is currently a 3rd-year Resident in Diagnostic Radiology at the
University of Utah, Salt Lake City, UT. He received his MD from
Eberhard-Karls University, Tübingen, Germany.
Breast cancer is a disease of high prevalence and mortality. It
is estimated that >211,300 women in the United States alone were
newly diagnosed with invasive breast cancer and >39,800 women
died of their disease in 2003,
making breast cancer the second leading cause, after lung cancer,
of malignancy-related death in women. Many factors influence
patient survival, a number of which cannot be changed, such as race
and age at diagnosis, while others can be affected. Because early
detection of breast cancer greatly improves the chances for
successful treatment and survival, it must be physicians' foremost
goal to diagnose breast cancer at the earliest stage possible.
These efforts must be balanced against the psychological and
financial burden that diagnostic measures could entail.
The leading imaging modality for breast cancer screening and
diagnosis is mammography. While the sensitivity of mammography is
reported to be 70% to 90%, the specificity is much lower.
Mammography is currently the primary imaging modality used to
decide which lesions require biopsy. The current recommended
positive predictive value for biopsies is 25% to 40%.
Fewer unnecessary biopsies are desirable for many reasons,
including economic factors and patient care aspects.
Magnetic resonance imaging (MRI) of the breast has proved to be
the most sensitive method of detecting implant rupture in the
augmented breast. In contrast, the use of MRI for the detection of
breast cancer is considerably less straightforward. Although it
offers the promise of a noninvasive problem-solving tool with a
mild increase in sensitivity and a substantial increase in
specificity, there is currently no consensus on MRI standards
regarding pulse sequences or interpretation criteria. Therefore,
many different imaging and interpretation approaches exist, leading
to a wide range of reported sensitivities and specificities.
While certain pulse sequences are designed to maximize spatial
resolution, others aim to detect dynamic enhancement
characteristics with high temporal resolution, at the cost of
spatial resolution. Newer approaches aim to obtain images with both
high spatial and temporal resolution. MR spectroscopy can detect
choline compounds, a known sign of malignancy.
This article will provide an overview of current and
investigational MRI approaches for detection of malignant breast
masses. This discussion presumes a basic understanding of MRI
physics. Readers desiring a review of basic imaging physics are
referred to any of several standard texts.
Because there is no universally accepted standard for MRI of the
breast, many studies employ different parameters regarding field
strength, coil configuration, pulse sequence choices (ie,
two-dimensional versus three-dimensional acquisition, temporal
versus spatial resolution, pulse sequences contain- ed in the MR
breast examination, etc.). However, no matter which protocol is
followed, high signal-to-noise ratio (SNR) and good spatial
resolution are desirable. The coil should be as close to the target
organ as possible to fulfill these demands. Different varieties of
breast coils exist. The most commonly used are double breast coils,
on which the patient rests in the prone position (Figure 1). Breast
coil design variations include open or closed design, compression
or no compression, and number of channels per breast. Multiple
channels are required to perform parallel imaging.
Most imaging protocols employ intravenous contrast. Unlike
mammography, where lesion detectibility is increased in the fatty
breast, fat may obscure an enhancing lesion in MRI, as it may
become isointense to fat after gadolinium administration. Thus,
most breast imaging protocols utilize some form of fat
Fat suppression strategies for breast magnetic resonance
Several fat-suppression methods exist. A commonly used method is
called fat saturation or chemical saturation and employs a
radiofrequency (RF) pulse to nullify fat signal.
Fat signal suppression with this technique relies on the fact that
fat and water have different precession frequencies, called Larmor
frequencies. The majority of the signal in nonfatty tissues comes
from water protons. Although the magnetic field of clinical MR
scanners is very homogenous, it is modulated by different tissues
to differing degrees. This effect is called chemical shift and
refers to the difference in magnetic field strength that the
protons of fat and nonfatty tissues experience. The resulting
difference in Larmor frequency of fat and water is 3.5 ppm
(approximately 220 Hz on a 1.5T scanner). Radiofrequency pulses can
be made more or less frequency-specific, meaning that they contain
a wider or narrower range of frequencies. A RF pulse with a wider
range of frequency will affect water and fat protons, while a RF
pulse with a narrow range of frequencies may only affect the fat
protons. The fat spins are turned into the transverse plane and are
dephased by application of a gradient, thus destroying the fat
signal (Figure 2).
