Dynamic Imaging of the Fetal Heart Using Metric

Optimized Gating
Christopher W. Roy,1 Mike Seed,2,3 Joshua F. P. van Amerom,1,3 Bahiyah Al Nafisi,2
Lars Grosse-Wortmann,2,3 Shi-Joon Yoo,2,3 and Christopher K. Macgowan1,2,3*
Purpose: Advances in fetal cardiovascular magnetic resonance
imaging have been limited by the absence of a reliable
cardiac gating signal. The purpose of this work was to develop
and validate metric-optimized gating (MOG) for cine imaging
of the fetal heart.
Theory and Methods: Cine MR and electrocardiogram data
were acquired in healthy adult volunteers for validation of the
MOG method. Comparison of MOG and electrocardiogram
reconstructions was performed based on the image quality for
each method, and the difference between MOG and electrocardiogram
trigger times. Fetal images were also acquired,
their quality evaluated by experienced radiologists, and the
theoretical error in the MOG trigger times were calculated.
Results: Excellent agreement between electrocardiogram and
MOG reconstructions was observed. The experimental errors
in adult MOG trigger times for all five volunteers were
6 (7, 25, 17, 8, and 13) ms. Fetal images captured normal and
diseased cardiac dynamics.
Conclusion: MOG for cine imaging of the fetal myocardium
was developed and validated in adults. Using MOG, the first
gated MR images of the human fetal myocardium were
obtained. Small moving structures were visualized during radial
contraction, thus capturing normal fetal cardiac wall motion and
permitting assessment of cardiac function. Magn Reson Med
70:1598–1607, 2013. VC 2013 Wiley Periodicals, Inc.
Key words: fetal magnetic resonance imaging; image metric;
gating; post-processing
Since its inception over 30 years ago, the field of fetal
magnetic resonance imaging (MRI) has grown considerably
and now provides valuable tools for assessing the
fetal anatomy including the brain, lungs, liver, and gastrointestinal
tract (1–4). However, fetal cardiovascular
MRI (CMR) remains relatively underdeveloped. CMR is
an important tool for the noninvasive assessment of cardiac
function, anatomy, and pathology in postnatal subjects.
Advances in fetal CMR may provide useful tools
for the in utero assessment of normal fetal cardiovascular
physiology and diseases that affect the fetal cardiovascular
system. Potential applications of fetal CMR include
congenital heart disease, placental insufficiency, and
other causes of fetal distress. Unfortunately, the development
of fetal CMR has been hampered by the difficulty
in measuring a fetal electrocardiogram (ECG) signal for
cardiac gating (5). The fetal ECG signal is typically an
order of magnitude weaker than that of the mother, and
can be difficult to detect in the presence of electrical
noise from the MRI (6). Current alternative strategies for
fetal gating include ultrasound based techniques and
self-gating. These strategies are described below followed
by the technique presented in this work: metric-optimized
gating (MOG).
Nonimaging ultrasound (fetal cardiotocography) is an
established technique for monitoring the fetal heart rate.
MR-compatible units show promise for fetal cardiac gating,
although they can be limited by interference between
the unit and MR system (7). Furthermore, the likelihood
of fetal movement during the scan may challenge the reliability
of the gating signal. This method has been used to
image the heart of anaesthetized fetal sheep but no direct
human application has been published to date (8).
Another approach to fetal cardiac gating is MRI “selfgating,”
where a periodic gating signal is extracted from
the MR data itself and is used to sort the data retrospectively
(9). Modern implementations exist for both radial
and Cartesian trajectories, and use a variety of signal
detection algorithms (10–12). Self-gating is attractive as
it requires no additional hardware, although the reliability
of the gating signal depends on slice prescription and
temporal resolution, and therefore may not accommodate
all k-space sampling trajectories. Previous publications
have used self-gating for cardiac imaging in the fetal
mouse, chick, and lamb (13–15). These techniques may
be applied to human fetal CMR, but as with ultrasound
gating, no direct application has been published to date.
Recent advances in non-gated acquisitions using real
time imaging may be applicable to fetal CMR; however,
these techniques require a combination of aggressive parallel
imaging with non-Cartesian k-space sampling, constraining
the available spatial resolution, temporal
resolution, and signal-to-noise ratio. Even in late gestation,
the diameter of a normal human left ventricle is only
15 mm and the diameters of major vessels are less than
8 mm (16). These small dimensions coupled with high
1Departments of Medical Biophysics and Medical Imaging, University of Toronto,
Toronto, Canada.
2Division of Cardiology, Department of Paediatrics, Labatt Family Heart
Centre, The Hospital for Sick Children, University of Toronto, Toronto,
Canada.
3Department of Diagnostic Imaging, The Hospital for Sick Children, University

of Toronto, Toronto, Canada.

Additional Supporting Information may be found in the online version of
this article.
Grant sponsor: Ontario Opportunity Trust Fund—Hospital for Sick Children
Foundation Student Scholarship Program and Alexander Graham Bell Canada
Graduate Scholarship (Natural Sciences and Engineering Research
Council of Canada).
*Correpondence to: Christopher Macgowan, Ph.D., Diagnostic Imaging,
555 University Ave., Toronto, ON, Canada, M5G 1X8. E-mail: christopher.
[email protected]
Received 15 July 2012; revised 28 November 2012; accepted 3 December
2012
DOI 10.1002/mrm.24614
Published online 4 February 2013 in Wiley Online Library (wileyonlinelibrary.
com).
