Synopsis
Keywords: Image acquisition: Artefacts
Image artifacts can arise from one or more of the following:
MR physics, MR system hardware, and the patient.
Given the wide range of possible MR artifacts this presentation will be limited
to the main relationships between data corruptions during image acquisition,
i.e., in raw data or k-space, and their
appearance in the reconstructed image after Fourier transformation. The
cause and effect of artifacts including Gibbs ringing, phase encode direction aliasing/foldover,
zippers, spikes and motion during image acquisition will be discussed, together
with failures in fat suppression and the effect of receiver bandwidth on
chemical shift artifact.
Introduction
MRI is undoubtedly the most powerful and versatile
diagnostic imaging method ever created. With this power, however, comes
complexity and, some features in an image may be unintended, unexpected, and
poorly understood. We generally refer to these features as artifacts since they
may lead to interpretive difficulties or errors. The best remedy for MRI
artifacts and their potential to produce diagnostic errors is to promote
understanding, so that they can be prevented or, alternatively, one can potentially
utilize the information content of an artifact to contribute to diagnostic
content.
Image artifacts can arise from one
or more of the following: MR physics; MR system hardware; the patient. Given
the wide range of MR artifacts I have chosen to limit this description of basic
artifacts to the main relationships between data corruptions during image
acquisition, i.e., in the raw data or k-space, and their appearance in
the reconstructed image, i.e., after Fourier transformation. For
descriptions of other artifacts, the reader is directed to the other
presentations in this session as well as review articles in the literature (1-8).k-space vs. image space
To understand the cause of
artifacts in reconstructed images it is necessary to understand the
relationship between k-space and image space. The most common method of
acquiring MR images is to use the spin-warp method (9), developed
at the University of Aberdeen in 1980. Following slice selective excitation that
tips the net magnetisation into the transverse plane, the spin warp method uses
a variable area phase encoding gradient to impart a phase twist, or warp, to
the spins in the phase encoding direction. This spatial variation in the spin
phase is retained when the MR echo is digitised in the presence of the
frequency encoding gradient. Each echo is then saved in the MR raw data or k-space.
The echoes acquired with the smallest area, i.e., the lowest ± amplitude phase
encoding gradients, that impart a 2π phase twist across the phase field-of-view
($$$FOV_y$$$), are placed
either side of the centre of k-space, i.e., at the $$$± 1\: k_y$$$ lines. The echoes acquired
with the largest ± amplitude phase encoding gradients, that impart a 2π phase
shift across adjacent voxels, are placed at either end of k-space, i.e., $$$± k_{y,max}/{2}$$$. The
other phase encoding gradient areas are linearly scaled between these two
extremes, such that the difference between adjacent phase encoding steps $$$\Delta k\ =1/{FOV_y}$$$, and $$$k_{y,max}=\frac{1}{2}(N_y-1)\Delta k_y$$$, where
$$$N_y$$$ is the number of phase encoding steps. The
maximum area of the phase encoding gradient is given by $$$A_{y,max}=\frac{\ 2\pi\ (N_y-1)}{\ \gamma\ FOV_y}$$$. Following phase encoding
the echo signal is then frequency encoded in the orthogonal direction. The
frequency encoding gradient causes the phase encoded spins to precess at
spatially varying frequencies along the direction of the gradient. Defining the spatial frequencies $$$ k_x=\gamma\int{G_x\left(t\right)}\: dt$$$ and $$$ k_y=\gamma\int{G_y\left(t\right)}\: dt$$$, then
for simple constant amplitude gradients the signal can be written as $$$S\left(k_x,k_y\right)=\iint{\rho\left(x,y\right)e^{-ik_xx}e^{-ik_yy}dxdy}$$$, where
$$$\rho\left(x,y\right)$$$ is the effective spatial distribution of the
proton density of the tissues. Hence the use of the spin -warp method
means that image reconstruction can be directly performed using a 2D inverse Fourier
transform.
