lustrin, M. H. Lev, C. S.
[8] G. H. Glover and N. J. Pelc, “Nonlinear Ogilvy, and J. M. Taveras, “Reduction of
partial volume artifacts in x-ray computed aneurysm clip artifacts on CT angiograms: a
tomography,” Med. Phys., vol. 7, no. 3, pp. technical note.,” AJNR. Am. J. Neuroradiol.,
238–248, 1980. vol. 20, no. 4, pp. 694–6, 1999.
[9] Y. Jin, V. Robinson, L. Gjesteby, G. Wang, J. [19] M. Lewis, A. P. Toms, K. Reid, and W.
Verburg, D. Giantsoudi, H. Paganetti, and B. Bugg, “CT metal artefact reduction of total
De Man, “Feasibility demonstration of high- knee prostheses using angled gantry
voltage clinical CT and impact on X-ray multiplanar reformation,” Knee, vol. 17, no.
penetration through metal objects,” in AAPM, 4, pp. 279–282, 2010.
2016. [20] A. Lakits, R. Prokesch, C. Scholda, R.
[10] W. A. Gray, S. Sekaran, J. A. Tanyi, and J. Nowotny, A. Kaider, and A. Bankier,
M. Holland, “Implications of Dental Artifacts “Helical and conventional CT in the imaging
on Radiotherapy Planning for Head and Neck of metallic foreign bodies in the orbit,” Acta
Cancer,” in Multidisciplinary Head & Neck Ophthalmol. Scand., vol. 78, no. 1, pp. 79–
Symposium, 2012. 83, 2000.
[11] J. L. Weese, M. S. Rosenthal, and H. Gould, [21] A. C. Mamourian, K. Erkmen, and D. J.
“Avoidance of artifacts on computerized Pluta, “Nonhelical acquisition CT angiogram
tomograms by selection of appropriate after aneurysmal clipping: In vitro testing
surgical clips,” Am. J. Surg., vol. 147, no. 5, shows diminished artifact,” Am. J.
pp. 684–687, 1984. Neuroradiol., vol. 29, no. 4, pp. 660–662,
[12] N. A. Ebraheim, R. Coombs, J. J. Rusin, and 2008.
W. T. Jackson, “Reduction of postoperative [22] R. P. Tolakanahalli, J. E. Medow, J. Hsieh,
CT artifacts of pelvic fractures by use of G.-H. Chen, and C. A. Mistretta, “Reduction
titanium implants,” Orthopedics, vol. 13, no. of metal artifacts using subtracted CT
12, pp. 1357–1358, 1990. projection data,” in SPIE 5535,
[13] N. Haramati, R. B. Staron, K. Mazel- Developments in X-Ray Tomography IV,
Sperling, K. Freeman, E. L. Nickoloff, C. 2004, vol. 5535, pp. 636–643.
Barax, and F. Feldman, “CT scans through [23] T. M. Link, W. Berning, S. Scherf, U.
metal scanning technique versus hardware Joosten, A. Joist, K. Engelke, and H. E.
composition,” Comput. Med. Imaging Daldrup-Link, “CT of metal implants:
Graph., vol. 18, no. 6, pp. 429–434, 1994. reduction of artifacts using an extended CT
[14] J. W. Lambert, P. M. Edic, P. Fitzgerald, A. scale technique,” J. Comput. Assist. Tomogr.,
S. Torres, and B. M. Yeh, “Complementary vol. 24, no. 1, pp. 165–172, 2000.
contrast media for metal artifact reduction in [24] M.-J. Lee, S. Kim, S.-A. Lee, H.-T. Song, Y.-
dual-energy CT,” in SPIE Medical Imaging, M. Huh, D.-H. Kim, S. H. Han, and J.-S. Suh,
2015, p. 941203. “Overcoming Artifacts from Metallic
[15] P. T. Knott, S. M. Mardjetko, R. H. Kim, T. Orthopedic Implants at High-Field-Strength
M. Cotter, M. M. Dunn, S. T. Patel, M. J. MR Imaging and Multi-detector CT 1,”
Spencer, A. S. Wilson, and D. S. Tager, “A Radiographics, vol. 27, no. 3, pp. 791–803,
comparison of magnetic and radiographic 2007.
[25] S. G. Moon, S. H. Hong, J.-Y. Choi, W. S. 1257, 2011.