Magnetic Resonance in Medicine 70:1598–1607 (2013)
VC 2013 Wiley Periodicals, Inc. 1598
heart rates (between 110 and 180 bpm) necessitate high
spatial and temporal resolutions to capture cardiac dynamics
(17).
In this work, we demonstrate a gating strategy for fetal
cardiac imaging called MOG. MOG is a post-processing
method which provides retrospective cardiac gating. The
technique was previously applied to phase-contrast MRI
measurements of fetal blood flow by Jansz et al. (18).
This method exploits the fact that nongated cardiac MR
images contain characteristic artifact that can be quantified
using an image metric. Here, we develop the MOG
technique for MR imaging of the human fetal myocardium
and quantify its accuracy. The approach is validated
experimentally in adult volunteers and used to
obtain the first gated high-resolution cine images of the
fetal human myocardium in both healthy subjects and
those with congenital heart disease.
THEORY
Cardiac Gating
To improve the spatial and temporal resolution of conventional
CMR, k-space data are typically acquired over
multiple heartbeats and combined into a set of cine
images through the use of cardiac gating. In gated scans,
an ECG (or plethysmograph) waveform triggers the continuous
acquisition of data. For each segment of k-space,
the time point in the cardiac cycle (cardiac phase) is
determined according to the acquisition time with
respect to the last ECG trigger. Data are then interpolated
to a uniformly distributed set of cardiac phases and Fourier
transformed to create a set of images spanning a representative
heart beat (19).
Synthetic Gating
In the absence of an ECG signal, cine images can be
acquired using a synthetic waveform to periodically trigger
the data acquisition. Using this approach, “pseudogating”
can be performed by estimating the mean heart rate of the
subject (20). If the heart rate is approximately constant and
the estimate is accurate, then every segment of k-space is
sampled at each phase of the cardiac cycle. These data can
then be reconstructed in the same manner as retrospectively
gated data with the heart rate estimate replacing the
ECG waveform. However, if the true heart rate varies
throughout the scan, the acquired k-space segments will be
incorrectly combined based on an inaccurate calculation of
the cardiac phase for each segment. The resulting
“misgated” images will contain artifact. Alternatively, one
can purposefully oversample each cardiac cycle, by setting
the period of the gating waveform to be longer than the cardiac
cycle (12). This guarantees that every row of k-space is
sampled at each phase of the cardiac cycle but the cardiac
phase associated with each segment remains unknown
and, thus, the resulting reconstructions will be misgated.
MOG
MOG was developed to retrospectively combine synthetically
gated, oversampled data (18). MOG employs an iterative
process as depicted in Figure 1. Upon collecting
synthetically gated, oversampled data, a hypothetical gating
waveform is created that spans the period of image acquisition.
Images are reconstructed using this
hypothetical triggering information, following the same
steps described above for cardiac gating. Artifact from
misgating is quantified using a metric for image quality,
calculated over a region-of-interest (ROI) in the image, and
the relative timings of the hypothetical triggers are
adjusted. The process is repeated iteratively until the
image metric is optimized.
The iterative process employed by the MOG technique
presents two challenges. First, an image metric is
FIG. 1. Schematic of the MOG method. a: K-space data are
acquired, triggered by a synthetic waveform. b: The data are retrospectively
interpolated and combined according to a hypothetical
waveform. Image reconstruction (two-dimensional inverse
Fourier Transform) is performed and an image metric (entropy)
quantifies artifact over a ROI in the reconstructed images. The hypothetical
waveform is adjusted and k-space interpolation, image
reconstruction, and image metric calculation is repeated. c: When
the entropy is minimized, the final set of cine images is output.
Dynamic Imaging of the Fetal Heart Using MOG 1599
required that is sensitive to the effects of misgated
images of the fetal heart. Second, a minimization algorithm
is required to optimize the hypothetical triggers in
a practical processing time.
Previous work in metric-based autocorrection has
identified several metrics that can quantify image quality
(21). Amongst them, the image entropy presents a functional
choice for identifying artifact in misgated cardiac
images. In this work we calculate the image entropy
using the following expression:
E¼
X
t
XBt
t
Bt
X
x
X
y
Sx;y;t
Bt
log
Sx;y;t
Bt
__
; [1]
where S is the pixel intensity with x, y indexing over
space, t indexing over time, and B is a normalization factor
adapted from Atkinson et al. (22):
Bt ¼
X
x
X
y
S2
x;y;t
h i1=2
: [2]
In misgated cardiac images the sources of artifact are
blurring and ghosting. Figure 2 illustrates the effective
decrease in temporal resolution that appears when comparing
gated and misgated image reconstructions. There
is also a decrease in contrast between the bright blood
pool and otherwise dark surrounding tissue. Image entropy
is sensitive to ghosting and blurring in areas of an
image that should otherwise be dark and, hence, is a
suitable metric for this application (22).
The hypothetical gating waveform used for MOG
reconstruction can be thought of as a multi-parameter
model with each parameter representing the timing of
one trigger. Thus, hereafter the relative timing of MOG
triggers are referred to as parameters. The simplest
method for optimizing multiple parameters is through a
brute force search that calculates the cost function
(entropy) for every combination of possible parameter
values. This method is computationally intensive but
has the advantage of being exhaustive, thereby increasing
the confidence in the minimum. Conversely, more efficient
optimization algorithms evaluate fewer parameter
combinations but require the cost function to vary
smoothly and to be well behaved to avoid solutions at
local minima (23). We demonstrate later that these conditions
are experimentally satisfied for the MOG method.