A second commonly used approach is subtraction of precontrast
images from postcontrast images acquired with the same imaging
This technique requires minimal patient motion and matching of
precontrast with postcontrast images. Many other fat suppression
techniques have been utilized, described elsewhere.
High spatial resolution magnetic resonance
Architectural features of a lesion, such as shape, margins, and
density, and of the surrounding parenchyma, such as architectural
distortion, play an important role in the interpretation of
mammograms. It follows intuitively to utilize some of these
criteria with breast MRI. While some of these criteria are easily
transferred to breast MRI, such as margins; others are unique to
standard mammography, such as lesion density. Breast MRI offers
additional diagnostic criteria, such as the presence of septations,
T1 and T2 signal characteristics, pattern, and degree of
enhancement. Different interpretation models have been proposed,
taking these features into consideration
(Figures 3 and 4).
Breast MRI for evaluation of architectural features requires
high spatial resolution. The most commonly used approach is a
spoiled gradient-recalled echo (SPGR) sequence before and after
This pulse sequence can be used with or without fat-suppression
techniques. Gradient-recalled echo sequences differ from standard
spin-echo sequences in two ways: First, instead of a 90º RF pulse,
a RF-pulse with a smaller flip angle is applied. Secondly, there is
no 180º refocusing pulse. This permits imaging with very short
repetition times (TR) and therefore fast imaging with high
resolution for the following reasons:
If a RF-pulse causes a smaller flip angle (the angle that is
formed between the magnetic vector before and after the RF pulse is
applied), a part of the magnetic vector is in the transverse plane,
while a portion of longitudinal magnetization remains. The
transverse component of the magnetic vector can be measured, while
the significant residual longitudinal magnetization allows for
faster recovery of the full longitudinal magnetization (Figure 5).
Therefore TR can be shortened. In standard spin-echo sequences, the
presence of a refocusing pulse puts limitations on the minimum time
after which the signal can be acquired (time of echo [TE]). This is
due to the fact that the pulse itself takes some time, and the
resulting echo should not overlap with the refocusing pulse to
prevent zipper artifacts. Thus, the absence of a refocusing pulse
allows a much shorter TE (Figure 6).
Contrast-enhanced sequences are typically T1-weighted (some
newer approaches differ from that and are mentioned later). While
in spin-echo sequences, the T1- or T2-weighting is determined by TE
and TR alone, additional factors are important in gradient-
recalled imaging, such as the flip angle and spoiling versus
refocusing. Gradient-recalled echo sequences show T2*-properties
rather than T2-properties. T2*is shorter than T2 due to additional
dephasing by a variety of factors, including field inhomogeneities,
and chemical shift. A difference in longitudinal magnetization must
exist when the next RF pulse
is applied to obtain T1-weighted se-quences. If the flip angle is
small, the residual magnetization after the RF pulse is only
marginally smaller than fully relaxed longitudinal magnetization.
Even differences in T1-property will not lead to a significant
difference in longitudinal magnetization and thus transverse
magnetization and signal after the RF pulse has been applied
(Figure 7A). If the flip angle is increased, the longitudinal
magnetization after the RF pulse is notably smaller than before the
pulse. Thus, differences in T1-property will lead to differences in
longitudinal magnetization and therefore transverse magnetization
and signal, after the RF pulse has been applied (Figure 7B). Also,
a certain time is required for the differences in longitudinal
magnetization to develop. Therefore the TE needs to be relatively
longer than that for T2*-weighted gradient-echo sequences.
Spoiling refers to dephasing of residual transverse magnetization
after the signal has been acquired. Because the TR in
gradient-recalled sequences is relatively short, residual
magnetization is present in the transverse plane by the time the
next RF pulse has to be applied. Because T2*-properties determine
the rate of transverse magnetization decay, differences of residual
transverse magnetization between different tissues are caused by
T2*. If the RF pulse is applied, before the signal is dephased, the
new transverse magnetization will add to the residual transverse
magnetization. This can be desirable when T2*-weighted images are
acquired, but is undesirable for T1-weighted images. To avoid the
interference between T2* and T1, the residual transverse
magnetization is dephased (spoiled) either via a gradient or by
applying successive RF-pulse with a phase offset in respect to each
other. The latter is called RF-spoiling and is used in most modern
MR scanners (Figure 6).
