Content uploaded by Charlene Liew
Author content
All content in this area was uploaded by Charlene Liew on Nov 04, 2017
Content may be subject to copyright.
ORIGINAL RESEARCH
Clinical Performance of High-Resolution
Late Gadolinium Enhancement Imaging
With Compressed Sensing
Tamer A. Basha, PhD,
1,2
Mehmet Akc¸akaya, PhD,
3,4
Charlene Liew, MD,
1
Connie W. Tsao, MD,
1
Francesca N. Delling, MD,
1
Gifty Addae, BS,
1
Long Ngo, PhD,
1
Warren J. Manning, MD,
1,5
and Reza Nezafat, PhD
1
*
Purpose: To evaluate diagnostic image quality of 3D late gadolinium enhancement (LGE) with high isotropic spatial
resolution (1.4 mm
3
) images reconstructed from randomly undersampled k-space using LOw-dimensional-structure
Self-learning and Thresholding (LOST).
Materials and Methods: We prospectively enrolled 270 patients (181 men; 55 614 years) referred for myocardial viability
assessment. 3D LGE with isotropic spatial resolution of 1.4 60.1 mm
3
was acquired at 1.5T using a LOST acceleration rate
of 3 to 5. In a subset of 121 patients, 3D LGE or phase-sensitive LGE were acquired with parallel imaging with an accelera-
tion rate of 2 for comparison. Two readers evaluated image quality using a scale of 1 (poor) to 4 (excellent) and assessed
for scar presence. The McNemar test statistic was used to compare the proportion of detected scar between the two
sequences. We assessed the association between image quality and characteristics (age, gender, torso dimension, weight,
heart rate), using generalized linear models.
Results: Overall, LGE detection proportions for 3D LGE with LOST were similar between readers 1 and 2 (16.30% vs.
18.15%). For image quality, readers gave 85.9% and 80.0%, respectively, for images categorized as good or excellent.
Overall proportion of scar presence was not statistically different from conventional 3D LGE (28% vs. 33% [P50.17] for
reader 1 and 26% vs. 31% [P50.37] for reader 2). Increasing subject heart rate was associated with lower image quality
(estimated slope 5–0.009 (P50.001)).
Conclusion: High-resolution 3D LGE with LOST yields good to excellent image quality in >80% of patients and identifies
patients with LV scar at the same rate as conventional 3D LGE.
Level of Evidence: 2
J. MAGN. RESON. IMAGING 2017;00:000–000.
Myocardial scar and fibrosis can be imaged using late gad-
olinium enhancement (LGE) cardiac magnetic reso-
nance (MR).
1
LGE has been utilized in identifying ventricular
tachycardia (VT) substrates, and predicting risk of sudden car-
diac death.
2–4
Furthermore, the presence and extent of LGE
or heterogeneous left ventricular (LV) scar strongly predicts
adverse cardiac events including appropriate implantable car-
dioverter defibrillator therapy.
5–8
However, challenges in LGE
imaging remain. The right ventricle (RV) can also develop
fibrosis, but this is difficult to detect due to limited spatial
resolution and its thin wall. Similarly, LGE enables imaging of
left atrial (LA) scar and fibrosis in patients with atrial fibrilla-
tion
9–12
, however, the low spatial resolution of current LGE
imaging sequences makes it challenging to reliably image LA
scar. Therefore, there is an unmet clinical need to further
improve LGE spatial resolution in a clinically feasible scan
time.
LV LGE imaging is commonly performed using 2D
breath-hold imaging with a slice thickness of 8–10 mm.
13
3D LGE acquisition is an alternative approach,
14–19
which
View this article online at wileyonlinelibrary.com. DOI: 10.1002/jmri.25695
Received Mar 28, 2016, Accepted for publication Feb 15, 2017.
This article was published online on 16 March 2017. An error was subsequently identified. This notice is included in the online and print versions to indicate
that both have been corrected 8 April 2017.
*Address reprint requests to: R.N., Beth Israel Deaconess Medical Center, 330 Brookline Ave., Boston, MA, 02215. E-mail: rnezafat@bidmc.harvard.edu
From the
1
Department of Medicine, Beth Israel Deaconess Medical Center and Harvard Medical School, Boston, Massachusetts, USA;
2
Systems and
Biomedical Engineering Department, University of Cairo, Cairo, Egypt;
3
Department of Electrical and Computer Engineering, University of Minnesota,
Minneapolis, Minnesota, USA;
4
Center for Magnetic Resonance Research, University of Minnesota, Minneapolis, Minnesota, USA; and
5
Department of
Radiology, Beth Israel Deaconess Medical Center and Harvard Medical School, Boston, Massachusetts, USA
V
C2017 International Society for Magnetic Resonance in Medicine 1
enables imaging with higher spatial resolution for better assess-
ment of LV scar heterogeneity, RV, or LA scar. However, 3D
high-resolution LGE requires long scan times, which has limit-
ed its clinical adoption. For example, in a recent study by
Andreu et al,
20
the reported 3D LGE scan time was 16.4 67.2
minutes for a spatial resolution of 1.4 31.4 31.4 mm
3
.
Besides patient comfort, a long scan time also creates additional
artifacts in 3D LGE, such as difficulty in nulling of healthy
myocardium due to temporal changes in contrast agent concen-
tration. Parallel imaging can reduce the scan time; however, the
acceleration rate is often limited to a factor of 2 because of the
LGE low signal-to-noise ratio (SNR). Compressed sensing (CS)
has been previously used to further reduce the 3D LGE scan
time.