METHODS
Adult Validation
Validation of the MOG method for cine imaging of the
heart was performed on five healthy adult volunteers.
Measurements were acquired with informed written consent
and approval from the hospital research ethic board.
Short axis images of the myocardium were acquired on a
3.0T MRI research scanner (TRIO, Siemens Healthcare—
Germany) using a multichannel cardiac coil. Three strategies
were performed: Cartesian sampling without parallel
imaging, Cartesian sampling with GRAPPA, and
radial sampling without parallel imaging (24). A steady
state free precession sequence was used for all image
acquisitions (scan parameters listed in Table 1). The
acquired spatial and temporal resolutions were chosen to
provide coverage of the heart similar to that of the fetal
acquisitions described below. For each scan, data were
acquired under breath-hold and the ECG waveform was
recorded for subsequent qualitative and quantitative
comparison between ECG and MOG reconstructions.
MOG performance was evaluated by reconstructing the
same data using either ECG trigger times or MOG parameters.
The image quality of MOG reconstructions was
compared qualitatively against ECG reconstructions.
Quantitative evaluation of MOG reconstructions was performed
through an error analysis of the MOG parameters.
An experimental error for MOG reconstructions was calculated
from the standard-deviation of the difference
between ECG trigger times and MOG parameters. Furthermore,
the theoretical errors in the MOG parameters
were calculated as follows: image noise was propagated
FIG. 2. The effect of misgating. a: Single frame from a retrospectively reconstructed set of adult images acquired with ECG
gating and Cartesian
sampling. b: M-mode display of intensities along the vertical dash line in (a) showing the temporal evolution of the signal. (c
and d)
Data reconstructed with an incorrect estimate of the heart rate (“misgated”), showing a loss of both spatial and temporal
resolution.
1600 Roy et al.
through Eq. [1] to provide an uncertainty in the image
entropy, which was then transformed into an uncertainty
in each MOG parameter. This transformation was based
on the curvature of the entropy landscape about its optimum,
with respect to each parameter. Details of the theoretical
error calculation are provided in Appendix.
Fetal Evaluation
Short axis images of the heart were acquired in a normal
fetal subject (35 weeks gestational age) and another with
a congenital heart defect (37 weeks gestational age). All
fetal scans were performed on a 1.5 T MRI clinical scanner
(Avanto, Siemens Healthcare—Germany) using a
multichannel abdominal coil and a steady state free precession
sequence (scan parameters listed in Table 1).
The mother of the healthy fetal subject was recruited as
part of a larger on-going study at our institution while
the congenital heart defect patient was directed to MR to
compliment fetal echocardiography data. In both cases,
scans were performed with informed written consent
and research ethic board approval.
Healthy fetal data were acquired under maternal breathhold
in a single slice acquisition with no parallel imaging.
These data were acquired to evaluate high spatial and temporal
resolution fetal cardiac imaging with the MOG
method. Accordingly, the scan length (14 s) does not necessarily
reflect a practical clinical protocol where considerations
given to fetal movement keep scan lengths as short as
possible. For the images acquired in a fetal subject with a
congenital heart defect, the scan length was shortened
through the use of parallel imaging and the images were
acquired with a free-breathing technique to allow adequate
time for a multislice acquisition (_5 s per slice). To minimize
scan time and avoid fetal motion artifacts, the spatial
and temporal resolution of the acquisition were reduced.
As with the adult analysis described above, fetal data
were evaluated both qualitatively and quantitatively. In
the absence of an ECG gating signal for comparison,
qualitative assessment of image quality was performed
by an expert radiologist and quantitative evaluation of
MOG performance was assessed by the theoretical error
in the MOG parameters.
MOG Algorithm
All image reconstructions were performed offline on a
personal computer using MATLAB (The Mathworks
Inc.). The duration of each adult scan spanned 7–15
heartbeats and fetal scans spanned 15–30 heartbeats,
resulting in hypothetical waveforms consisting of 7–15
and 15–30 parameters for adult and fetal reconstructions
respectively. The initial parameter values for the MOG
search algorithm were based on a one-dimensional (i.e.,
constant heart rate) brute force search to determine the
mean heart rate. Optimization of individual R-wave triggers
was then performed using the Nelder–Mead simplex
algorithm, adapted for entropy minimization (25).
Entropy was calculated over user-defined ROIs, each
prescribed using an initial temporally averaged image
reconstructed from the oversampled data. Each ROI was
chosen to cover the cardiac anatomy. For Cartesian sampling,
the ROI extended approximately 20% beyond the
heart in the phase encoding direction. This was done to
capture ghosting artifact resulting from misgating,
thereby improving the sensitivity of the MOG algorithm.
Conversely, misgating artifact in radial sampling is isotropic,
so ROIs were extended radially by approximately
20% beyond the heart for these data.
RESULTS
Adult Validation
Figure 3 displays single slice short-axis images obtained
in a representative healthy volunteer using three sampling
strategies. For each sampling strategy, a comparison
between ECG and MOG image reconstructions is
shown for two cardiac phases (Fig. 3a,c,e). Line plots of
the left-ventricle signal intensity versus time (M-mode
display) are also provided to compare the temporal dynamics
of ECG and MOG reconstruction (Fig. 3b,d,f).