The earlier mentioned advantages of gradient-recalled echo
sequences are achieved at the expense of much greater sensitivity
to susceptibility artifact. Also, if the echo time is chosen
inappropriately in sequences without fat suppression, the spins of
water and fat can possess opposite phase. This leads to reduced or
nulled signal of voxels containing fat and water, resulting in an
For earlier mentioned reasons, gradient-recalled sequences can
be acquired rapidly, which means that images with a large image
matrix can be obtained within short enough times, to sample
different time points after contrast enhancement.
High temporal resolution
magnetic resonance imaging
Solid tumors consist of three basic components: the cancer cells
themselves; blood vessels, accounting for approximately 1% to 10%
of tumor volume in tumors large enough to have their own blood
supply; and a collagen-rich matrix, the interstitium.
Many malignant tumors release a variety of cytokines and growth
factors, such as vascular endothelial growth factor (VEGF). These
induce growth of new vessels from already existing microvessels.
These vessels are abnormal in structure and function,
exhibiting increased vessel density and permeability. In addition,
they exhibit an increased amount of arteriovenous shunts. Although
many details have been observed regarding tumor morphologic
characteristics, the exact mechanism of altered contrast dynamics
remain unclear. It has been shown that many breast tumors show an
earlier onset of enhancement and a washout effect, which the
majority of benign breast masses do not exhibit. The commonly used
classification scheme distinguishes types 1a and 1b, type 2, and
type 3 time courses. Type 1a is continuous enhancement throughout
the entire acquisition. Type 1b indicates slowing down of
enhancement in the late postcontrast phase. A lesion classified as
type 2 shows a plateau after the initial enhancement, while a type
3 lesion shows loss of signal after the initial enhancement due to
Type 1 lesions are benign in 94% of the cases, while type 3 lesions
are malignant in 87% of the cases (Figure 8).
Type 2 lesions fall in between; however, they are more likely
malignant than benign.
In addition, the enhancement rates are also used by some for
characterization of a suspicious lesion. Again, different protocols
are used by different groups, and the reported sensitivities and
specificities to diagnose invasive cancer vary from 76% to 91% and
67% to 96%, respectively.
The majority of MRI protocols employing high temporal resolution
fall into two categories: T1-weighted SPGR with regular readout
or echoplanar imaging (EPI) readout.
The regular SPGR sequences are identical to the earlier
mentioned sequence (in the section on high spatial resolution
imaging), except for one key difference: to reduce the duration of
image acquisition, the image matrix is reduced. The number of
phase-encoding steps has a dominant effect on imaging time, while
the number of frequency-encoding steps minimally influences imaging
time. The desired resolution in phase-encoding direction equals the
number of phase-encoding steps needed in standard sequences. For
every phase-encoding step, the pulse sequence has to be repeated,
while varying the phase-encoding gradient. Each measurement
corresponds to one line in k-space. K-space is the data space that
contains spatial frequency information (Figure 9). The center line
in k-space represents zero-phase shift, thus signal acquisition in
the absence of a phase-encoding gradient. Fourier transformation
converts this data space into image space (the MR image). If
resolution in phase encoding direction is reduced by 50%, imaging
time will also be reduced by 50%, allowing for faster acquisition
and, therefore, higher temporal resolution.
Echoplanar imaging refers to a technique with which multiple
lines in k-space--multiple measurement--can be acquired rapidly
within one pulse sequence. While in regular gradient-echo and
spin-echo sequences the pulse sequence has to be repeated for every
line of k-space, while varying the phase- encoding gradient, in
EPI, the entire k-space can be acquired within one pulse sequence.
The maximum negative phase- encoding gradient is applied
(corresponding to the bottom line in k-space) and a measurement
acquired to allow aquisition of all of k-space. Immediately after
the measurement, a positive phase-encoding gradient is applied,
slightly reducing the negative phase shift. Another measurement is
acquired in the presence of a phase-encoding gradient (Figure 10).
The short application of a positive phase-encoding gradient and
measurement is repeated until the maximum positive phase shift is
reached (corresponding to the top-most line in k-space). In this
manner, a rapid acquisition of an entire slice is possible. The
earlier described technique is referred to as blipped EPI (short
blips of the phase-encoding gradient). For reasons that go beyond
the scope of this review, it can be desirable to acquire k-space
with several acquisitions, which is referred to as multishot EPI.
