| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
1 Départment Imagerie Médicale, Centre Hospitalo-Universitaire
Henri Mondor, Université de Créteil, Paris XII, 51 Ave. du Mal
de Lattre de Tassigny, Créteil 94010 Cedex, France.
2 Siemens SAS, Division Médicale, 39/47, Blvd. Ornano, Saint-Denis 93527
Cedex 2, France.
Received June 13, 2003;
accepted after revision September 23, 2003.
Address correspondence to L. Ruel
(lucile.ruel{at}hmn.ap-hop-paris.fr).
Abstract
 Top Abstract IntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
MATERIALS AND METHODS. Phantom imaging and lumbar spine studies of 15 healthy subjects were performed using T2-weighted turbo spin-echo sequences obtained with and without parallel imaging (generalized autocalibrating partially parallel acquisition [GRAPPA] and modified sensitive encoding [mSENSE]) on a 1.5-T magnet. The signal-to-noise ratio (SNR) and uniformity were measured in the phantom, and SNR and signal difference–noise ratio were evaluated in cerebrospinal fluid, vertebral bodies, and subcutaneous fat of the volunteers, using both techniques sequentially. Aliasing artifacts on GRAPPA and mSENSE images were visually evaluated. SNRs were compared using the Student's paired t test, with p values less than 0.05 considered significant.
RESULTS. In the phantom study, when the same number of coil elements were used (n = 3), SNR and uniformity values obtained with standard T2-weighted turbo spin-echo sequences were higher than those obtained with parallel sequences. The GRAPPA SNR obtained with three coil elements was higher than the standard T2-weighted SNR obtained with one coil element. Similar findings were noted regarding uniformity. In the lumbar spine, GRAPPA SNR values for fat, cerebrospinal fluid, and vertebral bodies were significantly higher than mSENSE SNR values, with a p value less than 0.01, but were not significantly different from T2-weighted turbo spin-echo SNR values. GRAPPA signal difference–noise ratio values were significantly higher than mSENSE signal difference–noise ratio values, with a p value less than 0.01, but were not significantly different from T2-weighted turbo spin-echo signal difference–noise ratio values. GRAPPA produced fewer aliasing artifacts than mSENSE.
CONCLUSION. In spine MRI, GRAPPA may be used to reduce scanning time and yields a higher SNR than mSENSE without any increase in aliasing artifacts and with an SNR similar to that obtained with standard T2-weighted turbo spin-echo.
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
SENSE (sensitive encoding) [1, 3] and SMASH (simultaneous acquisition of spatial harmonics) [2, 4] are different approaches to parallel imaging; mSENSE (modified SENSE) and GRAPPA (generalized autocalibrating partially parallel acquisition) [5] are recent variants of SENSE and SMASH, respectively, implemented on our MRI unit (Symphony, Siemens Medical Solutions, Erlangen, Germany).
Before using these new techniques routinely, radiologists must be aware of their influence on image quality. Therefore, our aim was to compare mSENSE, GRAPPA, and conventional reconstruction techniques in terms of SNR, signal difference–noise ratio, uniformity, and potential artifacts generated on T2-weighted images, both in vitro on phantom images and in vivo on spine MR images using a single manufacturer's MRI unit. Although parallel acquisition techniques have been developed by several other manufacturers, a comparison of MRI units was not performed, suggesting that our methodology and findings may not be applicable to other manufacturers' units.
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
For both phantom and subject experiments, three reconstruction techniques, including standard nonparallel mSENSE and GRAPPA, were tested after being applied to a standard T2 turbo spin-echo sequence (TR/TE, 4,390/136; echo-train length, 23; thickness, 4 mm; matrix, 512 x 302). An additional sequence was acquired with the same parameters using parallel imaging (mSENSE and GRAPPA techniques) with an accelerator factor of 2 and an acquisition of 46 central lines. All reconstruction techniques were compared in terms of SNR and image uniformity.