Qualitatively, these images show excellent agreement
between ECG and MOG reconstructions across all sampling
strategies.
The agreement between ECG and MOG reconstructions
was studied quantitatively by calculating the experimental
error in the MOG trigger times relative to ECG. The
Table 1
Scan Parameters for In Vivo Data Acquisition
Adult Fetal
Field strength (T) 3 3 3 1.5 1.5
Trajectory Cartesian Cartesian Radial Cartesian Cartesian
Parallel imaging Off GRAPPA Off Off GRAPPA
Acceleration factor N/A 2 N/A N/A 2
Reference lines N/A 24 N/A N/A 24
Flip angle (degrees) 45 45 45 56 56
TE/TR (ms/ms) 1.52/3.04 1.52/3.04 1.54/3.04 1.33/3.14 1.33/3.14
Views per segment 15 15 15 8 16
Temporal resolution (ms) 46 46 46 25 50
FOV (mm2) 400 _ 396 400 _ 396 400 _ 400 300 _ 220 273 _ 340
Matrix size 195 _ 192 195 _ 192 180 _ 192 240 _ 176 180 _ 224
Spatial resolution (mm2) 2.1 _ 2.1 2.1 _ 2.1 2.1 _ 2.1 1.3 _ 1.3 1.5 _ 1.5
Number of slices 1 1 1 1 10
Slice Thickness (mm) 8 8 8 4 4
Scan length (s) 11 8 11 14 53
Dynamic Imaging of the Fetal Heart Using MOG 1601
experimental errors across all three sampling strategies
for each adult volunteer were: 67, 625, 617, 68, and
613 ms, demonstrating that the MOG algorithm provided
precise trigger times for reconstruction.
Figure 4 shows the theoretical error in each parameter
for a representative adult Cartesian reconstruction and a
radial MOG reconstruction. For brevity, only the theoretical
errors for Cartesian and radial data without parallel
imaging are shown; however, the theoretical errors for
Cartesian data with parallel imaging exhibited a similar
trend. Differences in the range of theoretical errors for
Cartesian (62_16 ms) and radial (68_12 ms) data can
be explained by the curvature of the entropy landscapes
used to calculate the theoretical errors. Figure 5 shows
representative one-dimensional entropy landscapes, relative
to their optimal values, corresponding to the same
data used to produce Figure 4. For the Cartesian data,
landscape curvature varied across the acquisition,
reflecting differences in the k-space data associated with
each parameter. Because a linear-filling acquisition was
used, central rows of k-space were collected near the
midpoint of the scan and changes in the corresponding
parameters resulted in greater curvature of the entropy
landscape and consequently a smaller theoretical error.
Conversely, landscape curvature was consistent between
parameters for the radial acquisition and hence the theoretical
errors shown in Figure 4 demonstrate less variation
between parameter values when compared to
Cartesian data. Despite the differences in relative curvature,
all entropy landscapes were smooth with welldefined
minima, supporting the use of the Nelder–Mead
Simplex optimization method.
Fetal Evaluation
Figure 6 displays short axis images of a healthy human
fetal heart during end-systole and end-diastole. The
cross-sections of the antero-lateral and postero-medial
papillary muscles are visible in the bright blood pool of
the left ventricle (arrow in Fig. 6b). A normal thin rim of
pericardial fluid appears as a thin bright band surrounding
the two ventricles in the end-systolic images. The Mmode
display shown in Figure 6c is comparable to the
adult data displayed in Figure 3. Radial contraction and
relaxation of the fetal ventricles is well represented and
thus, ventricular function can be assessed from these
images. Finally, Figure 6d shows an M-mode display of
the data before the MOG method is applied. Without the
MOG method, the ungated data produce blurry images
with poorly defined ventricles and thus, prevent assessment
of ventricular function. Figure 7 shows short axis
images of the fetal myocardium for a subject with a congenital
heart defect. Disproportionate left (arrow in Fig.
7b) and right ventricles indicate a degree of left ventricular
hypoplasia. Also, comparison of the end-systolic and
end-diastolic images combined with the M-mode display
in Figure 7c shows reduced contractility of the left ventricle
when compared to that of the right ventricle. Aortic
and mitral valve stenosis resulting in reduced filling
and pressure overload of the left ventricle explain this
ventricular dysfunction, which was confirmed postnatally
in agreement with these MRI findings.
Although the lack of an external gating signal prevented
direct measurement of experimental error in the
MOG trigger times for fetal data, the theoretical errors in
the optimized fetal parameters were calculated. The theoretical
error in the fetal MOG parameters ranged from
63 ms to 634 ms. As with the adult Cartesian data, the
theoretical error in parameters that affected the central
rows of K-space were lower than those affecting the
periphery.
DISCUSSION
In this work, MOG was developed and validated to retrospectively
reconstruct oversampled MRI data for fetal
cardiac imaging. Using this technique, the first gated
FIG. 3. Comparison of MOG and ECG adult volunteer reconstructions.
End-systolic and diastolic images are shown for each sampling
strategy: (a and b) Cartesian without parallel imaging, (c and
d) Cartesian sampling with GRAPPA, and (e and f) Radial sampling
without parallel imaging. In all cases the temporal evolution
of signal along the white dashed line is plotted.