18,21
The LOw-dimensional-structure Self-learning and
Thresholding (LOST) reconstruction technique has been
shown to enable 3D LGE acceleration rates up to rate 3.
21,22
The combination of parallel imaging and LOST has also been
shown to further the accelerate imaging up to rate 6 in coronary
MRI.
23
However, CS-based image reconstructions (eg, LOST)
are not yet commercially available. Image reconstruction there-
fore needs to be performed manually offline, which is not ideal
in clinical settings. Furthermore, prior literature assessing
diagnostic image quality of high-resolution 3D LGE with
accelerated data acquisition and CS reconstruction is lacking.
Materials and Methods
All imaging sequences were implemented on a 1.5T Philips Achieva
(Philips Healthcare, Best, The Netherlands) system with a
32-channel cardiac phased-array receiver coil. The research protocol
was approved by our Institutional Review Board and written
informed consent was obtained from all participants in a HIPAA-
compliant manner.
Patient Study
In this prospective study, we recruited patients referred for clinical
cardiac MR exam with LGE assessment. Indications for imaging
included known or suspected coronary disease, nonischemic cardio-
myopathy, or atrial fibrillation. LGE images were acquired 10–20
minutes after a bolus (2 ml/s) infusion of 0.2 mmol/kg of Gd-DTPA
(Magnevist, Bayer Schering Pharma, Berlin, Germany), or 0.1–0.2
mmol/kg of Gd-BOPTA (MultiHance, Bracco, Rome, Italy). All
patients with estimated glomerular filtration rate (eGFR) between
30–60 mL/min/1.73m
2
received 0.1 mmol/kg of Gd-BOPTA.
A free-breathing electrocardiogram (ECG)-triggered navigator-
gated inversion-recovery gradient echo imaging sequence was used
for all acquisitions. Prior to each scan, a Look-Locker acquisition was
used to select the inversion time to null the LV myocardial signal. To
simplify the clinical protocol, the ECG trigger delay was selected
to be the longest trigger time, thereby accommodating the LGE
sequence acquisition window. For patients with heart rates >80 bpm
(n529), imaging was performed in systole. A respiratory navigator
(2D spiral pencil-beam) placed on the dome of the right hemidiaph-
ragm was used for respiratory motion compensation, utilizing pro-
spective real-time correction. A navigator with fixed scan time was
used to adaptively change the acquisition window throughout the
scan to achieve a fixed navigator efficiency of 60%.
24
A gradient echo
imaging sequence with the following parameters was used for 3D
LGE: repetition time / echo time (TR/TE) 55.2/2.6 msec, field of
view (FOV) 5320 3320 3100–120 mm
3
, flip angle 5258. Imag-
ing was performed axially, covering the entire heart. Saturation bands
along the phase-encode direction were used to reduce foldover arti-
facts. A right–left phase-encoding direction was used to reduce respi-
ratory artifacts from the chest wall. The spatial resolution varied from
1.0–1.5 mm
3
(except for one patient who was imaged with a spatial
resolution of 2 mm
3
to accommodate for large patient size) with
mean 6standard deviation of 1.4 60.1 mm
3
.
In a subset of patients, we also acquired 3D LGE or 3D phase-
sensitive inversion recovery (PSIR) images. These images were
acquired with a uniform undersampling pattern and were recon-
structed on the scanner using Philips SENSE reconstruction imple-
mentation. The ordering of acquisition was randomly changed in
different patients. The typical imaging parameters for 3D LGE were:
spatial resolution 51.5 31.5 33mm
3
,FOV5320 3320 3100–
120 mm
3
, flip angle 5258, TR/TE 55.3/2.5 msec, acceleration rate
of 2, and for 3D PSIR LGE: spatial resolution 51.5 31.5 3
5mm
3
,FOV5320 3320 3100–120 mm
3
, TR/TE 55.7/2.7
msec, flip angle 5158, and acceleration rate of 2. The imaging
parameters for 3D LGE and 3D PSIR LGE were our standard clinical
imaging protocol and were not modified in our study.
LOST Data Acquisition and Reconstruction
3D random undersampling was implemented for the accelerated
acquisitions.
21,22
A pseudorandom k-space undersampling pattern
was generated by fully sampling a part of the central k-space (corre-
sponding to 15–20% of k
y
lines and 25% of k
z
lines), and randomly
discarding the edges. This pattern was stored as a lookup table, and a
profile reordering was performed to sort the selected k
y
-k
z
lines based
on their k
y
and k
z
location in a radial fashion.
25
The acceleration rate
was 3 in 150 patients, 4 in 25 patients, 5 in 92 patients, and 6 in
three patients. The acceleration rate was selected by the technologists.
In patients with longer imaging protocols, a higher acceleration rate
was chosen. The total acquisition times were 6:25 minutes for
3-fold accelerated acquisitions and 4 minutes for 5-fold accelerated
acquisitions at a heart rate of 70 bpm. The exact scan time varied
between different subjects to accommodate for changes in imaging
field of view, spatial resolution, and heart rate.
The detailed explanation for the reconstruction algorithm and
parameters has been previously described.
21,22
The randomly under-
sampled k-space data are reconstructed using an iterative B
1
-weighted
LOST algorithm.