In Vitro Phantom MRI
The phantom experiments were carried out on a cylindric plastic recipient
filled with 1.24 g of hydrated NiSO4 and 2.62 g of NaCl per liter
of distilled water. The phantom (37.5 cm in length and 15.5 cm in diameter)
was straddled by three coils. The phantom was positioned parallel to
12.5-cm-long coil elements. Phantom imaging procedures were repeated three
times. Acquisition times were 142 sec for turbo spin-echo nonparallel
sequences and 88 sec for GRAPPA and mSENSE turbo spinecho sequences.
In the first experiment (experiment E1), both standard nonparallel mSENSE and GRAPPA images were obtained with a 415-mm field of view, with three coil elements switched on. A second experiment (experiment E2) was carried out with only one coil element switched on for the turbo spin-echo nonparallel sequence and three coil elements switched on for GRAPPA and mSENSE turbo spin-echo sequences to assess the influence of the number of coils on image quality using parallel imaging. A 160-mm field of view was selected in this second experiment accounting for the maximum possible field of view covered by one coil element.
A complementary experiment was performed to assess the influence of the chosen phantom on SNR differences that could possibly account for SNR variations between patients in in vivo studies: three phantoms were used including a spherical phantom (240-mm diameter and filled with 1.25 g of hydrated NiSO4), an elliptic and hollow body load phantom (filled with 3 g of hydrated MnCl2 and 5 g of NaCl) and a cylindric plastic phantom (20 cm in length, 12 cm in diameter, and filled with 1.25 g of hydrated NiSO4 and 5 g of NaCl). First, SNR was measured in the cylindric phantom with (experiment E3) and without (experiment E4) the body load phantom placed next to it. Second, SNR was measured in the cylindric phantom placed inside the body load phantom with (experiment E5) and without (experiment E6) the spherical phantom placed next to it. All experiments were repeated three times.
In Vivo Spine MRI
Imaging was performed with a 380-mm field of view using three array coil
elements for nonparallel turbo spin-echo sequences and four coil elements for
parallel turbo spin-echo sequences. Phase encoding was performed in the
head–feet direction. Both nonparallel turbo spin-echo images and mSENSE
and GRAPPA images of the lumbar spine were obtained from 15 volunteers (seven
women and eight men; mean age, 43 years; age range, 23–76 years).
Acquisition times were 217 sec for turbo spin-echo nonparallel sequences and
126 sec for turbo spin-echo parallel sequences. A 30% phase oversampling was
used to limit the risk of aliasing for parallel and nonparallel imaging.
In keeping with the Declaration of Helsinki, full written informed consent was obtained from all the subjects before scanning.
Image Analysis
In vitro phantom MRI.—On images obtained after either
nonparallel mSENSE or GRAPPA reconstruction, the mean signal intensity (SI)
and background noise SD were respectively calculated across two rectangular
regions of interest (ROIs): one fitted to the size of the whole phantom and
the other positioned outside the phantom.
SNR was calculated as follows:
 | (1) |
Image uniformity was calculated using a formula modified from Price et al.
[6]:
 | (2) |
An image was generated by subtracting mSENSE images from GRAPPA images for visual assessment of residual signal to illustrate signal differences within the entire volume.
In vivo spine MRI.—With mSENSE, GRAPPA, and nonparallel turbo spin-echo images, one mid-sagittal image was selected to identify the largest spinal canal size for all ROI measurements. The mean SI and mean SD were calculated using four circular ROIs located in background, in cerebrospinal fluid, in subcutaneous fat, and in the vertebral body. The ROIs were fitted to the studied regions (Fig. 1). The size and position of the ROIs were identical for each given location and subject for mSENSE, GRAPPA, and nonparallel turbo spin-echo images. SNRs were calculated using equation 1.
|
Signal difference–noise ratio was calculated as-follows:
 | (3) |
The signal difference–noise ratio is an SNR that reflects the ability to differentiate objects on the basis of signal intensity. This parameter is display-independent [7].