The inverse relationship
between k-space and image space (Figure 1) means
that the spacing between phase encode steps ($$$\Delta k_y$$$) or
frequency encode samples ($$$\Delta k_x$$$) dictates
the field-of-view (FOV) in the phase encode and frequency encoding direction respectively.
Similarly, this inverse relationship means that the spatial resolution in the
phase encode direction is dictated by the extent of k-space, i.e.,$$$\Delta y=1/2k_{y,max}$$$ , with a
similar equation for the spatial resolution in the frequency encoding direction.Gibbs ringing
The Fourier transform assumes that the signal being
transformed is periodic and extends infinitely in all directions. However, in
practice, the acquired MRI data has a finite extent and is truncated in
comparison to the idealised situation. The consequence of this truncation is
that the Fourier transform cannot accurately represent the transition of a
signal with a sharp discontinuity, such as an edge. The Gibbs phenomenon is a
well-known mathematical property where, during the Fourier transformation of a
signal with such a sharp discontinuity, oscillatory overshoots or undershoots
occur near the transition region. These oscillations manifest as bright and
dark lines parallel to, but decaying in amplitude with distance from, the
discontinuity. In the spatial domain,
this is known as Gibbs ringing (Figure 2).
Josiah Willard Gibbs was Professor of mathematical
physics at Yale from 1871 until his death in 1903. Working in relative
isolation, he became the earliest theoretical scientist in the United States to
earn an international reputation and was praised by Albert
Einstein as "the greatest mind in American history”. In 1899, inspired
by correspondence in Nature between A. A. Michelson (Chicago) and A.
E. H. Love (Oxford) about the convergence of the Fourier series of the square
wave function, Gibbs (after making a mistake in his initial communication) eventually
described this overshoot at the point of
a discontinuity (10). In 1906, the American
mathematician Maxime Bôcher gave a detailed mathematical analysis of the
overshoot, coining the term "Gibbs phenomenon" (11)
and bringing the term into widespread use.
To mitigate Gibbs ringing, one common approach is to apply a
low-pass filter to the raw k-space data that reduces the amplitude of
the high-frequency components at the edge of k-space. This data
apodisation can help reduce the ringing, but at the cost of some loss of
sharpness or fine details in the reconstructed image. Alternatively higher
spatial frequencies can be acquired, i.e., more $$$N_x$$$ and/or $$$N_y$$$ samples can be acquired, resulting in a
smaller voxel size and hence a higher spatial resolution. This has the effect
of moving the oscillations towards the discontinuity but does not decrease their
magnitude.Phase-encode direction aliasing or foldover
MRI allows
different FOVs to be selected in the frequency and phase encoding direction. In
the frequency encoding direction, the FOV and the readout (receiver) bandwidth
are selected by the operator. These two pieces of information allow the system
to calculate the desired frequency encoding gradient strength. The readout
bandwidth information is also used by the system to electronically filter out
any frequencies outside of the bandwidth. However, in the phase encoding
direction, the signal encoding is achieved through the phase imparted to the
spins (9), so there is no direct analogue to frequency filtering. The desired phase FOV can also be chosen by
the operator but in this case any tissue that exists outside this FOV will be
encoded incorrectly. The minimum ± phase encoding gradient amplitudes are calculated
to achieve a 2π phase shift across the selected FOV. This means that any tissue beyond
the FOV will be phase aliased (Figure 3). For example, if the FOV is 30cm then tissue at
32cm will appear at 2cm, i.e., it will appear on the other side of the image.
Usually, phase
encode aliasing, sometimes known as foldover, is easy to recognize, but the
resulting image can be confusing if the aliased body part is completely outside
the FOV so that continuity from one side of the image to the other is difficult
to recognize. Examples include an arm or hand mimicking a lesion or enhancing
vessels in the mediastinum resembling an enhancing tumour in the breast. Simply
swapping the phase and frequency directions, so that all the tissue along the
phase axis is within the FOV may be sufficient to eliminate these artifacts.