23
LOST reconstruction uses an initial image esti-
mate to adaptively identify 2D image blocks of similar signal content,
and to generate an adaptive sparse representation for the data. These
blocks are declared to be similar if this distance is less than a prede-
fined fixed threshold. Such similar blocks were then grouped into the
similarity cluster of that voxel. The adaptive sparse representation was
generated by applying a 3D fast Fourier transform (FFT) to each
similarity cluster. Dealiasing of the data was performed via shrinkage
of the 3D FFT coefficients of the similarity clusters. The iterative
B
1
-weighted reconstruction approach uses the coil sensitivity infor-
mation for data consistency during the reconstruction. For faster
implementation, coil compression
26
is applied to these 3D coil
sensitivity maps to compress the 32-channel coil data to 10 “virtual”
Journal of Magnetic Resonance Imaging
2 Volume 00, No. 00
channels. The coil compression was performed only once, prior
to the iterative portion of the algorithm. At every iteration of the
B
1
-weighted LOST algorithm: 1) the current combined-coil image
estimate is thresholded using LOST; 2) individual coil images are
generated by the voxel-wise multiplication of the coil sensitivity map
of that coil and the combined image; 3) data consistency is enforced
by replacing the acquired k-space locations with the acquired lines for
each of these individual coil images; and 4) a new combined-coil
image estimate is generated by summing the voxelwise product of the
data-consistent coil images and the complex conjugate of the coil
sensitivity maps across the coil dimension. LOST reconstruction was
implemented in MatLab (MathWorks, Natick, MA), with the
adaptive learning and nonlinear shrinkage portions implemented in
C11. The same reconstruction parameters were used in all cases and
acceleration rates, allowing for fully automated reconstructions.
21,24
Automated Reconstruction Framework
A software platform was developed to automate the communica-
tion and data transfer between the scanner and the remote process-
ing units. Figure 1 shows a schematic diagram of these connections
illustrating the sequence of the program workflow. After a scan was
completed, the operator used the program to request the recon-
struction of the scanned data, which initiates the following fully
automated workflow: 1) the program inquires the scanner database
for the scan information including the raw data and the imaging
parameters; 2) the database returns the desired information; 3) raw
data are packaged along with the imaging parameters and recon-
struction options, and are then sent to a remote processing unit for
image reconstruction; 4) during the processing, the program peri-
odically connects with the remote processing unit and updates the
progress of the operation; and 5) upon reconstruction completion,
the reconstructed images are automatically pulled from the remote
processing unit, packaged into a Digital Imaging and Communica-
tions in Medicine (DICOM) format, and then pushed to the scan-
ner database, from where it can also be sent to the hospital Picture
Archiving and Communication System (PACS). In this workflow,
multiple datasets can be sent at the same time or sequentially. All
workflow processes are implemented such that they are performed
in the background and do not block operator interaction with the
scanner console. Before sending the data, the software generates
and attaches a random unique identifier (UID) for the specific
data, and then uses this UID to inquire about progress, retrieve
the results when ready, and push them into the scanner database
with the proper patient and series information to comply
with HIPAA, especially if the remote system is located outside the
imaging facility or using an available web-based cloud computational
system. The operator can choose to have the remote processing occur
on the scanner machine or at a different remote station on the net-
work, or even on a CPU cluster or GPU server on the same network.
In our implementation, a CPU cluster was used for performing the
reconstruction. Once the reconstructed images are available in the
scanner database or in the PACS, they can be viewed and stored
similar to images reconstructed by the vendor reconstruction system.
The software program was developed in MatLab.
Image Analysis
LOST reconstructed 3D LGE images were written into DICOM
format and were used for subjective assessment. In patients with
available clinical 3D LGE or PSIR images, reconstructed images
from the scanner vendor software were available and were used for
assessment. Subjective qualitative assessment was performed by two
independent readers (each with 10 years of experience in interpreting
clinical cardiac MR exams, C.T., F.D.), blinded to patient history.
These readers were not aware of the specifics of the reconstructions;
however, because of variations in contrast and resolution, they could
differentiate between LOST reconstructed images vs. conventional
SENSE reconstruction. To minimize this impact, image assessment
for LOST and SENSE reconstruction were performed in two separate
sessions. Readers reviewed each dataset in Osirix (Pixmeo SARL,
Geneva, Switzerland) and were free to adjust the window leveling.
Both original and multiplanar reformatted images were used for
subjective assessment.
For each dataset, the presence of LGE was assessed using a three-
point scale (absent with confidence 50, present with confidence 51,
unable to interpret/inconclusive 52). Furthermore, each dataset was
qualitatively assessed using a four-point ordinal system based on overall
image quality to assess for viability/scar: 1, poor (large artifacts/no con-
fidence in interpreting LGE in >50% of the myocardium); 2, fair
(moderate artifacts/LGE interpretable in 50–70% of the myocardium);
3, good (small artifacts/LGE interpretable in 75–90% of the
myocardium); 4, excellent (no artifacts/LGE interpretable in >90% of
the myocardium).
To assess the potential impact of various patient characteristics on
image quality, we extracted the following parameters for each patient:
age,gender,heartrate,weight,torso dimension (ie, posteroanterior
[back to chest wall] and transverse [right to left]).
Statistical Analysis
We first analyzed data for all subjects with 3D LGE with LOST.