Additionally, a consensus qualitative grading of aliasing artifacts was performed on digital images by two radiologists who were unaware of SNR measurements, using a 3-point scale: 1, major artifacts (aliasing artifacts located within the lumbar spine or the spinal canal, or both); 2, minor artifacts (aliasing artifacts located elsewhere); and 3, no artifacts. Both radiologists were unaware of the reconstruction technique used to generate the images they were scoring. Window settings were left open to all radiologists.
Statistical Evaluation
SNR and signal difference–noise ratio differences among nonparallel,
GRAPPA, and mSENSE images were analyzed with the Student's paired t
test and were considered significant when p was less than 0.05.
Interobserver reliability of this grading was determined using Cohen's kappa coefficient [8], and interpretation was performed according to the guidelines of Landis and Koch [9]. Values of 0.81–1.00 indicated excellent or perfect agreement; 0.61–0.80, substantial agreement; 0.41–0.60, moderate reliability; 0.21–0.40, fair reliability; and 0.00–0.20, poor reliability.
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
SNR values obtained with nonparallel mSENSE and GRAPPA reconstructions are represented graphically in Figure 2.
|
In the first experiment, image uniformity was 0.077 with standard T2 turbo spin-echo to 0.061 with both mSENSE and GRAPPA imaging. In the second experiment, uniformity was 0.157 on GRAPPA images and 0.014 on mSENSE images compared with 0.142 on standard T2 turbo spin-echo images. The differences in SNR values for each parallel reconstruction technique among the different phantoms are summarized in Figures 3 and 4.
|
|
The subtracted (GRAPPA – mSENSE) image illustrated a residual signal and aliasing artifacts observed with the mSENSE technique (Fig. 5A, 5B, 5C) in the first experiment.
|
|
|
In Vivo Spine MRI
Mean ROIs considered in the 15 subjects were 984 mm2 for the
background, 26 mm2 for cerebrospinal fluid, 216 mm2 for
subcutaneous fat, and 469 mm2 for the vertebral body
(Fig. 1).
SNR values obtained for the 15 subjects with T2-weighted turbo spin-echo GRAPPA and mSENSE reconstructions are summarized in Table 1. SNR values were significantly higher with the GRAPPA technique than with the mSENSE technique (p = 9 x 10–7 for cerebrospinal fluid, p = 4 x 10–7 for subcutaneous fat, and p = 1 x 10–6 for the vertebral body). GRAPPA SNR values were not significantly different from standard turbo spin-echo SNR values (p = 0.07 for cerebrospinal fluid, p = 0.15 for subcutaneous fat, and p = 0.22 for the vertebral body). SNR values for mSENSE were significantly lower than standard T2 SNR values (p = 4 x 10–7 for cerebrospinal fluid, p = 4 x 10–8 for subcutaneous fat, and p = 4 x 10–7 for the vertebral body).
|
Signal difference–noise ratio values obtained for the 15 subjects with T2-weighted turbo spinecho GRAPPA and mSENSE reconstructions are summarized in Table 2. Signal difference–noise ratio values were significantly higher with the GRAPPA technique than with the mSENSE technique (p = 1 x 10–5 for subcutaneous fat–cerebrospinal fluid, p = 6 x 10–7 for fat–vertebral body, and p = 2 x 10–6 for cerebrospinal fluid–vertebral body). GRAPPA signal difference–noise ratio values were not significantly different from standard turbo spin-echo signal difference–noise ratio values (p = 0.28 for subcutaneous fat–cerebrospinal fluid, p = 0.17 for fat–vertebral body, and p = 0.10 for cerebrospinal fluid–vertebral body). Signal difference–noise ratio values for mSENSE were significantly lower than standard T2 signal difference–noise ratio values (p = 1 x 10–6 for subcutaneous fat–cerebrospinal fluid, p = 1 x 10–8 for fat–vertebral body, and p = 3 x 10–7 for cerebrospinal fluid–vertebral body).
|
Table 3 shows aliasing artifacts observed with the mSENSE and GRAPPA techniques in the 15 subjects. Aliasing artifacts were always present on mSENSE images and were major in seven subjects. Two major and four minor aliasing artifacts were seen on GRAPPA images. No artifacts were seen on GRAPPA images of nine subjects. When aliasing artifacts were present on both mSENSE and GRAPPA images, they were more pronounced on mSENSE images. Examples of major and minor aliasing artifacts are shown in Figure 6A, 6B, 6C.
|
|
|
|
Interobserver agreement was excellent (
= 0.86) for GRAPPA images
and substantial (k = 0.77) for mSENSE images.