1602 Roy et al.
cine MR images of the human fetal myocardium were
presented, demonstrating a new MR tool for assessing
congenital heart disease and investigating compromised
fetal circulation in other conditions such as placental
insufficiency. Validation was performed in adults using
short axis images of the heart acquired using three sampling
strategies. Qualitative and quantitative analyses
showed excellent agreement between MOG and ECG
reconstructions of the same data.
The comparable quality of MOG and ECG image reconstructions
in adults, combined with the precision of the
optimized MOG trigger times (experimental errors: 67,
625, 617, 68, and 613 ms), provided a strong validation
of the MOG technique for imaging the fetal myocardium.
Furthermore, the mean theoretical errors in the
adult MOG parameters were 8, 14, 13, 9, and 8 ms for
each volunteer respectively. The theoretical error was dependent
on our ability to capture cardiac wall motion
FIG. 4. Theoretical error as a function of MOG parameter for adult data acquired using (a) Cartesian and (b) radial sampling.
FIG. 5. Entropy as a function of timing offset. a–c: Demonstration of the change in entropy that occurs when Cartesian k-
space data
are interpolated and reconstructed based on R-wave triggers that are shifted from their true ECG values. d–f: shows the same
process
but for radial data.
Dynamic Imaging of the Fetal Heart Using MOG 1603
(temporal resolution), inherent image quality (SNR,
spatial resolution), and metric sensitivity (Appendix).
The temporal resolution of the adult MR acquisition
(46 ms) exceeded that needed for accurate interpolation
of cardiac wall motion (i.e., the majority of cardiac
motion is bounded by 620 Hz) and thus, the theoretical
error in our adult data was mainly a function of SNR
and metric sensitivity (26–28).
To our knowledge, no study has been performed to
determine the temporal bandwidth of human fetal cardiac
motion and there is not yet a standard fetal CMR protocol.
Therefore, a temporal resolution of <40 ms was used to
capture fetal end-systole (29). Still, it is possible that our
current temporal resolution may not fully capture fetal cardiac
wall motion which, when combined with the lower
SNR of the fetal scan, may contribute to greater theoretical
errors in fetal MOG parameters. However, the mean and
range of fetal theoretical errors (16 ms and 63_34 ms)
were comparable to those in adults (10 ms and 2–29 ms).
Therefore the image quality of both fetal volunteer scans
allowed for clinical assessment of cardiac function.
Despite the success of the presented MOG algorithm,
potential improvements have been identified. The ROI
used for the entropy calculation was user-defined, based
on an initial temporally averaged reconstruction of the
oversampled data. Potential segmentation solutions exist
for removing this user-dependence; however the optimal
ROI for MOG was not investigated. As previously mentioned
in the Methods section, the present approach
attempts to maximize metric sensitivity by capturing
dynamic fetal cardiac information within the ROI while
avoiding possible artifact from maternal cardiac or respiratory
motion. Avoiding maternal artifact may be more
challenging in radial scans because artifact is isotropic.
Another potential improvement involves the image
metric. Image entropy was used in this application and
demonstrated good sensitivity to artifact associated with
misgating. Previous work in metric-based autocorrection
has focused on correcting artifact from patient motion in
images of the brain, shoulder, and angiography by iteratively
adjusting k-space (22,30,31). Although a number of
different image metrics have been proposed in the literature,
the image entropy and the entropy of the gradient of
the image are prevalently identified as appropriate metrics
for quantifying motion artifact (21). McGee et al.(32)
describe the entropy of the gradient as an appropriate
metric for images wherein high resolution structures are
distorted within the image. Conversely, the image
FIG. 6. MOG reconstructions of a healthy human fetal heart. a: A short axis view at the mid ventricular level is shown at end-
systole
and end-diastole. b: Magnified end diastolic image from (a) with the left ventricle denoted by a white arrow. c: Plot of signal
intensity
versus time for the dashed line in (b). d: Same as (c) but before application of MOG, showing obscured radial contraction due
to artifact
from cardiac motion.
FIG. 7. MOG reconstructions of human fetal subject with congenital heart disease. a: A short axis view at the mid ventricular
level is
shown at end-systole and end-diastole. b: Magnified end diastolic image from (a) with the left ventricle denoted by a white
arrow. c:
Plot of signal intensity versus time for the dashed line in (b).
1604 Roy et al.
entropy is appropriate for images with ghosting outside
the moving object. Furthermore, Atkinson et al. (22) state
that entropy minimization removes blurring from structures
that would otherwise be dark. Given that the fetal
myocardium and surrounding anatomy is naturally dark
relative to the bright blood pool in steady state free precession
images, entropy was a sensitive metric for artifact
due to misgating. Other metrics may hold potential for
this application, but given the demonstrated precision of
the MOG parameters it is likely that improvements to the
image acquisition (e.g., higher SNR and shorter scan duration)
would provide greater benefit.
Finally, the MOG algorithm required full image reconstruction
at each iteration so the entropy could be calculated
over an ROI. In this work, the Nelder–Mead
Simplex method was used and convergence was typically
reached in 1000 iterations. For Cartesian sampled
data without parallel imaging, MOG processing lasted 10
min per slice (MATLAB, standard desktop computer).
For image reconstructions with a higher computational
demand (parallel imaging, radial data), total MOG processing
time lasted approximately 2 h per slice. The
Nelder–Mead Simplex method produced excellent
results, but further investigation into alternative optimization
algorithms, computing hardware, and improved
code efficiency may result in faster post-processing
times, justifying the use of MOG in a clinical setting.