The proportion of the detected LGE presence (“Present with Con-
fidence”) was reported for each reader, as well as the agreement
kappa statistic and 95% confidence interval (CI). Stratification by
acceleration rate was carried out to yield these estimates within
each rate stratum (acceleration rate 3, 4, and 5). We performed
similar analysis for image quality by grouping “good” and
“excellent” image quality in one category and estimating the pro-
portion of “good or excellent” for overall image quality and by
stratification of rate. The kappa statistic was computed. Second, we
analyzed data for the sample of 121 subjects with data from both
sequences (3D LGE with LOST, and conventional 3D LGE). We
FIGURE 1: Integration of the automated reconstruction frame-
work into the clinical workflow: (1) the scanner database is que-
ried for scan raw data and imaging parameters, (2) the
database returns the requested information, (3) raw data are
packed and sent to the Central Processing Unit (CPU) cluster
for LOST reconstruction, (4) progress and results are updated
upon request from the operator, (5) reconstructed images are
pushed to the scanner database, from where they can also be
sent to the hospital PACS.
Basha et al.: High-Resolution LGE With Compressed Sensing
Month 2017 3
also reported the same statistics as in the sample of 270 subjects.
Additionally, we compared the “Present with Confidence” of LGE
between the two sequences using the McNemar test statistic to
account for dependency in the sample (since subjects in this group
were imaged with both sequences). We used SAS software v. 9.4
(Cary, NC) for data management and statistical analysis.
Results
We recruited 270 patients (181 men; 54.9 614.1 years)
who were imaged using high-resolution 3D LGE with
LOST. We also acquired 3D LGE or 3D PSIR images in a
subset of 121 patients. The reconstruction framework
allowed successful automation of image reconstruction in all
cases without the need for any interaction with research sci-
entists on the investigator’s team. Figure 2 shows an example
of 3D LGE with high isotropic resolution, as well as 3D
PSIR LGE image acquired in a 65-year-old-man with a his-
tory of coronary artery disease and ventricular tachycardia.
Improved spatial resolution allows better depiction of
complex scar geometry in the patient. In high-resolution
images, several additional small focal enhancements can be
seen that were not seen in 3D PSIR images due to limited
FIGURE 2: Reformatted 3D LGE images (3D high-resolution LGE with isotropic spatial resolution with LOST rate 5 and 3D PSIR
LGE with nonisotropic spatial resolution) in a 65-year-old man with coronary artery disease, symptomatic nonsustained ventricular
tachycardia, and presyncope referred for evaluation of LV scar. Subendocardial LGE in the left anterior descending coronary distri-
bution can be seen in all images. Additionally, smaller scar in remote areas is visible at different slice locations with significantly
better depiction of scar in 3D high-resolution LGE with isotropic spatial resolution. Smaller areas of scar are not visible in 3D PSIR
LGE with nonisotropic spatial resolution.
FIGURE 3: Reformatted 3D LGE of a 67-year-old man with hypertrophic cardiomyopathy. Focal hypertrophy of the mid inferior
left ventricular septum. High-resolution 3D high-resolution LGE images acquired with LOST rate 3 demonstrate a more detailed
depiction of scar compared with 3D PSIR LGE.
Journal of Magnetic Resonance Imaging
4 Volume 00, No. 00
spatial resolution. The isotropic spatial resolution also allows
reformatting in any desired orientation (as shown) with
identical image quality as the original axial images. Because
of nonisotopic spatial resolution of PSIR LGE, there is
additional partial voluming and lower image quality in
reformatted images. Viability images from a patient with
hypertrophic cardiomyopathy (HCM) are shown in Fig. 3.
The isotropic spatial resolution of accelerated 3D LGE
shows the complex nature of the scar in this patient includ-
ing papillary muscle. While hyperenhancement can be
assessed with 3D PSIR LGE, the visualization of complex
scar and scar patchiness are better seen with high-resolution
3D LGE with isotropic spatial resolution. Figure 4 shows
viability images in a 62-year-old male patient with sarcoido-
sis and LV aneurysms. Higher spatial resolution images
improve detection of hyperenhancement region, as well as
confirm the involvement of the RV.
Table 1 shows the proportions of “Present with Con-
fidence” for the presence of detected LGE in 3D LGE with
LOST; 16.30% and 18.15% for readers 1 and 2, respectively.
These proportions agree well with a kappa statistic of 0.78
(95% CI: 0.68–0.88). These proportions are also similar
when the sample was stratified by acceleration rate (3–5). For
acceleration rate 4, the sample size is small, and compared to
rates 3 and 5, the proportions are smaller. Similar analysis was
carried out for image quality. The proportion of “Good or
Excellent” is high, 85.93% for reader 1, and 80.00% for read-
er 2, with a fair kappa statistic of 0.43 (95% CI 0.22–0.57).
Reader 1 scored 4, 34, 88, and 144 as poor, fair, good, and
excellent, respectively, while reader 2 scored 7, 47, 135, and
81 as poor, fair, good, and excellent, respectively. Stratification
by rate also yielded similar proportions.