However, if there are good reasons to place the direction of the phase encoding
axis along the long axis of the patient, e.g., to control the direction of
motion artifacts (see Patient motion below), or if both axes extend beyond the desired
FOV then it may be necessary to employ no-phase-wrap techniques. These methods extend the original FOV in the
phase encode direction, commonly by a factor of two, whilst commensurately
reducing the number of signal averages to maintain the pixel size and
signal-to-noise ratio. Tissue signal outside of the extended FOV aliases into
the regions of extended FOV and are discarded resulting in the original FOV
image being displayed but without any aliasing artifact. However, if the body
part is sufficiently large, aliasing may still degrade an image.
With 3D
Fourier transform acquisitions, a second orthogonal phase encoding gradient is
used to encode slices. Even though the excitation is often slab selective the
profile of the excitation slab is not sharp at the edges, which means that some
signal will be excited outside the range of the slice axis phase encoding.
Therefore, data can be aliased in this direction as well as the axis of
in-plane phase encoding. Because the aliased structures are at a level distant
from the reconstructed image, this form of phase aliasing can be more confusing
to understand. Viewing the top and bottom slices can usually clarify this
artifact. RF interference/zipper artifacts
Many image artifacts result from undesired RF signals that
corrupt the MR data. This undesired RF may arise externally or internally to
the MR system. External sources are usually minimized by locating the magnet
inside an RF shielded enclosure or cabin. This cabin provides a high degree of
immunity from external interference and prevents leakage of the high-power RF
from the MR system. However, if there is a problem with this enclosure then
external RF can enter the MR receiver system. If this interference is periodic
with either single or multiple frequencies, then the artifact typically appears
as a bright and dark alternating pattern in the phase encoding direction that
is often referred to as a zipper artifact.
A more uncommon form of zipper artifact may be seen parallel
to the frequency encoding direction along a line at magnet isocentre (Figure 4). This artifact
may be seen with spin echo-based sequences and is due to the transverse
magnetisation created by the imperfect 180°. This free induction decay (FID)
signal may then overlap the beginning of the digitised spin echo signal. Since
this FID is created after phase encoding its amplitude is constant during each phase
encoding, i.e., it can be considered as a DC offset. Following Fourier
transformation this DC offset manifests as a line through the centre of the
gradient field, i.e., not necessarily through the centre of the image. The FID is usually “crushed” by applying
balanced dephasing gradient lobes either side of the 180° slice selection
gradient, with the gradient lobe on the right doing the dephasing and the
gradient on the left balancing the phase shift imparted by the right lobe. In addition,
the RF excitation pulse is usually “chopped”, i.e., the 90°
excitation pulse has a 180° phase shift applied on alternate TRs. This has the
effect of pushing the DC artifact to the edges of the image FOV, where the
images lines can be blanked if required.Spike noise
The
periodic RF interference described above should be contrasted with short
intense artifacts that affect only single points in k-space, called
“spike-noise” or “white pixel noise” (12). One or more spikes of detected external noise
may produce a patterned degradation of the image. Depending upon where this
point occurs in k-space the overall image, after Fourier transformation,
will demonstrate a superimposed banding artifact corresponding to that spatial
frequency (Figure 5). A spike occurring near the centre of k-space will have a low
frequency banding whilst a spike occurring near the edge of k-space will
generate a high frequency banding. This cross-hatching appearance is sometimes
referred to as a corduroy (single spike) or herringbone (multiple spikes)
artifact. Causes of such artifact can include static electricity from clothing
or blankets, or random noise from electrical sources such as damaged filament
light bulbs, or electrical discharges from gradient cable connections. Patient motion
The MRI acquisition process assumes that the underlying
tissue does not change in location or signal intensity during the acquisition.