Ultrasound is the current gold standard for assessment
of the fetal heart. It is non-invasive, relatively cheap, accessible,
and capable of real-time imaging at high spatial
resolution (33). Still, there is increasing interest in the
use of MRI for studying fetal hemodynamics (34). PCMRI
is the gold standard for the noninvasive measurement
of blood flow in adults and our previous work has
shown that MOG can enable PC-MRI measurements in
utero (18,35). Through the combined MOG applications
of PC-MRI and cine imaging of the myocardium, a more
comprehensive assessment of fetal cardiac function may
be available, making in utero MRI a valuable compliment
to ultrasound for assessment of the fetal heart.
Current applications of postnatal cardiac imaging with
MRI include delineation of cardiac anatomy and evaluation
of cardiac function through ventricular volumetry (36–39).
There remain significant challenges in adapting these applications
to in utero image of the fetal heart. Constraints on
scan length, due to the possibility of fetal motion combined
with the size of fetal cardiac structures and the speeds at
which they move, necessitate new approaches, as the temporal
and spatial resolutions provided by current fetal imaging
protocols restrict high quality cardiac imaging to late
gestation or the use of sedation to prevent gross fetal
motion. Nevertheless, the development of a gating technique
for fetal CMR is an important step towards high quality
imaging that can not only capture the dynamics of the
fetal cardiac cycle, but provide new methods for studying
diseases that disrupt the fetal circulation.
CONCLUSION
A novel gating approach CMR of the fetal heart was presented
and validated through adult volunteer data. With
this method, the first gated high resolution dynamic
images of the human fetal heart were acquired using conventional
cine techniques.
APPENDIX
Theoretical Error Calculation
This appendix presents a method for determining the theoretical
error in MOG parameters. For a given parameter,
the uncertainty in the image entropy is determined by
propagating the image noise through Eq. [1]. The uncertainty
in the image entropy is then transformed to provide
an estimate of the theoretical error in each MOG parameter,
based on the curvature of the entropy landscape about
the optimum. To simplify several steps in the error propagation
calculation below, Eq. [1] is written as follows:
E¼
X
t
X
i
si;t
c
_ ln
si;t
Bt
__
: [A1]
Here, the spatial indices x and y have been combined
into one spatial coordinate i. Si,t is the intensity of the
magnitude image at spatial location i and time t and a
third variable C is introduced:
C¼
X
t
Bt :
Now, we define a function of three variables
ei;t ¼ eðSi;t ; Bt ; CÞ:
eðSi;t ; Bt ; CÞ ¼
Si;t
C
_ ln
Si;t
Bt
__
:
Equation [A1] is then written as:
E¼
X
t
X
i
ei;t :
Hereafter, r is used to denote the uncertainty of each
expression. The uncertainty in E, r (E), then becomes:
sðEÞ ¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X
t
X
i
_
sðei;tÞ
_2
r
: [A2]
The uncertainty in ei,t is given by:
sðei;tÞ ¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffif
fiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
@ei;t
@si;t
_ sðSi;tÞ
_ _2
þ
@ei;t
@Bt
_ sðBtÞ
_ _2 @ei;t
@C
_ sðCÞ
_ _2
_
s
[A3]
The error in Si,t is the image noise in space and time.
The uncertainties in Bt and C are given by the following:
sðBi;tÞ ¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X
i
_
Si;t _ _sðSi;tÞ
_
X
i
si;t
2
2
vuuut
:
sðCi;tÞ ¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X
t
X
i
_
Si;t _ _sðSi;tÞ
_
X
i
si;t
2
20
B@
1
C A:
vuuuut
Dynamic Imaging of the Fetal Heart Using MOG 1605
The partial derivatives of e
i,t

are:
@ei;t
@Si;t
¼
1
C
_ ln
Si:t
Bt
__
þ1
__
;
@ei;t
@Bt
¼_
Si;t
Bt _ C
;
@ei;t
@C
¼_
Si:t
C2 _ ln
Si:t
Bt
__
:
Equation [A1] then becomes:
sðei;tÞ ¼
1
C
_ ln
Si;t
Bt
__
þ1
__
_ sðSi;tÞ
_ _2
"
þ
Si;t
Bt _ C
_ _2
_
X
i
_
Si;t _ sðSi;tÞ
_2
X
i
Si;t
2
þ
Si;t
C2 ln
Si;t
Bt
_ _ __2
_
X
t
X
i
_
Si;t _ sðSi;tÞ
_2
X
i
Si;t
2
3
75
12

and the full expression for the uncertainty in the

image entropy is:
sðEÞ ¼
X
t
X
i
1
C
_ ln
Si;t
Bt
__
þ1
__
_ sðSi;tÞ
_ _2
0
B@
2
64
þ
Si;t
Bt _ C
_ _2
_
P
i
_
Si;t _ sðSi;tÞ
_2
P
iSi;t
2
þ
Si;t
C2 _ ln
Si;t
Bt
_ _ __2
_
X
t
P
i
_
Si;t _ sðSi;tÞ
_2
P
iSi;t
2
1
CA
3
75
12

: [A4]
To explain the dependency of Eq. [A4] on relevant
imaging parameters, several simplifying assumptions can
be made. First, assuming the image noise is approximately
constant over an ROI [i.e., sðSi;tÞ _ sðSÞ] for all
spatial and time indices, sðEÞ simplifies to give:
sðEÞ _
sðSÞ
C
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffif
fiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiff
iffi
X
t
X
i
ln
Si;t
Bt
__
þ1
_ _2
þ
Si;t
Bt
_ _2
þ
Si;t _ ln Si;t
Bt
__
C
0
@
1
A
2
_T
0
B@
1
CA
vuuuut
;
where T is the total number of time indices. Furthermore,
the second and third terms of the above expression
are experimentally small relative to the first, and can be
removed:
sðEÞ _
sðSÞ
C
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X
t
X
i
ln
Si;t
Bt
__
þ1
_ _2
vuut
:
Finally, ln Si;t
Bt
__
<< 0 and so,
sðEÞ _
sðSÞ
C
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X
t
X
i
ln
Si;t
Bt
_ _ __2
vuut
:
The error in the image entropy is then inversely proportional
to the signal-to-noise ratio, scaled by a second
term. The scaling term is minimized when an image consists
of a single bright pixel against a dark background. If
the image is blurred, the scaling term grows, amplifying
the uncertainty in the image entropy. Thus, the uncertainty
in the image entropy is governed by the image
noise but for images that are inherently blurry (poor resolution,
misgating, etc.) the contribution of the image
noise to the uncertainty in the image entropy increases.