Table 2 shows the analysis for the 121 subjects with
data for both 3D LGE with LOST, and conventional 3D
LGE. For reader 1, the overall proportion of “Present with
Confidence” is 28.93% for 3D LGE with LOST, and is not
statistically different from 33.88% for conventional 3D
LGE (P50.157, McNemar test). We stratified this sample
FIGURE 4: Viability images from a 62-year-old man with cardiac sarcoidosis, comparing high-resolution 3D LGE with 1.2 mm
3
isotropic
resolution with LOST rate 4 with 3D PSIR LGE with nonisotropic resolution. 3D high-resolution LGE significantly improves visualization
of scar and shows involvement of the RV free wall as well as signal enhancement in the RV papillary muscles and apical LV aneurysm.
Basha et al.: High-Resolution LGE With Compressed Sensing
Month 2017 5
by acceleration rate 3, 4, 5, and found no significant statisti-
cal difference between the two proportions. Numerically,
3D LGE with LOST estimates are slightly less than those of
the conventional 3D LGE overall, and within each stratum
of acceleration rate. The agreement between the two readers
via kappa statistic is high at 0.90 (95% CI: 0.81–0.99) for
“Present with Confidence” for LGE detection. Conventional
3D LGE also yields a high kappa of 0.87 (95% CI: 0.77–
0.96). For image quality, “Good or Excellent” is 85.95% for
reader 1 for 3D LGE with LOST, and 84.30% for conven-
tional 3D LGE (P50.637). For reader 2, there is also no
statistical difference between the two approaches (78.51%
vs. 85.95%, P50.083). The agreement is fair with kappa
of 0.36 (95% CI: 0.15–0.56) for 3D LGE with LOST. For
conventional 3D LGE, the kappa is high 0.80 (95% CI:
0.65–0.96).
The baseline characteristics of the study population
and torso size measurements are shown in Table 3. Increas-
ing subject heart rate, weight, and torso dimension are asso-
ciated negatively with subjective image quality (Table 4),
but only heart rate reached statistical significance.
Discussion
We developed and implemented a clinically feasible image
reconstruction framework for accelerated data acquisition
and reconstruction using the LOST technique. The image
quality of 3D LOST-accelerated LGE was evaluated in 270
patients with known or suspected cardiovascular disease
referred for clinical viability assessment. We demonstrated
that using this approach, 3D LGE with high isotropic spa-
tial resolution is clinically feasible with acceptable image
quality in most patients.
Despite the growth in development of CS-based image
reconstruction techniques, there have been very limited
efforts in clinical translation and evaluation.
27–34
Sharma
et al
29
evaluated CS-based reconstruction for routine neuro-
imaging sequences and found that CS was able to moderate-
ly accelerate certain neuroimaging sequences without severe
loss of clinically relevant information. For those sequences
with coarser spatial resolution and/or at a higher accelera-
tion factor, artifacts degraded the quality of the recon-
structed image to the point where they are of minimal
or no clinical value. Vasanawala et al
30
showed the feasibility
TABLE 1. Subjective Assessment of 3D LGE With LOST Reconstruction (N 5270) for Two Readers and Inter-
reader Agreement for Images Acquired via Different Acceleration Rates
Hyperenhancement on LGE
Present with
confidence
for all rates
(N5270)
Present with
confidence
for rate 53
(N5150)
Present with
confidence for
rate 54
(N525)
Present with
confidence
for rate 55
(N592)
Reader 1 3D LGE with LOST 16.30% 15.33% 8.00% 20.65%
Reader 2 - 3D LGE with LOST 18.15% 16.00% 12.00% 22.83%
Image quality
Good or
excellent for
all rates
(N5270)
Good or
excellent for
rate 53
(N5150)
Good or
excellent for
rate 54
(N525)
Good or
excellent for
rate 55
(N592)
Reader 1 - 3D LGE with LOST 85.93% 88.00% 96.00% 79.35%
Reader 2 - 3D LGE with LOST 80.00% 79.33% 96.00% 76.09%
Interreader agreement
Kappa 95% Confidence
Interval
3D LGE with LOST Score 0.78 0.68-0.88
3D LGE with LOST Image Quality 0.43 0.29-0.57
Journal of Magnetic Resonance Imaging
6 Volume 00, No. 00
of improved pediatric imaging to achieve higher resolution
and/or faster imaging in 34 patients in a clinical setting. Hsiao
et al
31
assessed the potential of CS in 4D phase-contrast MRI
for the evaluation of valvular insufficiency and intracardiac
shunts in 34 patients with congenital heart disease and
demonstrated that the CS 4D phase-contrast sequence can
TABLE 2. Subjective Assessment of the Subset of Patients With Both 3D LGE With LOST Reconstruction and
Conventional 3D LGE (N5121) for Two Readers and Interreader Agreement for Images Acquired via Different
Acceleration Rates (3 to 5)
Hyperenhancement on LGE
Present with
confidence
for all rates
(N5121)
Present with
confidence
for rate 53
(N548)
Present with
confidence
for rate 54
(N513)
Present with
confidence
for rate 55
(N560)
Reader 1 - 3D LGE with LOST 28.93% 33.33% 7.69% 30.00%
Reader 1 - Conventional 3D LGE 33.88% 35.42% 30.77% 33.33%
P-value Comparison of LOST and
Conventional LGE for Reader 1
a
0.157 0.739 0.083 0.414
Reader 2 - 3D LGE with LOST 26.45% 29.17% 7.69% 28.33%
Reader 2 - Conventional 3D LGE 31.40% 35.42% 23.08% 30.00%
P-value Comparison of LOST and
Conventional LGE for Reader 2
a
0.134 0.317 0.157 0.655
Image quality
Good or
Excellent for
All Rates
(N5121)
Good or
Excellent for
Rate 53
(N548)
Good or
Excellent for
Rate 54
(N513)
Good or
Excellent for
Rate 55
(N560)
Reader 1 - 3D LGE with LOST 85.95% 85.42% 100.00% 83.33%
Reader 1 - Conventional 3D LGE 84.30% 85.42% 100.00% 80.00%
P-value Comparison of LOST and
Conventional Image Quality for Reader 1
a
0.637 1.000 1.000 0.564
Reader 2 - 3D LGE with LOST 78.51% 75.00% 92.71% 78.33%
Reader 2 - Conventional 3D LGE 85.95% 87.50% 100.00% 81.67%
P-value Comparison of LOST and
Conventional Image Quality for Reader 2
a
0.083 0.109 1.000 0.564
Interreader agreement
Kappa
b
95% Confidence
Interval
3D LGE with LOST Score 0.90 0.81-0.99
Conventional 3D LGE Score 0.87 0.77-0.96
3D LGE with LOST Image Quality 0.36 0.15-0.56
Conventional 3D LGE Image Quality 0.80 0.65-0.96
a
McNemar test.