Any changes can cause modulation of the MRI signal, creating artifacts (13, 14). Since a single frequency encoding is very
rapid, motion does not generally cause artifacts in the frequency encoding
direction. However, the interval between phase encoding steps in a conventional
sequence, i.e., every TR, is more temporally commensurate with the rates of
physiological motion. Hence motion artifacts, regardless of the direction of
the motion, will manifest as artifacts in the phase encoding direction, usually
as “ghosting” whereby replicates of the moving structure are propagated along
the phase encoding direction. Any type of movement can give rise to such
ghosting appearances, including whole body motion, respiration, cardiac motion,
and blood flow. Many artifact reduction strategies focus on either removing the
motion (e.g., breath-holding), compensating for the motion (e.g.,
ECG-triggering, respiratory triggering, or reordering phase encoding),
suppressing the signal of moving tissue, e.g., spatial saturation (15),
signal averaging (16)
or by using rapid imaging techniques. In some situations, it may be possible to
swap the phase and frequency directions (see Phase-encode direction aliasing above) to
change the direction of the artifact and stop it from overlying a particular
area of interest.
Respiratory motion causes an anterior-posterior motion of
the chest wall (in-plane) as well as inferior-superior motion of the liver
(through-plane) resulting in a modulation of the MRI signal amplitude with
time. After Fourier transformation this modulated signal results in “sidebands”
or ghosts, with the spacing of the ghosts proportional to the periodicity of
the motion and inversely proportion to the TR. These sidebands appear as ghosts
in the reconstructed image (Figure 6). Note that with the use of surface coil arrays the
signal from subcutaneous fat is very high and can exacerbate the appearance of
the ghosts.
Similar amplitude modulation artifacts result from pulsatile
flow in a vessel (17).
Although the vessel itself is not moving the changing signal intensity with
differential flow related signal enhancement throughout the cardiac cycle
result in ghosting artifacts. It should
also be noted that any system hardware instabilities can also cause signal
modulation and hence ghosting in images.
Another category of motion induced artifact that can cause
signal loss and ghosting arises from the phase shifts accumulated by spins
moving during the application of the imaging gradients (18).
This category of motion artifact can be greatly reduced by using gradient
moment nulling techniques, called flow-compensation by some vendors, which
involves using more complex gradient waveforms, usually on the slice select and
frequency encoding axes, so that the phase accumulated by moving spins is zero,
or nulled, at the end of the gradient waveform (19, 20).
The simplest form of gradient moment nulling involves correcting the phase
shift due to spins moving with a constant velocity, known as the first order
moment, which is the largest source of phase errors (Figure 7). Compensating for higher
orders of motion, such as acceleration or jerk, requires more complex and
longer duration gradient waveforms, meaning that they may cause more problems than
they address. It should be noted that even complete correction of gradient
moment errors does not affect amplitude modulation, so that ghosting is still
likely from pulsatile flow. Some forms of motion artifact result from changes in tissue
position that do not necessarily interfere with the process of phase encoding,
and therefore do not produce ghost artifact along the phase encoding axis. As
with any method of imaging, including conventional radiography, motion during
image acquisition can produce artifacts. For example, image blurring or complex
signal may occur when tissue occupies various positions during the period of
image acquisition. Chemical shift selective fat saturation failure
A good static magnetic field (B0) uniformity is
required for all MR imaging. Manufacturing tolerances and the local
installation environment means that the main magnetic field needs to be
optimized at installation by the manufacturer who “shims” the magnet either
passively, using small pieces of metal distributed around the room temperature
bore of the magnet, or actively by adjusting the currents through additional
superconducting coils within the magnet cryostat.
Magnetic field uniformity is usually reported as the root
mean square (rms) variation from the nominal field strength, e.g., 3.0 T, in
either parts-per-million (ppm), or as a frequency variation in Hz, over many
points within a given diameter spherical volume (DSV). A typical 3.0 T manufacturer’s shim would
have a rms value of 50 Hz or 0.4 ppm over a 45 cm DSV. Since heterogeneous
susceptibility within the human body will affect local magnetic field
uniformity, manufacturers also provide methods of patient specific shimming.
This involves rapidly mapping the static magnetic field uniformity, usually
using a gradient echo-based phase difference technique, and adjusting the
offsets on the x, y and z gradients to remove any linear field variations (21).
Some systems may be equipped with additional room-temperature, i.e.,
non-superconducting, higher order shim coils that can further improve the
uniformity, particularly over small volumes.