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
|
Unlike SENSE, SMASH is based on a modified image reconstruction in the k-space [2, 4, 13, 14] (i.e., before fast Fourier transform). SMASH exploits the sensitivity profiles of each element of a coil array. First, MRI signal data are acquired simultaneously from each individual component of the coil array. Next, the signal data for each individual component are linearly combined to produce spatial harmonics that are used to create several k-space lines simultaneously. These composite signals are then interleaved to fill a full k-space matrix covering the appropriate chosen field of view. This matrix is Fourier-transformed to produce the final image (Fig. 8). In GRAPPA [5], based on a variable density–AUTO–SMASH technique [15], several lines are acquired in the center of the k-space. The central calibration lines are acquired during the entire acquisition and are added to the whole k-space acquisition.
|
Parallel reconstruction techniques such as SENSE and SMASH offer reduced imaging time, which is strongly dependent on the number of coil array elements. The shorter imaging time results from the measurement of fewer echoes than in traditional nonparallel techniques. However, SENSE and SMASH both have inherent disadvantages, including loss of SNR.
Using parallel imaging, the SNR can be calculated as follows:
 | (4) |
Our results confirm that the SNR of parallel imaging is lower than that of standard T2-weighted turbo spin-echo imaging for a given number of coil elements [2, 13]. However, using three coil elements, the GRAPPA SNR is not significantly different from the standard T2-weighted turbo spin-echo SNR performed with only one coil. Our study thus highlights the fact that using more coil elements improves the SNR in parallel imaging.
Moreover, we found, in both phantom and subject studies, that the GRAPPA SNR was higher than the mSENSE SNR (19–36% superior for the phantom and 33% superior for the lumbar spine). Madore and Pelc [16], simulating an ideal case, found that the original SMASH technique reduced the SNR more than the SENSE technique. Two reasons for this apparent discrepancy between their results and our findings are possible. Parallel image reconstruction is sensitive to calibration, which can influence the SNR [17, 18]. In the parallel-imaging technique with autocalibration, as with the GRAPPA technique, the SNR is theoretically higher when reconstruction is performed before fast Fourier transformation. Madore and Pelc's results could have been hindered by their selection of a static phantom with good SNR, thus favoring SENSE reconstruction techniques. For instance, the SENSE reconstruction technique requires full sensitivity maps for each coil element, whereas the GRAPPA technique makes use of averaged sensitivity maps. Thus, comparing each reconstruction technique in vitro and in vivo probably provides more relevant information on the performance of both techniques than in vitro simulation alone. Additionally, Madore and Pelc studied conventional SMASH and non-modified SMASH (i.e., GRAPPA).
When studying the lumbar spine, we found that the GRAPPA SNR and signal difference–noise ratio were not significantly different from the standard T2-weighted turbo spin-echo SNR. This was unexpected, because parallel imaging usually reduces the SNR [1, 2, 13]. In nonparallel imaging, the number of coil elements is limited to the useful field of view [19] because of the aliasing risk. For spinal examination using nonparallel reconstruction techniques, a maximum of three coil elements are used, with a field of view of 380 mm. In our study, we used four coil elements for parallel imaging and three for nonparallel imaging. Because the SNR is proportional to the square root of the number of coil elements [1, 5], using more coils with parallel imaging might partially explain the increased SNR observed with GRAPPA over standard turbo spin-echo.