b
Kappa based on 2 by 2 table present with confidence versus otherwise, good or excellent.
Basha et al.: High-Resolution LGE With Compressed Sensing
Month 2017 7
augment conventional cardiac MR imaging by improving sen-
sitivity for and depiction of hemodynamically significant
shunts and valvular regurgitation. Roujol et al evaluated the
performance of CS-accelerated ECG gated pulmonary vein
MRA in 19 patients with atrial fibrillation.
28
Mann et al
assessed the utility of CS in quantification of the fat fraction
in 11 patients with type 2 diabetes and concluded that accept-
able image quality can be achieved with acceleration up to a
factor of 3.8.
27
These exploratory studies demonstrated the
potential and clinical feasibility of the CS-based approach for
accelerated imaging; however, evaluations were only per-
formed in a small and selected number of patients (ie, between
20–40).
Despite excellent image quality, the long reconstruc-
tion time remains one of the main limitations of this tech-
nique, with a typical reconstruction time of 1 hour for
high-resolution 3D LGE. In our implementation, a 3D vol-
ume reconstructed by zero filling of the undersampled data
will be available immediately after imaging. Because of the
fully sampled center of k-space, we found these images to
be helpful in predicting the image quality of the CS recon-
structed images, and often scar can be seen in these images.
While this workflow does not work on cases where immedi-
ate access to reconstructed data is needed (eg, real-time
MRI), immediate reconstruction is not necessary in our spe-
cific application of 3D LGE. In our clinical practice, images
from each patient are typically evaluated the next day,
allowing sufficient time for image reconstruction for clinical
evaluation. Further research to reduce the reconstruction
time and development of alternative reconstruction techniques
is warranted.
Our study only evaluated the image quality of
isotropic 3D LGE with high resolution in our tertiary care
medical center’s referral population. The clinical impact of
high spatial resolution imaging for different diseases should
be carefully assessed in future studies. Not all patients
may need imaging with high spatial resolution; however, for
some cardiovascular diseases this may be clinically war-
ranted. For example, a higher spatial resolution may allow
better visualization of the LV scar morphology, which may
facilitate assessment for ventricular arrhythmia. In addition,
a volumetric isotropic 3D LGE allows for image reconstruc-
tion in any orientation and multiplanar visualization of the
data, which further improves confidence in detection of scar
by confirming the scar in different views. Further studies are
needed to establish the clinical impact of high-resolution
isotropic 3D LGE. We also found that there were differ-
ences in scoring between the two readers for the LOST
accelerated imaging dataset. While the proportion of good
to excellent quality was similar between the two readers, one
of our reader’s scores were more positive and a majority of
images were scored as excellent.
Image reconstruction for CS is commonly performed
offline, which has hindered evaluation of its various recon-
struction algorithms in a clinical environment. Other stand-
alone reconstruction frameworks have been described. A flexi-
ble real-time MRI reconstruction platform (ie, RTHawks) has
been widely applied in real-time MRI.
35
A real-time display
and reconstruction system was also developed at the National
Institutes of Health and is widely used for interventional pro-
cedures.
36
Recently, a flexible stand-alone MRI image recon-
struction framework
37,38
has been developed that provides
additional flexibility for reconstruction and use of cloud com-
puting to facilitate image reconstruction. These reconstruction
platforms are designed mainly for situations where real-time
or immediate reconstruction of MRI data is needed, such as
image guidance or real-time imaging.
Our study has several limitations. We did not perform a
direct comparison between 3D and 2D LGE imaging, where
the latter is the current clinical standard at many medical cen-
ters. Currently, all patients imaged for LV viability at our
TABLE 3. Baseline Characteristics of the Study Popu-
lation (N5270)
Baseline patient characteristics N5270
Age years (range) 55 (18-85)
Male subjects 181 (66.7%)
Heart rate (/min) 66 613
Weight (kg) 83 615
Chest posteroanterior
dimension (cm)
23.1 63.4
Chest transverse
dimension (cm)
35.3 63.6
TABLE 4. Relationship Between Subjective Average
Qualitative Image Assessment Scores and Patient
Factors
Patient Factors Change in mean of
average image score
per unit change in
patient factor
a
P-value
Heart rate -0.009 (0.003) 0.001
Weight -0.003 (0.003) 0.323
Chest AP
dimension
-0.012 (0.015) 0.421
Chest transverse
dimension
-0.027 (0.015) 0.064
a
Slope estimate and standard error from multivariable generalized
linear model.