Uniformity of B0 is important for chemical shift
selective fat saturation techniques. Human adipose tissue is composed primarily
of triglycerides, which are a subgroup of lipid molecules. Triglycerides comprise multiple groups of protons
(CH3, CH2, CH=CH, etc) with the nuclear shielding effect resulting in a range
of chemical shifts from 0.9 to 5.3 ppm lower than the water (H20)
proton resonance. The most abundant triglyceride resonance comes from the
methylene (CH2) groups, in which the protons resonate at
approximately 3.4 ppm lower than those in water. Hence at 1.5 T the methylene
peak appears at 3.4 ppm * 64 MHz = 220 Hz lower than the water resonant
frequency. Chemical shift selective saturation pulses at 1.5 T are therefore centred
around -220 Hz from water. These pulses have a bandwidth that is sufficiently
narrow to avoid saturating the water signal. A typical 1.5 T chemical shift
saturation pulse has a bandwidth of approximately 150 Hz. Therefore, to obtain
good fat saturation, it is necessary that the main magnetic field does not vary
by more than 150 Hz across the entire field-of-view and across all slices in a
multi-slice acquisition. If this is not the case the fat saturation will be
suboptimal (Figure 8).
Magnetic field non-uniformity due to patient susceptibility
is particularly problematic at air tissue interfaces, which are especially
abundant in various locations within the chest, including the breast. Failure
of fat saturation is almost inevitable when the air tissue interface is
oriented along the z-axis of the magnetic field (22-24).
One particularly common example is adipose tissue situated anterior to the
liver, immediately inferior to the lung base. Heterogeneous susceptibility
within the patient can also cause the water frequency to be shifted so that it
falls within the bandwidth of the fat saturation pulse. In this situation the
water signal can also become saturated.
This effect has also been reported in contrast-enhanced thoracic
angiography (25). Chemical shift artifact
The
frequency encoding gradient is used to spatially localize signals based on
their precessional frequency. The
natural difference in frequency of 3 ppm between the protons in water and fat
means that after Fourier transformation they will be spatially shifted in the
frequency encoding direction (Figure 9) (26,
27). The amount of shift will
depend upon the receiver bandwidth per pixel and the chemical shift at the
given field strength. For example, a typical receiver bandwidth of ±16 kHz over
256 pixels in the frequency encoding direction is equivalent to $$$32000/256 = 125 Hz/pixel$$$. At 1.5 T
the chemical shift induced frequency offset is approximately 220 Hz (see Chemical
shift selective fat saturation failure above), so water and fat will be
spatially shifted by $$$220/125 = 1.76\: pixels$$$. Reducing the receiver bandwidth will increase
the chemical shift offset and vice versa. Note that moving from 1.5 T to 3.0 T
will double the offset for a given bandwidth.Summary
Image
artifacts usually arise during MR acquisition; therefore, it is important to
have a good understanding of the relationship between raw data, or k-space
and their appearance in the reconstructed image after the inverse Fourier transformation. Artifacts can arise from one or more of MR physics, the MR
system hardware, as well as the patient.Acknowledgements
Thanks to colleagues at the University of Cambridge Department of Radiology and the Cambridge University Hospitals Department of Imaging.References
- Bellon EM, Haacke EM, Coleman PE,
Sacco DC, Steiger DA, Gangarosa RE. MR artifacts: a review. Ajr.
1986;147(6):1271-81.
-
Bernstein MA, Huston J, 3rd, Ward
HA. Imaging artifacts at 3.0T. J Magn Reson Imaging. 2006;24(4):735-46.
-
Mirowitz SA. MR imaging artifacts.
Challenges and solutions. Magnetic resonance imaging clinics of North America.
1999;7(4):717-32.
-
Morelli JN, Runge VM, Ai F,
Attenberger U, Vu L, Schmeets SH, et al. An image-based approach to
understanding the physics of MR artifacts. Radiographics. 2011;31(3):849-66.
-
Stadler A, Schima W, Ba-Ssalamah A,
Kettenbach J, Eisenhuber E. Artifacts in body MR imaging: their appearance and
how to eliminate them. European radiology. 2007;17(5):1242-55.