Jun, H.-G. Kang, H.-S. Kim, and H. S. Kang, [34] E. Pessis, R. Campagna, J.-M. Sverzut, F.
“Metal artifact reduction by the alteration of Bach, M. Rodallec, H. Guerini, A. Feydy,
technical factors in multidetector computed and J.-L. Drapé, “Virtual Monochromatic
tomography: a 3-dimensional quantitative Spectral Imaging with Fast Kilovoltage
assessment,” J. Comput. Assist. Tomogr., vol. Switching: Reduction of Metal Artifacts at
32, no. 4, pp. 630–633, 2008. CT,” Radiographics, vol. 33, no. 2, pp. 573–
[26] I. S. Lee, H. J. Kim, B. K. Choi, Y. J. Jeong, 583, 2013.
T. H. Lee, T.-Y. Moon, and D. W. Kang, “A [35] P. Sukovic and N. H. Clinthorne, “Penalized
pragmatic protocol for reduction in the metal weighted least-squares image reconstruction
artifact and radiation dose in multislice for dual energy X-ray transmission
computed tomography of the spine: cadaveric tomography,” IEEE Trans. Med. Imaging,
evaluation after cervical pedicle screw vol. 19, no. 11, pp. 1075–1081, 2000.
placement,” J. Comput. Assist. Tomogr., vol. [36] K. Rajendran, M. F. Walsh, N. J. A. de
31, no. 4, pp. 635–641, 2007. Ruiter, A. I. Chernoglazov, R. K. Panta, A. P.
[27] L. J. Schreiner, M. Rogers, G. Salomons, and H. Butler, P. H. Butler, S. T. Bell, N. G.
A. Kerr, “Metal artifact suppression in Anderson, T. B. F. Woodfield, S. J.
megavoltage computed tomography,” in Tredinnick, J. L. Healy, C. J. Bateman, R.
SPIE Medical Imaging 2005, 2005, vol. Aamir, R. M. N. Doesburg, P. F. Renaud, S.
5745, no. 613, pp. 637–645. P. Gieseg, D. J. Smithies, J. L. Mohr, V. B.
[28] M. Wu, A. Keil, D. Constantin, J. Star-Lack, H. Mandalika, A. M. T. Opie, N. J. Cook, J.
L. Zhu, and R. Fahrig, “Metal artifact P. Ronaldson, S. J. Nik, A. Atharifard, M.
correction for x-ray computed tomography Clyne, P. J. Bones, C. Bartneck, R. Grasset,
using kV and selective MV imaging,” Med. N. Schleich, and M. Billinghurst, “Reducing
Phys., vol. 41, no. 12, p. 121910, 2014. beam hardening effects and metal artefacts in
[29] H. Jeon, D. Park, H. Youn, J. Nam, J. Lee, spectral CT using Medipix3RX,” J. Instrum.,
W. Kim, Y. Ki, Y. H. Kim, J. H. Lee, D. vol. 9, no. 03, p. P03015, 2014.
Kim, and H. K. Kim, “Generation of hybrid [37] R. A. Nasirudin, K. Mei, P. Panchev, A.
sinograms for the recovery of kV-CT images Fehringer, F. Pfeiffer, E. J. Rummeny, M.
with metal artifacts for helical tomotherapy,” Fiebich, and P. B. Noël, “Reduction of Metal
Med. Phys., vol. 42, no. 8, pp. 4654–4667, Artifact in Single Photon-Counting
2015. Computed Tomography by Spectral-Driven
[30] M. Lewis, K. Reid, and A. P. Toms, Iterative Reconstruction Technique,” PLoS
“Reducing the effects of metal artefact using One, vol. 10, no. 5, 2015.
high keV monoenergetic reconstruction of [38] J. Hsieh, “Adaptive streak artifact reduction
dual energy CT (DECT) in hip in computed tomography resulting from
replacements,” Skeletal Radiol., vol. 42, no. excessive x-ray photon noise,” Med. Phys.,
2, pp. 275–282, 2013. vol. 25, no. 11, pp. 2139–47, 1998.
[31] F. Bamberg, A. Dierks, K. Nikolaou, M. F. [39] R. M. Rangayyan and R. Gordon, “Streak
Reiser, C. R. Becker, and T. R. C. Johnson, Preventive Image Reconstruction with ART
“Metal artifact reduction by dual energy and Adaptive Filtering,” IEEE Trans. Med.
computed tomography using monoenergetic Imaging, vol. MI–1, no. 3, pp. 173–178,
extrapolation,” Eur. Radiol., vol. 21, no. 7, 1983.
pp. 1424–1429, 2011. [40] M. Kachelrieß, O. Watzke, and W. A.
[32] F. G. Meinel, B. Bischoff, Q. Zhang, F. Kalender, “Generalized multi-dimensional
Bamberg, M. F. Reiser, and T. R. C. Johnson, adaptive filtering for conventional and spiral
“Metal Artifact Reduction by Dual-Energy single-slice, multi-slice, and cone-beam CT,”
Computed Tomography Using Energetic Med. Phys., vol. 28, no. 4, pp. 475–490,
Extrapolation,” Invest. Radiol., vol. 47, no. 7, 2001.
pp. 406–414, 2012. [41] O. Watzke, “Metallartefakreduktion in der
[33] C. Zhou, Y. E. Zhao, S. Luo, H. Shi, L. Computertomographie,” Friedrich-
Zheng, L. J. Zhang, and G. Lu, Alexander-Universitat Erlangen-Nurnberg,
“Monoenergetic imaging of dual-energy CT 2002.
reduces artifacts from implanted metal [42] D. Prell, Y. Kyriakou, M. Kachelriess, and
orthopedic devices in patients with factures,” W. A. Kalender, “Reducing metal artifacts in
Acad. Radiol., vol. 18, no. 10, pp. 1252– computed tomography caused by hip