Journal of Magnetic Resonance Imaging
8 Volume 00, No. 00
medical center are imaged using 3D LGE. The spatial resolu-
tion varied between different patients to accommodate patient
size and heart rate. Due to the inherent denoising in CS, we
cannot directly perform SNR or CNR measurements. We
used fixed image reconstruction parameters for all studies and
did not optimize the parameters based on acceleration rate.
In conclusion, 3D LGE with LOST yields good to
excellent image quality in >80% of patients and identifies
patients with LV scar at the same rate as conventional 3D
LGE. High isotropic spatial resolution (1–1.4 mm
3
) detects
smaller scar and enables multiplanar reconstruction of 3D
LGE to facilitate scar visualization.
Acknowledgments
Contract grant sponsor: National Institutes of Health (NIH);
contract grant numbers: R01EB008743, R01HL129185-01,
R01HL127015, R21HL127650, NIH K99HL111410 and
R00HL111410 (to M.A.); Contract grant sponsor: American
Heart Association (AHA); contract grant number:
15EIA22710040.
We thank Sophie Berg, Kraig V. Kissinger, and Beth Goddu
for patient recruitment and scanning.
References
1. Kim RJ, Fieno DS, Parrish TB, et al. Relationship of MRI delayed con-
trast enhancement to irreversible injury, infarct age, and contractile
function. Circulation 1999;100:1992–2002.
2. Schmidt A, Azevedo CF, Cheng A, et al. Infarct tissue heterogeneity
by magnetic resonance imaging identifies enhanced cardiac arrhyth-
mia susceptibility in patients with left ventricular dysfunction. Circula-
tion 2007;115:2006–2014.
3. Yan AT, Shayne AJ, Brown KA, et al. Characterization of the peri-
infarct zone by contrast-enhanced cardiac magnetic resonance imag-
ing is a powerful predictor of post-myocardial infarction mortality. Cir-
culation 2006;114:32–39.
4. Roes SD, Borleffs CJ, van der Geest RJ, et al. Infarct tissue heteroge-
neity assessed with contrast-enhanced MRI predicts spontaneous ven-
tricular arrhythmia in patients with ischemic cardiomyopathy and
implantable cardioverter-defibrillator. Circ Cardiovasc Imaging 2009;2:
183–190.
5. Assomull RG, Prasad SK, Lyne J, et al. Cardiovascular magnetic reso-
nance, fibrosis, and prognosis in dilated cardiomyopathy. J Am Coll
Cardiol 2006;48:1977–1985.
6. Wu KC, Weiss RG, Thiemann DR, et al. Late gadolinium enhancement
by cardiovascular magnetic resonance heralds an adverse prognosis
in nonischemic cardiomyopathy. J Am Coll Cardiol 2008;51:2414–
2421.
7. Iles L, Pfluger H, Lefkovits L, et al. Myocardial fibrosis predicts appro-
priate device therapy in patients with implantable cardioverter-
defibrillators for primary prevention of sudden cardiac death. J Am
Coll Cardiol 2011;57:821–828.
8. Rayatzadeh H, Tan A, Chan RH, et al. Scar heterogeneity on cardio-
vascular magnetic resonance as a predictor of appropriate implant-
able cardioverter defibrillator therapy. J Cardiovasc Magn Reson
2013;15:31.
9. McGann CJ, Kholmovski EG, Oakes RS, et al. New magnetic reso-
nance imaging-based method for defining the extent of left atrial wall
injury after the ablation of atrial fibrillation. J Am Coll Cardiol 2008;
52:1263–1271.
10. Peters DC, Wylie JV, Hauser TH, et al. Recurrence of atrial fibrillation
correlates with the extent of post-procedural late gadolinium enhance-
ment: a pilot study. JACC Cardiovasc Imaging 2009;2:308–316.
11. Marrouche NF, Wilber D, Hindricks G, et al. Association of atrial tissue
fibrosis identified by delayed enhancement MRI and atrial fibrillation
catheter ablation: the DECAAF study. JAMA 2014;311:498–506.
12. Moghari MH, Peters DC, Smink J, et al. Pulmonary vein inflow artifact
reduction for free-breathing left atrium late gadolinium enhancement.
Magn Reson Med 2011;66:180–186.
13. Kramer C, Barkhausen J, Flamm S, et al. Standardized cardiovascular
magnetic resonance (CMR) protocols 2013 update. J Cardiovasc
Magn Reson 2013;15:91.
14. Nguyen TD, Spincemaille P, Weinsaft JW, et al. A fast navigator-
gated 3D sequence for delayed enhancement MRI of the myocardi-
um: comparison with breathhold 2D imaging. J Magn Reson Imaging
2008;27:802–808.
15. Peukert D, Laule M, Taupitz M, Kaufels N, Hamm B, Dewey M. 3D
and 2D delayed-enhancement magnetic resonance imaging for detec-
tion of myocardial infarction: preclinical and clinical results. Acad
Radiol 2007;14:788–794.