In multishot EPI, the individual acquisitions are interleaved. The
overall contrast of the EPI sequence depends on the root sequence
(the sequence before the echoplanar readout--such as a T2-weighted
spin-echo or a T1-weighted gradient-echo). Many other
EPI-techniques are available with different types of k-space
Combining speed and resolution
The benefits of high temporal resolution and of high spatial
resolution studies have been shown earlier. It is easy to
understand that, ideally, one would want an imaging sequence that
combines high temporal resolution with high spatial resolution to
use morphologic as well as dynamic enhancement to characterize
lesions. While the benefits of combining two different sets of
diagnostic criteria is intuitive, it has also been shown that
higher spatial resolution allows for better definition of dynamic
contrast enhancement characteristics.
The reason is that smaller lesions can be volume-averaged, with
tissues exhibiting benign enhancement characteristics. The benign
tissue can then mask the smaller focus of malignant enhancement
pattern. Different approaches are used, which so far have not been
very commonly utilized, but will likely be more frequently seen in
the future. One approach samples the center of k-space more
frequently than the periphery,
the second approach utilizes parallel imaging approaches.
K-space (or data space) differs in several respects from the
image space. While in the regular image the position along the x-
and y-axis refers to spatial location, in k-space every point
contains information about the entire image (spatial frequency
information). One of the consequences of this is that the center of
k-space contains information regarding predominantly the overall
image contrast, while the periphery of k-space contains information
about interfaces. Thus, after enhancement, the center of k-space
will contain most of the information regarding changes in signal
due to enhancement. The periphery of k-space changes little after
contrast application, because the interfaces between structures do
not change. If the periphery of k-space can be sampled less
frequently than the center of k-space, the imaging time is shorter,
because fewer lines in k-space are sampled. Saranathan et al
sampled the entire k-space with every fourth acquisition. Outer
k-space data from the full data set are then combined with each of
the previous fractional data sets to generate four full-resolution
images at high temporal resolution (Figure 11).
Song et al
report a different technique utilizing earlier described properties
of k-space. K-space can be sampled in a radial fashion (Figure 12).
The more lines that are acquired, the higher the resolution will
be. Lower resolution images can be sampled more rapidly (fewer
lines in k-space) and, therefore, more frequently. If the radial
lines between each acquisition are slightly rotated in respect to
the prior acquisition, multiple acquisitions can be combined to
form a single image with high spatial and low temporal resolution
(Figure 12). If the individual acquisitions are
Fourier-transformed, images with low spatial, but high temporal,
resolution are obtained. The individual acquisitions can thus be
combined in different ways to form images of varying temporal and
spatial resolution (Figure 13).
Another innovative approach is using parallel imaging
techniques, such as sensitivity encoding (SENSE) (Figure 14),
simultaneous acquisition of spatial harmonics (SMASH), or
variations of these techniques.
The concept of parallel imaging described here illustrates SENSE
imaging. If k-space is undersampled (for instance only every other
line in k-space is acquired), a reduced field of view and aliasing,
also called wrap-around artifact, will occur (Figure 15).
Obviously, the imaging time will be reduced if k-space is
undersampled. In parallel imaging, a multiple-element coil is used.
The signal is acquired by each element separately. Each coil
element has a certain sensitivity profile. This means that spins
closer to the coil element will create a stronger signal than the
same spins at a greater distance from this coil. Two coil elements
obtain different measurements of the same object due to their
different location and a particular sensitivity profile. Thus,
spatial information is encoded in the signal by the coil
sensitivity. If the sensitivity profile of the coil elements is
known, this information can be used to combine the measurements of
the individual coil elements and unwrap the image, resulting in an
image with the full field of view and no aliasing (Figure 16). The
parallel imaging approach can be combined with any pulse sequence.
The possible acceleration factor depends on the number of coil
elements and receiver channels. Parallel imaging can be used either
to acquire an image faster, or with higher resolution or larger
field of view within the same scan time.
Magnetic resonance spectroscopy for breast mass
Imaging of the breast for breast cancer detection aims to
fulfill two major goals: high sensitivity for detection of breast
lesions and reliable differentiation of benign from malignant
lesions (specificity). MR spectroscopy offers an adjunctive tool
for lesion characterization utilizing other characteristics
(choline content) than earlier mentioned techniques. Single-voxel
proton-spectroscopy is most widely used and, although not a tool
for screening an entire breast, it can play a role in improving
specificity (Figure 17).
In addition, proton MR spectroscopy may be a useful tool to
evaluate response to neoadjuvant chemotherapy.