endoprostheses using a physics-based method.,” J. Nucl. Med., vol. 41, no. 11, pp.
approach,” Invest. Radiol., vol. 45, no. 11, 1913–1919, 2000.
pp. 747–754, 2010. [54] R. L. Morin and D. E. Raeside, “A pattern
[43] F. E. Boas and D. Fleischmann, “Evaluation recognition method for the removal of
of two iterative techniques for reducing metal streaking artifact in computed tomography,”
artifacts in computed tomography,” Radiology, vol. 141, no. 1, pp. 229–233,
Radiology, vol. 259, no. 3, pp. 894–902, 1981.
2011. [55] A. H. Mahnken, R. Raupach, J. E.
[44] P. Seitz and P. Rüegsegger, “Anchorage of Wildberger, B. Jung, N. Heussen, T. G.
femoral implants visualized by modified Flohr, R. W. Gunther, and S. Schaller, “A
computed tomography,” Arch. Orthop. New Algorithm for Metal Artifact Reduction
Trauma. Surg., vol. 100, no. 4, pp. 261–266, in Computed Tomography: In Vitro and In
1982. Vivo Evaluation After Total Hip
[45] J. M. Verburg and J. Seco, “CT metal artifact Replacement,” Invest. Radiol., vol. 38, no.
reduction method correcting for beam 12, pp. 769–775, 2003.
hardening and missing projections,” Phys. [56] D. J. Moseley, J. H. Siewerdsen, and D. A.
Med. Biol., vol. 57, no. 9, pp. 2803–2818, Jaffray, “High-contrast object localization
2012. and removal in cone-beam CT,” in SPIE
[46] S. Schüller, S. Sawall, K. Stannigel, M. Medical Imaging 2005, 2005, vol. 5745, pp.
Hülsbusch, J. Ulrici, E. Hell, and M. 40–50.
Kachelrieß, “Segmentation-free empirical [57] J. Wei, L. Chen, G. a Sandison, Y. Liang, and
beam hardening correction for CT,” Med. L. X. Xu, “X-ray CT high-density artefact
Phys., vol. 42, no. 2, p. 794, 2015. suppression in the presence of bones,” Phys.
[47] E. Meyer, C. Maas, M. Baer, R. Raupach, B. Med. Biol., vol. 49, pp. 5407–5418, 2004.
Schmidt, and M. Kachelriess, “Empirical [58] M. Yazdi, L. Gingras, and L. Beaulieu, “An
scatter correction (ESC): A new CT scatter adaptive approach to metal artifact reduction
correction method and its application to metal in helical computed tomography for radiation
artifact reduction,” in IEEE Nuclear Science therapy treatment planning: experimental and
Symposuim & Medical Imaging Conference, clinical studies.,” Int. J. Radiat. Oncol. Biol.
2010, pp. 2036–2041. Phys., vol. 62, no. 4, pp. 1224–31, 2005.
[48] R. M. Lewitt and R. H. T. Bates, “Image- [59] W. J. H. Veldkamp, R. M. S. Joemai, A. J.
reconstruction from projections. III. van der Molen, and J. Geleijns,
Projection completion methods (theory),” “Development and validation of
Optik (Stuttg)., vol. 50, no. 3, pp. 189–204, segmentation and interpolation techniques in
1978. sinograms for metal artifact suppression in
[49] T. Hinderling, P. Ruegsegger, M. Anliker, CT,” Med. Phys., vol. 37, no. 2, pp. 620–628,
and C. Dietschi, “Computed Tomography 2010.
Reconstruction from Hollow Projections: An [60] H. Yu, K. Zeng, D. K. Bharkhada, G. Wang,
Application to In Vivo Evaluation of M. T. Madsen, O. Saba, B. Policeni, M. A.
Artificial Hip Joints,” J. Comput. Assist. Howard, and W. R. K. Smoker, “A
Tomogr., vol. 3, no. 1, pp. 52–57, 1979. Segmentation-Based Method for Metal
[50] G. H. Glover and N. J. Pelc, “An algorithm Artifact Reduction,” Acad. Radiol., vol. 14,
for the reduction of metal clip artifacts in CT no. 4, pp. 495–504, 2007.
reconstructions.,” Med. Phys., vol. 8, no. 6, [61] S. Zhao, D. D. Robertson, G. Wang, B.
pp. 799–807, 1981. Whiting, and K. T. Bae, “X-ray CT metal
[51] P. Seitz and P. Rüegsegger, “CT Bone artifact reduction using wavelets: an
Densitometry Of The Anchorage Of application for imaging total hip prostheses,”
Artificial Knee Joints,” J. Comput. Assist. IEEE Trans. Med. Imaging, vol. 19, no. 12,
Tomogr., vol. 9, no. 3, p. 621, 1985. pp. 1238–1247, 2000.
[52] W. A. Kalender, R. Hebel, and J. [62] A. Mehranian, M. R. Ay, A. Rahmim, and H.
Ebersberger, “Reduction of CT artifacts Zaidi, “X-ray CT metal artifact reduction
caused by metallic implants,” Radiology, vol. using wavelet domain L0 sparse
164, no. 2, pp. 576–577, 1987. regularization,” IEEE Trans. Med. Imaging,
[53] P. P. Bruyant, J. Sau, and J. J. Mallet, “Streak vol. 32, no. 9, pp. 1707–22, 2013.
artifact reduction in filtered backprojection [63] E. Meyer, F. Bergner, R. Raupach, T. Flohr,
using a level line-based interpolation and M. Kachelrieß, “Normalized metal
artifact reduction (NMAR) in computed vol. 38, no. 4, pp. 2275–2281, 2011.
tomography,” Med. Phys., vol. 37, no. 10, pp. [74] C. Xu, F. Verhaegen, D. Laurendeau, S. a