The uncertainty in the image entropy is transformed into
an uncertainty in each MOG parameter according to the
following: first, MOG parameters are perturbed about their
optimum values creating one-dimensional entropy landscapes
(Fig. 8). Second, the entropy landscapes are fit with
a parabola. Finally, the uncertainty in each parameter is
given by the change in parameter value required to raise
the entropy an amount r(E) above its optimal value. The
theoretical error is then dependent on the uncertainty in
the image entropy and the curvature of the entropy landscape
which, in turn, is dependent on the underlying
motion and sensitivity of the image entropy to that motion.
REFERENCES
1. Kim D-H, Chung S, Vigneron DB, Barkovich AJ, Glenn OA. Diffusionweighted
imaging of the fetal brain in vivo. Magn Reson Med
2008;59:216–220.
FIG. 8. Entropy as a function of parameter offset. The theoretical
error is given by the timing offset required to raise the entropy an
amount r(E) above its minimum value.
1606 Roy et al.
2. Kim D-hyun, Vahidi K, Caughey AB, Coakley FV, Vigneron DB,
Kurhanewicz J, Mow B, Joe BN. In vivo (1) H magnetic resonance
spectroscopy of amniotic fluid and fetal lung at 1.5 T: technical challenges.
JMRI 2008;28:1033–1038.
3. S_rensen A, Holm D, Pedersen M, Tietze A, Stausb_l-Gr_n B, Duus
L, Uldbjerg N. Left-right difference in fetal liver oxygenation during
hypoxia estimated by BOLD MRI in a fetal sheep model. Ultrasound
Obstet Gynecol 2011;38:665–672.
4. Cannie M, Jani J, Dymarkowski S, Deprest J. Fetal magnetic resonance
imaging: luxury or necessity? Ultrasound Obstet Gynecol
2006;27:471–476.
5. Brugger PC. In: Prayer D, editor. Fetal MRI. Berlin, Heidelberg:
Springer Berlin Heidelberg; 2011 pp. 247–257.
6. Peters M, Crowe J, Pi_eri JF, Quartero H, Hayes-Gill B, James D, Stinstra
J, Shakespeare S. Monitoring the fetal heart non-invasively: a
review of methods. J Perinat Med 2001;29:408–416.
7. Feinberg Da, Giese D, Bongers DA, Ramanna S, Zaitsev M, Markl M,
G€unther M. Hybrid ultrasound MRI for improved cardiac imaging
and real-time respiration control. Magn Reson Med 2010;63:290–296.
8. Yamamura J, Kopp I, Frisch M, Fischer R, Valett K, Hecher K, Adam
G, Wedeg€artner U. Cardiac MRI of the fetal heart using a novel triggering
method: initial results in an animal model. JMRI
2012;35:1071–1076.
9. Spraggins TA. Wireless retrospective gating: application to cine cardiac
imaging. Magn Reson Imaging 1990;8:675–681.
10. Crowe ME, Larson AC, Zhang Q, Carr J, White RD, Li D, Simonetti
OP. Automated rectilinear self-gated cardiac cine imaging. Magn
Reson Med 2004;52:782–788.
11. Thompson RB, McVeigh ER. Flow-gated phase-contrast MRI using radial
acquisitions. Magn Reson Med 2004;52:598–604.
12. Larson AC, White RD, Laub G, McVeigh ER, Li D, Simonetti OP. Selfgated
cardiac cine MRI. Magn Reson Med 2004;51:93–102.
13. Hiba B, Richard N, Thibault H, Janier M. Cardiac and respiratory
self-gated cine MRI in the mouse: comparison between radial and
rectilinear techniques at 7T. Magn Reson Med 2007;58:745–753.
14. Holmes WM, McCabe C, Mullin JM, Condon B, Bain MM. Images in
cardiovascular medicine. Noninvasive self-gated magnetic resonance
cardiac imaging of developing chick embryos in ovo. Circulation
2008;117:e346–e347.