Phosphorus 31 spectroscopy has also been used in evaluation of
However, the signal is lower, because there is less
P in the breast than protons, and special hardware is required. A
recent review of several MR spectroscopy studies showed an average
sensitivity and specificity of 92%.
In a subgroup of patients (¾40 years of age) the sensitivity and
specificity approached 100%. It has been shown that there are large
differences in the metabolite composition of breast tumors and
surrounding normal breast tissue.
Choline is the most widely used marker for malignancy of breast
lesions and has been shown to be a valid marker in vivo and ex
In fine-needle biopsy specimens, the sensitivity and specificity
for distinguishing benign lesions from invasive cancer with proton
MR spectroscopy was 95% and 96%, respectively.
Choline is commonly viewed as a marker of membrane turnover;
however, it has been shown that some breast cancer cell lines
exhibit a in-creased expression of choline kinase, leading to
increased choline/phosphocholine levels.
This technique has certain limitations: Choline is detectable in a
majority of breast-feeding women without the presence of a
and ductal carcinoma in situ shows only a mild elevation of choline
and is unreliably or not diagnosed with in vivo MR-spectroscopy.
The most commonly used single-voxel MR spectroscopy sequences
are point-resolved spectroscopy (PRESS)
and stimulated acquisition mode (STEAM).
Both sequences use 3 RF pulses that are made slice-selective in 3
orthogonal planes. Only the spins experiencing all 3 pulses will
have a spin-echo at the time of measurement. In this way, a target
volume is selected (Figure 18). STEAM uses three 90º pulses. The
time between the second and third pulse is called mixing time (TM).
During TM, the transverse magnetization created by the first 90º RF
pulse is oriented antiparallel to the z-axis and, therefore, decays
with T1-relaxation. Because T1 is longer than T2, less
magnetization decays (Figure 19). Despite this, STEAM has about
half the signal of PRESS, which uses one 90º RF pulse and two 180º
Different TEs (31 to 450 msec) are used, with the longer TEs
commonly being a multiple of 135 msec. TE is chosen this way to
allow the highest sensitivity for detection of lactate. Lactate has
a double peak at slightly different resonant frequencies. The
individual components of the doublet peak show an in- and
out-of-phase phenomenon in respect to each other with a frequency
of approximately 7 Hz. Thus, lactate will be in-phase every 135
msec (1Ž7 sec). This is due to j-coupling effects. Lactate is not
routinely used as a criterion in evaluation of breast lesions,
although it has been shown to be elevated in breast cancer cell
The standard MR spectroscopy sequences are set up in such a fashion
to allow detection of lactate.
Suppression of water signal is vital for proton MR spectroscopy
to detect small metabolite quantities. Water-elimination
Fourier-transform technique (WEFT) and chemical shift selective
saturation (CHESS) are the most commonly used techniques. WEFT is a
frequency-selective inversion recovery sequence. A 180º inversion
pulse affecting only the water spins is applied, aligning the water
spins antiparallel to the main magnetic field (antiparallel to the
longitudinal axis). T1-relaxation of water then takes place. When
the longitudinal magnetization of water traverses zero, the regular
90º RF pulse is applied. CHESS utilizes a water-selective 90º RF
pulse, flipping only the water spins into the transverse plane. The
water spins in the transverse plane are then dephased in the
transverse plane (Figure 20). This is often repeated with dephasing
gradients applied along different planes, before the regular 90º RF
pulse is applied. In both techniques water will not have greatly
reduced longitudinal magnetization at the time the regular 90º RF
pulse is applied and will, therefore, yield little signal.
A variety of other MR techniques have been applied to breast
cancer evaluation, but are not currently widely used and,
therefore, are not discussed here. Such techniques include MR
MR perfusion scanning,
and steady-state gradient-recalled echo with balanced gradients
Some of these techniques also require additional hardware or
certain system performance characteristics. MRI of breast cancer is
not limited to evaluation of the primary breast lesion. Different
techniques have been used for staging and evaluation of nodal
disease, including utilization of ultrasmall superparamagnetic iron
Breast MR imaging has the potential to become a standard for
breast cancer evaluation. It offers high sensitivity and
specificity. However, at this point, there is no consensus on an
imaging standard for MRI of the breast and many different
approaches are utilized: Evaluation of morphologic features,
contrast dynamics, and choline content of the lesion are most
commonly used. The combination of these techniques holds the most
promise to offer the highest sensitivity and specificity for breast
mass evaluation, supplementing the established techniques of
mammography and ultrasound.