5482–5493, 2010. Enger, and L. Beaulieu, “An algorithm for
[64] J. J. Liu, S. R. Watt-Smith, and S. M. Smith, efficient metal artifact reductions in
“A description for computed tomography permanent seed,” Med. Phys., vol. 38, no. 1,
based on sinusoidal curves,” J. Xray. Sci. pp. 47–56, 2011.
Technol., vol. 11, pp. 205–218, 2003. [75] H. Xue, L. Zhang, Y. Xiao, Z. Chen, and Y.
[65] J. J. Liu, S. R. Watt-Smith, and S. M. Smith, Xing, “Metal artifact reduction in dual energy
“Metal artifact reduction for CT based on CT by sinogram segmentation based on
sinusoidal description,” J. Xray. Sci. active contour model and TV inpainting,” in
Technol., vol. 13, no. 2, pp. 85–96, 2005. IEEE Nuclear Science Symposium
[66] J. Gu, L. Zhang, G. Yu, Y. Xing, and Z. Conference, 2009, pp. 904–908.
Chen, “X-ray CT metal artifacts reduction [76] Y. Zhang, Y. F. Pu, J. R. Hu, Y. Liu, and J.
through curvature based sinogram L. Zhou, “A new CT metal artifacts reduction
inpainting,” J. Xray. Sci. Technol., vol. 14, algorithm based on fractional-order sinogram
no. 2, pp. 73–82, 2006. inpainting,” J. Xray. Sci. Technol., vol. 19,
[67] Y. Zhang, L. Zhang, X. R. Zhu, A. K. Lee, no. 3, pp. 373–384, 2011.
M. Chambers, and L. Dong, “Reducing metal [77] Q. Wang, L. Li, L. Zhang, Z. Chen, and K.
artifacts in cone-beam CT images by Kang, “A novel metal artifact reducing
preprocessing projection data,” Int. J. Radiat. method for cone-beam CT based on three
Oncol. Biol. Phys., vol. 67, no. 3, pp. 924– approximately orthogonal projections,” Phys.
932, 2007. Med. Biol., vol. 58, no. 1, pp. 1–17, 2013.
[68] R. M. S. Joemai, P. W. de Bruin, W. J. H. [78] J. Toftegaard, W. Fledelius, D. Seghers, M.
Veldkamp, and J. Geleijns, “Metal artifact Huber, M. Brehm, E. S. Worm, U. V.
reduction for CT: Development, Elstrøm, and P. R. Poulsen, “Moving metal
implementation, and clinical comparison of a artifact reduction in cone-beam CT scans
generic and a scanner-specific technique,” with implanted cylindrical gold markers,”
Med. Phys., vol. 39, no. January, p. 1125, Med. Phys., vol. 41, no. 12, p. 121710, 2014.
2012. [79] C. R. Crawford, J. G. Colsher, N. J. Pelc, and
[69] B. Kratz, I. Weyers, and T. M. Buzug, “A A. H. R. Lonn, “High speed reprojection and
fully 3D approach for metal artifact reduction its applications,” in SPIE Medical Imaging II,
in computed tomography,” Med. Phys., vol. 1988, pp. 311–318.
39, no. 11, p. 7042, 2012. [80] R. Naidu, I. Bechwati, S. Karimi, S.
[70] L. Yu, H. Li, J. Mueller, J. M. Kofler, X. Liu, Simanovsky, and C. Crawford, “Method of
A. N. Primak, J. G. Fletcher, L. S. and system for reducing metal artifacts in
Guimaraes, T. Macedo, and C. H. images generated by x-ray scanning devices,”
McCollough, “Metal artifact reduction from 6,721,387 B1, 2004.
reformatted projections for hip prostheses in [81] H. K. Tuy, “A post-processing algorithm to
multislice helical computed tomography: reduce metallic clip artifacts in CT images,”
techniques and initial clinical results.,” Eur. Radiol., vol. 3, no. 2, pp. 129–134,
Invest. Radiol., vol. 44, no. 11, pp. 691–696, 1993.
2009. [82] K. Y. Jeong and J. B. Ra, “Metal artifact
[71] H. Li, L. Yu, X. Liu, J. G. Fletcher, and C. H. reduction based on sinogram correction in
McCollough, “Metal artifact suppression CT,” in IEEE Nuclear Science Symposium
from reformatted projections in multislice Conference, 2009, pp. 3480–3483.
helical CT using dual-front active contours,” [83] C. S. Olive, M. R. Klaus, V. Pekar, K. Eck,
Med. Phys., vol. 37, no. 10, p. 5155, 2010. and L. Spies, “Segmentation aided adaptive
[72] S. Tohnak, A. J. H. Mehnert, M. Mahoney, filtering for metal artifact reduction in radio-
and S. Crozier, “Dental CT metal artefact therapeutic CT images,” in SPIE Medical
reduction based on sequential substitution,” Imaging 2004, 2004, vol. 5370, pp. 1991–
Dentomaxillofacial Radiol., vol. 40, no. 3, 2002.
pp. 184–190, 2011. [84] M. Bal and L. Spies, “Metal artifact reduction
[73] M. Yazdi, M. A. Lari, G. Bernier, and L. in CT using tissue-class modeling and
Beaulieu, “An opposite view data adaptive prefiltering,” Med. Phys., vol. 33,
replacement approach for reducing artifacts no. 8, p. 2852, 2006.
due to metallic dental objects.,” Med. Phys., [85] C. Lemmens, D. Faul, and J. Nuyts,
“Suppression of metal artifacts in CT using a [97] E. Meyer, R. Raupach, B. Schmidt, A. H.