15. Yamamura J, Frisch M, Ecker H, Graessner J, Hecher K, Adam G,
Wedeg€artner U. Self-gating MR imaging of the fetal heart: comparison
with real cardiac triggering. Eur Radiol 2011;21:142–149.
16. DeVore GR. Assessing fetal cardiac ventricular function. Semin Fetal
Neonatal Med 2005;10:515–541.
17. Sandmire HF, DeMott RK. Electronic fetal heart rate monitoring:
research guidelines for interpretation. Am J Obstet Gynecol
1998;179:276–277.
18. Jansz MS, Seed M, van Amerom JFP, Wong D, Grosse-Wortmann L,
Yoo S-J, Macgowan CK. Metric optimized gating for fetal cardiac
MRI. Magn Reson Med 2010;64:1304–1314.
19. Lenz GW, Haacke EM, White RD. Retrospective cardiac gating: a
review of technical aspects and future directions. Magn Reson Imaging
1989;7:445–455.
20. Haacke EM, Lenz GW, Nelson AD. Pseudo-gating: elimination of periodic
motion artifacts in magnetic resonance imaging without gating.
Magn Reson Med 1987;4:162–174.
21. McGee KP, Manduca A, Felmlee JP, Riederer SJ, Ehman RL. Image
metric-based correction (autocorrection) of motion effects: analysis of
image metrics. JMRI 2000;11:174–181.
22. Atkinson D, Hill DL, Stoyle PN, Summers PE, Keevil SF. Automatic
correction of motion artifacts in magnetic resonance images using an
entropy focus criterion. IEEE Trans Med Imaging 1997;16:903–910.
23. Bevington P, Robinson DK. Data reduction and error analysis for the
physical sciences, 2nd ed. New York: McGraw-Hill; 1994.
24. Griswold MA, Jakob PM, Heidemann RM, Nittka M, Jellus V, Wang J,
Kiefer B, Haase A. Generalized autocalibrating partially parallel
acquisitions (GRAPPA). Magn Reson Med 2002;47:1202–1210.
25. Lagarias J, Reeds J, Wright M. Convergence properties of the Nelder-
Mead simplexmethod in low dimensions. SiamJ Optim 1998;9:112–147.
26. Foo TK, Bernstein MA, Aisen AM, Hernandez RJ, Collick BD, Bernstein
T. Improved ejection fraction and flow velocity estimates with
use of view sharing and uniform repetition time excitation with fast
cardiac techniques. Radiology 1995;195:471–478.
27. Holdsworth DW, Norley CJ, Frayne R, Steinman DA, Rutt BK. Characterization
of common carotid artery blood-flow waveforms in normal
human subjects. Physiol Meas 1999;20:219–240.
28. Kramer CM, Barkhausen J, Flamm SD, Kim RJ, Nagel E. Standardized
cardiovascular magnetic resonance imaging (CMR) protocols, society
for cardiovascular magnetic resonance: board of trustees task force on
standardized protocols. J Cardiovasc Magn Reson 2008;10:35–45.
29. Tsyvian P, Malkin K, Wladimiroff JW. Assessment of fetal left cardiac
isovolumic relaxation time in appropriate and small-for-gestationalage
fetuses. Ultrasound Med Biol 1995;21:739–743.
30. Manduca A, McGee K, Welch E, Felmlee J, Ehman RL. Autofocusing
of clinical shoulder MR images for correction of motion artifacts.
MICCAI 1998:598–605.
31. McGee KP, Felmlee JP, Jack CR, Manduca A, Riederer SJ, Ehman RL.
Autocorrection of three-dimensional time-of-flight MR angiography of
the Circle of Willis. AJR Am J Roentgenol 2001;176:513–518.
32. McGee K, Felmlee J, Ehman R. Autocorrection of MR images using
multiple metrics. US Patent 6,265,874. 2001.
33. Godfrey ME, Messing B, Cohen SM, Valsky DV, Yagel S. Functional
assessment of the fetal heart: a review. Ultrasound Obstet Gynecol
2012;39:131–144.
34. Loomba RS, Chandrasekar S, Shah PH, Sanan P. The developing role of
fetal magnetic resonance imaging in the diagnosis of congenital cardiac
anomalies: a systematic review. Ann Pediatr Card 2011;4:172–176.
35. Lotz J, Meier C, Leppert A, Galanski M. Cardiovascular flow measurement
with phase-contrast MR imaging: basic facts and implementation.
Radiographics 2002;22:651–671.
36. Thiele H, Nagel E, Paetsch I, Schnackenburg B, Bornstedt A, Kouwenhoven
M, Wahl A, Schuler G, Fleck E. Functional cardiac MR imaging
with steady-state free precession (SSFP) significantly improves endocardial
border delineation without contrast agents. JMRI 2001;14:362–367.
37. Rathi VK, Biederman RWW. Imaging of ventricular function by cardiovascular
magnetic resonance. Curr Cardiol Rep 2004;6:55–61.
38. Alfakih K, Reid S, Jones T, Sivananthan M. Assessment of ventricular
function and mass by cardiac Magn Reson Imaging Eur Radiol
2004;14:1813–1822.
39. Peshock RM, Willett DL, Sayad DE, Hundley WG, Chwialkowski MC,
Clarke GD, Parkey RW. Quantitative MR imaging of the heart. Magn
Reson Imaging Clin N Am 1996;4:287–305.
Dynamic Imaging of the Fetal Heart Using MOG 1607