16. Peters DC, Appelbaum EA, Nezafat R, et al. Left ventricular infarct
size, peri-infarct zone, and papillary scar measurements: A comparison
of high-resolution 3D and conventional 2D late gadolinium enhance-
ment cardiac MR. J Magn Reson Imaging 2009;30:794–800.
17. Moghari MH, Chan RH, Hong SN, et al. Free-breathing cardiac MR
with a fixed navigator efficiency using adaptive gating window size.
Magn Reson Med 2015;73:1555–1561.
18. Roujol S, Basha TA, Akcakaya M, et al. 3D late gadolinium enhance-
ment in a single prolonged breath-hold using supplemental oxygena-
tion and hyperventilation. Magn Reson Med 2014;72:850–857.
19. Saranathan M, Rochitte CE, Foo TK. Fast, three-dimensional free-
breathing MR imaging of myocardial infarction: a feasibility study.
Magn Reson Med 2004;51:1055–1060.
20. Andreu D, Ortiz-P
erez JT, Fern
andez-Armenta J, et al. 3D delayed-
enhanced magnetic resonance sequences improve conducting chan-
nel delineation prior to ventricular tachycardia ablation. Europace
2015;17:938–945.
21. Akcakaya M, Rayatzadeh H, Basha TA, et al. Accelerated late gadolin-
ium enhancement cardiac MR imaging with isotropic spatial resolution
using compressed sensing: initial experience. Radiology 2012;264:
691–699.
22. Akcakaya M, Basha TA, Goddu B, et al. Low-dimensional-structure
self-learning and thresholding: Regularization beyond compressed
sensing for MRI reconstruction. Magn Reson Med 2011;66:756–767.
23. Akcakaya M, Basha TA, Chan RH, Manning WJ, Nezafat R. Accelerat-
ed isotropic sub-millimeter whole-heart coronary MRI: Compressed
sensing versus parallel imaging. Magn Reson Med 2014;71:815–822.
24. Moghari MH, Chan RH, Hong SN, et al. Free-breathing cardiac MR
with a fixed navigator efficiency using adaptive gating window size.
Magn Reson Med 2012;68:1866–1875.
25. Akcakaya M, Basha TA, Chan RH, et al. Accelerated contrast-
enhanced whole-heart coronary MRI using low-dimensional-structure
self-learning and thresholding. Magn Reson Med 2012;67:1434–1443.
26. Huang F, Vijayakumar S, Li Y, Hertel S, Duensing GR. A software
channel compression technique for faster reconstruction with many
channels. Magn Reson Imaging 2008;26:133–141.
27. Mann LW, Higgins DM, Peters CN, et al. Accelerating MR imaging liver
steatosis measurement using combined compressed sensing and paral-
lel imaging: a quantitative evaluation. Radiology 2016;278:247–256.
28. Roujol S, Foppa M, Basha TA, et al. Accelerated free breathing ECG trig-
gered contrast enhanced pulmonary vein magnetic resonance angiogra-
phy using compressed sensing. J Cardiovasc Magn Reson 2014;16:91.
29. Sharma SD, Fong CL, Tzung BS, Law M, Nayak KS. Clinical image
quality assessment of accelerated magnetic resonance neuroimaging
using compressed sensing. Invest Radiol 2013;48:638–645.
Basha et al.: High-Resolution LGE With Compressed Sensing
Month 2017 9
30. Vasanawala SS, Alley MT, Hargreaves BA, Barth RA, Pauly JM, Lustig
M. Improved pediatric MR imaging with compressed sensing. Radiol-
ogy 2010;256:607–616.
31. Hsiao A, Lustig M, Alley MT, Murphy MJ, Vasanawala SS. Evaluation
of valvular insufficiency and shunts with parallel-imaging com-
pressed-sensing 4D phase-contrast MR imaging with stereoscopic
3D velocity-fusion volume-rendered visualization. Radiology 2012;
265:87–95.
32. Jaspan ON, Fleysher R, Lipton ML. Compressed sensing MRI: a
review of the clinical literature. Br J Radiol 2015;88:20150487.
33. Zhang T, Chowdhury S, Lustig M, et al. Clinical performance of con-
trast enhanced abdominal pediatric MRI with fast combined parallel
imaging compressed sensing reconstruction. J Magn Reson Imaging
2014;40:13–25.
34. Akcakaya M, Hu P, Chuang ML, et al. Accelerated noncontrast-
enhanced pulmonary vein MRA with distributed compressed sensing.
J Magn Reson Imaging 2011;33:1248–1255.
35. Santos JM, Wright GA, Pauly JM. Flexible real-time magnetic resonance
imaging framework. Conf Proc IEEE Eng Med Biol Soc 2004;2:1048–1051.
36. Guttman MA, Kellman P, Dick AJ, Lederman RJ, McVeigh ER.
Real-time accelerated interactive MRI with adaptive TSENSE and
UNFOLD. Magn Reson Med 2003;50:315–321.
37. Hansen MS, Sorensen TS. Gadgetron: an open source framework for
medical image reconstruction. Magn Reson Med 2013;69:1768–1776.
38. Xue H, Inati S, Sorensen TS, Kellman P, Hansen MS. Distributed MRI
reconstruction using Gadgetron-based cloud computing. Magn Reson
Med 2015;73:1015–1025.
Journal of Magnetic Resonance Imaging
10 Volume 00, No. 00