reconstruction procedure that combines MAP Mahnken, and M. Kachelriess, “Adaptive
and projection completion,” IEEE Trans. Normalized Metal Artifact Reduction
Med. Imaging, vol. 28, no. 2, pp. 250–60, (ANMAR) in Computed Tomography,” in
2009. 2011 IEEE Nuclear Science Symposium
[86] D. Prell, Y. Kyriakou, M. Beister, and W. A. Conference Record, 2011, pp. 2560–2565.
Kalender, “A novel forward projection-based [98] M. M. Lell, E. Meyer, M. A. Kuefner, M. S.
metal artifact reduction method for flat- May, R. Raupach, M. Uder, and M.
detector computed tomography,” Phys. Med. Kachelriess, “Normalized metal artifact
Biol., vol. 54, no. 21, pp. 6575–6591, 2009. reduction in head and neck computed
[87] D. Prell, Y. Kyriakou, T. Struffert, A. tomography,” Invest. Radiol., vol. 47, no. 7,
Dörfler, and W. a Kalender, “Metal artifact pp. 415–421, 2012.
reduction for clipping and coiling in [99] B. P. Medoff, W. R. Brody, M. Nassi, and A.
interventional C-arm CT,” AJNR. Am. J. Macovski, “Iterative convolution
Neuroradiol., vol. 31, no. 4, pp. 634–9, 2010. backprojection algorithms for image
[88] Philips Healthcare, “Metal Artifact Reduction reconstruction from limited data,” J. Opt.
for Orthopedic Implants (O-MAR),” White Soc. Am., vol. 73, no. 11, p. 1493, 1983.
Paper, 2012. . [100] G. Wang, D. L. Snyder, J. O’Sullivan, and M.
[89] S. Karimi, P. Cosman, C. Wald, and H. W. Vannier, “Iterative deblurring for CT
Martz, “Segmentation of artifacts and metal artifact reduction,” IEEE Trans. Med.
anatomy in CT metal artifact reduction,” Imaging, vol. 15, no. 5, pp. 657–664, 1996.
Med. Phys., vol. 39, no. 10, p. 5857, 2012. [101] G. Wang, M. W. Vannier, and P.-C. Cheng,
[90] S. Karimi, H. Martz, and P. Cosman, “Metal “Iterative X-ray Cone-Beam Tomography for
artifact reduction for CT-based luggage Metal Artifact Reduction and Local Region
screening,” J. Xray. Sci. Technol., vol. 23, no. Reconstruction,” Microsc. Microanal., vol. 5,
4, pp. 435–451, 2015. no. 1, pp. 58–65, 1999.
[91] J. Wang, S. Wang, Y. Chen, J. Wu, J.-L. [102] G. Wang, T. Frei, and M. Vannier, “Fast
Coatrieux, and L. Luo, “Metal artifact iterative algorithm for metal artifact
reduction in CT using fusion based prior reduction in X-ray CT,” Acad. Radiol., vol. 7,
image,” Med. Phys., vol. 40, no. 8, p. 81903, no. 8, pp. 607–614, 2000.
2013. [103] J. August and T. Kanade, “Fast streaking
[92] M. Li, J. Zheng, T. Zhang, Y. Guan, P. Xu, artifact reduction in CT using constrained
and M. Sun, “A prior-based metal artifact optimization in metal masks,” in Medical
reduction algorithm for x-ray CT,” J. Xray. Image Computing and Computer-Assisted
Sci. Technol., vol. 23, no. 2, pp. 229–241, Intervention–MICCAI 2004, Springer, 2004,
2015. pp. 1044–1045.
[93] T. Heußer, M. Brehm, L. Ritschl, S. Sawall, [104] B. De Man, “Method and apparatus for the
and M. Kachelrieß, “Prior-based artifact reduction of artifacts in computed
correction (PBAC) in computed tomography images,” US 7444010 B2, 2008.
tomography,” Med. Phys., vol. 41, no. 2, p. [105] B. De Man, J. Nuyts, P. Dupont, G. Marchal,
021906, 2014. and P. Suetens, “Reduction of metal streak
[94] Y. Zhang, H. Yan, X. Jia, J. Yang, S. B. artifacts in X-ray computed tomography
Jiang, and X. Mou, “A hybrid metal artifact using a transmission maximum a posteriori
reduction algorithm for x-ray CT,” Med. algorithm,” IEEE Trans. Nucl. Sci., vol. 47,
Phys., vol. 40, no. 4, p. 041910, 2013. no. 3, pp. 977–981, 2000.
[95] P. Bannas, Y. Li, U. Motosugi, K. Li, M. [106] B. De Man, “Iterative reconstruction for
Lubner, G.-H. Chen, and P. J. Pickhardt, reduction of metal artifacts in computed
“Prior Image Constrained Compressed tomography,” University of Leuven, 2001.
Sensing Metal Artifact Reduction (PICCS- [107] N. Menvielle, “Reduction des artefacts
MAR): 2D and 3D Image Quality metalliques en tomographie a rayons X,”
Improvement with Hip Prostheses at CT Ecole Polytechnique de Montreal, 2004.
Colonography,” Eur. Radiol., pp. 1–8, 2015. [108] B. De Man, J. Nuyts, P. Dupont, G. Marchal,
[96] J. Müller and T. M. Buzug, “Spurious and P. Suetens, “An iterative maximum-
structures created by interpolation-based CT likelihood polychromatic algorithm for CT,”
metal artifact reduction,” in SPIE Medical IEEE Trans. Med. Imaging, vol. 20, no. 10,
Imaging, 2009, p. 72581Y–72581Y. pp. 999–1008, 2001.