-
Zhuo J, Gullapalli RP. AAPM/RSNA
physics tutorial for residents: MR artifacts, safety, and quality control.
Radiographics. 2006;26(1):275-97.
-
Graves MJ, Mitchell DG. Body MRI
artifacts in clinical practice: a physicist's and radiologist's perspective. J
Magn Reson Imaging. 2013;38(2):269-87.
-
Krupa K, Bekiesinska-Figatowska M.
Artifacts in magnetic resonance imaging. Pol J Radiol. 2015;80:93-106.
-
Edelstein WA, Hutchison JM, Johnson
G, Redpath T. Spin warp NMR imaging and applications to human whole-body
imaging. Physics in medicine and biology. 1980;25(4):751-6.
-
Gibbs JW. Fourier's Series. Nature.
1899(April 27):606.
-
Bôcher M. Introduction to the theory
of Fourier's series. Annals of Mathematics. 1906;7(3):81-152.
-
Foo TF, Grigsby NS, Mitchell JD,
Slayman BE. SNORE: spike noise removal and detection. IEEE transactions on
medical imaging. 1994;13(1):133-6.
-
Barish MA, Jara H. Motion artifact
control in body MR imaging. Magnetic resonance imaging clinics of North
America. 1999;7(2):289-301.
-
Wood ML, Henkelman RM. MR image
artifacts from periodic motion. Medical physics. 1985;12(2):143-51.
-
Felmlee JP, Ehman RL. Spatial
presaturation: a method for suppressing flow artifacts and improving depiction
of vascular anatomy in MR imaging. Radiology. 1987;164(2):559-64.
-
Gazelle GS, Saini S, Hahn PF, Goldberg
MA, Halpern EF. MR imaging of the liver at 1.5 T: value of signal averaging in
suppressing motion artifacts. Ajr. 1994;163(2):335-7.
-
Perman WH, Moran PR, Moran RA,
Bernstein MA. Artifacts from pulsatile flow in MR imaging. Journal of computer
assisted tomography. 1986;10(3):473-83.
-
van Dijk P. Direct cardiac NMR imaging
of heart wall and blood flow velocity. Journal of computer assisted tomography.
1984;8(3):429-36.
-
Ehman RL, Felmlee JP. Flow artifact
reduction in MRI: a review of the roles of gradient moment nulling and spatial
presaturation. Magn Reson Med. 1990;14(2):293-307.
-
Haacke EM, Lenz GW. Improving MR image
quality in the presence of motion by using rephasing gradients. Ajr.
1987;148(6):1251-8.
-
Schneider E, Glover G. Rapid in vivo
proton shimming. Magn Reson Med. 1991;18(2):335-47.
-
Anzai Y, Lufkin RB, Jabour BA, Hanafee
WN. Fat-suppression failure artifacts simulating pathology on
frequency-selective fat-suppression MR images of the head and neck. AJNR Am J
Neuroradiol. 1992;13(3):879-84.
-
Axel L, Kolman L, Charafeddine R,
Hwang SN, Stolpen AH. Origin of a signal intensity loss artifact in
fat-saturation MR imaging. Radiology. 2000;217(3):911-5.
-
Yoshimitsu K, Varma DG, Jackson EF.
Unsuppressed fat in the right anterior diaphragmatic region on fat-suppressed
T2-weighted fast spin-echo MR images. J Magn Reson Imaging. 1995;5(2):145-9.
-
Siegelman ES, Charafeddine R, Stolpen
AH, Axel L. Suppression of intravascular signal on fat-saturated
contrast-enhanced thoracic MR arteriograms. Radiology. 2000;217(1):115-8.
-
Babcock EE, Brateman L, Weinreb JC,
Horner SD, Nunnally RL. Edge artifacts in MR images: chemical shift effect.
Journal of computer assisted tomography. 1985;9(2):252-7.
-
Dwyer AJ, Knop RH, Hoult DI. Frequency shift artifacts in
MR imaging. Journal of computer assisted tomography. 1985;9(1):16-8.