[109] B. Hamelin, Y. Goussard, D. Gendron, J.-P. [119] J. A. O’Sullivan and J. Benac, “Alternating
Dussault, G. Cloutier, G. Beaudoin, and G. minimization algorithms for transmission
Soulez, “Iterative CT reconstruction of real tomography,” IEEE Trans. Med. Imaging,
data with metal artifact reduction,” in IEEE vol. 26, no. 3, pp. 283–297, 2007.
International Symposium on Biomedical [120] R. J. Murphy, S. Yan, J. A. O’Sullivan, D. L.
Imaging, 2008, no. 1, pp. 1453–1456. Snyder, B. R. Whiting, D. G. Politic, G.
[110] K. Van Slambrouck and J. Nuyts, “Metal Lasio, and J. F. Williamson, “Pose estimation
artifact reduction in computed tomography of known objects during transmission
using local models in an image block- tomographic image reconstruction,” IEEE
iterative scheme,” Med. Phys., vol. 39, no. Trans. Med. Imaging, vol. 25, no. 10, pp.
11, p. 7080, 2012. 1392–1404, 2006.
[111] M. Oehler and T. M. Buzug, “Statistical [121] J. W. Stayman, Y. Otake, J. L. Prince, A. J.
image reconstruction for inconsistent CT Khanna, and J. H. Siewerdsen, “Model-based
projection data,” Methods Inf Med, vol. 46, tomographic reconstruction of objects
no. 3, pp. 261–269, 2007. containing known components,” IEEE Trans.
[112] B. Kratz and T. Buzug, “Metal artifact Med. Imaging, vol. 31, no. 10, pp. 1837–
reduction in computed tomography using 1848, 2012.
nonequispaced Fourier transform,” in IEEE [122] G. Henrich, “A simple computational method
Nuclear Science Symposium Conference, for reducing streak artifacts in CT images,”
2009, pp. 2720–2723. Comput. Tomogr., vol. 4, no. 1, pp. 67–71,
[113] S. Aootaphao, C. Pintavirooj, and S. 1980.
Sotthivirat, “Penalized-likelihood [123] H. Soltanian-Zadeh, J. P. Windham, and J.
reconstruction for metal artifact reduction in Soltanianzadeh, “CT artifact correction: an
cone-beam CT,” in Engineering in Medicine image-processing approach,” in Medical
and Biology Society, 2008. EMBS 2008. 30th Imaging 1996, 1996, pp. 477–485.
Annual International Conference of the [124] M. Bal, H. Celik, K. Subramanyan, K. Eck,
IEEE, 2008, pp. 2733–2736. and L. Spies, “A radial adaptive filter for
[114] X. Zhang, J. Wang, and L. Xing, “Metal metal artifact reduction,” in SPIE Medical
artifact reduction in x-ray computed Imaging 2005, 2005, vol. 5747, pp. 2075–
tomography (CT) by constrained 2082.
optimization,” Med. Phys., vol. 38, no. 2, pp. [125] O. Watzke and W. A. Kalender, “A
701–711, 2011. pragmatic approach to metal artifact
[115] A. Kalvin and B. Williamson, “Using scout reduction in CT: Merging of metal artifact
images to reduce metal artifacts in CT with reduced images,” Eur. Radiol., vol. 14, no. 5,
application to revision total hip surgery,” in pp. 849–856, 2004.
SPIE Medical Imaging 1997, 1997, vol. [126] E. Meyer, R. Raupach, M. Lell, B. Schmidt,
3034, pp. 1017–1028. and M. Kachelrieß, “Frequency split metal
[116] A. D. Kalvin, “System and Method for artifact reduction (FSMAR) in computed
Reducing Reconstruction Artifacts in tomography,” Med. Phys., vol. 39, no. 4, p.
Computed Tomography Images,” 5,933,471, 1904, 2012.
1999. [127] A. Mouton, N. Megherbi, G. T. Flitton, S.
[117] G.-H. Chen, J. Tang, and S. Leng, “Prior Bizot, and T. P. Breckon, “A novel intensity
image constrained compressed sensing limiting approach to Metal Artefact
(PICCS): a method to accurately reconstruct Reduction in 3D CT baggage imagery,” in
dynamic CT images from highly Proceedings - International Conference on
undersampled projection data sets,” Med. Image Processing, ICIP, 2012, pp. 2057–
Phys., vol. 35, no. 2, pp. 660–663, 2008. 2060.
[118] D. L. Snyder, J. A. O’Sullivan, B. R. [128] H. Ballhausen, M. Reiner, U. Ganswindt, C.
Whiting, R. J. Murphy, J. Benac, J. A. Belka, and M. Söhn, “Post-processing sets of
Cataldo, D. G. Politte, and J. F. Williamson, tilted CT volumes as a method for metal
“Deblurring subject to nonnegativity artifact reduction,” Radiat. Oncol., vol. 9, no.
constraints when known functions are present 1, p. 114, 2014.
with application to object-constrained [129] J. Y. Huang, J. R. Kerns, J. L. Nute, X. Liu,
computerized tomography,” IEEE Trans. P. A. Balter, F. C. Stingo, D. S. Followill, D.
Med. Imaging, vol. 20, no. 10, pp. 1009– Mirkovic, R. M. Howell, and S. F. Kry, “An
1017, 2001. evaluation of three commercially available
metal artifact reduction methods for CT

imaging,” Phys. Med. Biol., vol. 60, no. 3, p.
1047, 2015.
[130] M. Kidoh, T. Nakaura, S. Nakamura, S.
Tokuyasu, H. Osakabe, K. Harada, and Y.
Yamashita, “Reduction of dental metallic
artefacts in CT: value of a newly developed
algorithm for metal artefact reduction (O-
MAR),” Clin. Radiol., vol. 69, no. 1, pp.
e11–e16, 2014.
[131] Y. Xi, Y. Jin, B. De Man, and G. Wang,
“High-kVp Assisted Metal Artifact
Reduction for X-ray Computed
Tomography,” 2016.
[132] L. Liu, “Model-based iterative
reconstruction: a promising algorithm for
today’s computed tomography imaging,” J.
Med. Imaging Radiat. Sci., vol. 45, no. 2, pp.
131–136, 2014.
[133] S. Boudabbous, D. Arditi, E. Paulin, A.
Syrogiannopoulou, C. Becker, and X.
Montet, “Model-Based Iterative
Reconstruction (MBIR) for the reduction of
metal artifacts on CT,” Am. J. Roentgenol.,
vol. 205, no. 2, pp. 380–385, 2015.
[134] G. Delso, S. Wollenweber, A. Lonn, F.
Wiesinger, and P. Veit-Haibach, “MR-driven
metal artifact reduction in PET/CT,” Phys.
Med. Biol., vol. 58, no. 7, p. 2267, 2013.
[135] A. K. Louis and A. Rieder, “Incomplete data
problems in X-ray computerized
tomography,” Numer. Math., vol. 56, no. 4,
pp. 371–383, 1989.
[136] P. Maass, “The interior Radon transform,”
SIAM J. Appl. Math., vol. 52, no. 3, pp. 710–
724, 1992.
[137] D. Silver, A. Huang, C. J. Maddison, A.
Guez, L. Sifre, G. van den Driessche, J.
Schrittwieser, I. Antonoglou, V.
Panneershelvam, M. Lanctot, S. Dieleman,
D. Grewe, J. Nham, N. Kalchbrenner, I.
Sutskever, T. Lillicrap, M. Leach, K.
Kavukcuoglu, T. Graepel, and D. Hassabis,
“Mastering the game of Go with deep neural
networks and tree search,” Nature, vol. 529,
no. 7587, pp. 484–489, Jan. 2016.

Metal Artifact Reduction in CT: Where Are We After Four Decades?

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Metal Artifact Reduction in CT: Where Are We After Four Decades?

Uploaded by

Copyright:

Available Formats

This article has been accepted for publication in a future issue of this journal, but has not been

Metal Artifact Reduction in CT:

*Corresponding Authors: deman@ge.com and ge-wang@ieee.org

1. INTRODUCTION greatly attenuate low-energy x-rays. As the

monoenergetic extrapolation” by Bamberg et

Figure 1. Total annual publications on CT MAR

Small metal Minor shading in

3. METAL ARTIFACT REDUCTION that employ techniques in more than one of

chrome artificial hips by Haramati et al. with medium-level characteristics. The

3.2.2 Dual-energy CT conclusions when extrapolating single

model of an iterative reconstruction technique by correcting for beam hardening

To improve projection completion projection is hollow, the density distribution

gradient parameter for regularization, which A tissue-class model is a segmentation of

3.5 Iterative Reconstruction function is deterministic, meaning it is not

data consistency, while smoothness is Like other statistical reconstruction

maximum-likelihood. Murphy et al. aim to reconstructed, and do not rely on access to

3.6.2 Hybrid hardening and image noise. However, this

Figure 12. Images of two cobalt-chrome

Meyer et al. further introduced a processing correction may be sufficient, as

[25] S. G. Moon, S. H. Hong, J.-Y. Choi, W. S. 1257, 2011.

“Suppression of metal artifacts in CT using a [97] E. Meyer, R. Raupach, B. Schmidt, A. H.

metal artifact reduction methods for CT

You might also like