It has been reported that SMASH images can be reconstructed with limited artifacts [20]. However, even small artifacts can be deleterious when imaging spinal nerves or vertebral bodies. In our study, GRAPPA images bore fewer aliasing artifacts than mSENSE images. We acknowledge that our study engendered subjective perception of aliasing artifacts; therefore, major artifacts were defined only regarding their position relative to the spine because this position could be a potential limitation to the use of parallel imaging in diagnosis. However, interobserver agreement was high when assessing spine images for aliasing artifacts. Furthermore, uniformity figures observed in the phantom studies confirmed these findings, with improved uniformity after GRAPPA reconstruction procedures.
Our study has several limitations. First, we have compared only two parallel reconstruction techniques provided by a single manufacturer. Because parallel-imaging techniques are implemented differently on different platforms, our results may not be indicative of those obtained on a system from a different manufacturer or even a different software version [21]. Moreover, we acknowledge that the SNR for a particular sequence is influenced by the characteristics of the radiofrequency coil used. We probably favored the parallel reconstruction technique by using linear arrays [19, 20]. It is thus difficult to generalize our results to other coil designs. Successful image reconstruction in parallel imaging requires that the sensitivity profiles of the receiving coil array vary along the phase-encoding direction. The number of array coils in the phase-encoding direction has to be at least as large as the acceleration factor. Thus, in parallel imaging, the signal is acquired simultaneously by at least two receiver coil elements. If this requirement is not observed, artifacts and major losses in SNR in the reconstructed image will result. In spine imaging, parallel acquisition with the head–feet phase-encoding direction is optimized because the elements of the spine coil are aligned in this direction.
Because the full algorithms of each reconstruction technique are not publicly disclosed, it is difficult to fully explain the differences we have observed. Improved SNR observed with GRAPPA could be related to the autocalibration performed before fast Fourier transform. However, our study highlights the fact that on a single machine, selecting an mSENSE type or GRAPPA type of reconstruction may provide images of different quality. The purpose of this study was to compare two parallel-imaging methods. Other MRI methods are available to reduce image acquisition time by reducing the number of acquired echoes. Comparing these methods with parallel reconstruction techniques is beyond the scope of our study and warrants additional evaluation. However, one of the major inherent advantages of parallel imaging is that the time-saving process is used without altering the spatial resolution.
Differences in SNR were observed among patients for each parallel-imaging technique. Although we were unable to fully assess the influence of patient body habitus on the calibration process and the effectiveness of the parallel-imaging techniques, our in vitro results with phantoms of different size, shape, and content suggest that in GRAPPA, just as in mSENSE, this influence should be limited.
We deliberately excluded from our study patients with spinal abnormalities. Therefore, it is difficult to conclude from our study the feasibility of using these parallel techniques in diagnostic procedures. However, our aim was to compare these two techniques, both in vitro, using a dedicated phantom, and in vivo, using the most standardized and reproducible procedures. Including patients with spinal abnormalities, possibly causing discomfort during the imaging procedures, could have biased our results. Feasibility determination in diagnostic studies is thus beyond the scope of our study, but reducing scanning time without excessive SNR deterioration could be beneficial, especially in patients in whom imaging time is limited by pain or discomfort.
In conclusion, GRAPPA parallel imaging yields a higher SNR than mSENSE parallel imaging on a given imager using linear phased coil arrays. Moreover, in lumbar spine MRI with several coil elements, GRAPPA parallel imaging offers a similar SNR to traditional T2-weighted MRI and fewer aliasing artifacts than mSENSE parallel imaging.
Acknowledgments
We thank David Young for his careful reading of and useful comments on this
manuscript.
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
CiteULike 
Complore 
Connotea 
Del.icio.us 
Digg 
Reddit 
Technorati What's this?
This article has been cited by other articles:
|
|
 |
|
  Y.A. Bhagat, D.J. Emery, S. Naik, T. Yeo, and C. Beaulieu Comparison of Generalized Autocalibrating Partially Parallel Acquisitions and Modified Sensitivity Encoding for Diffusion Tensor Imaging AJNR Am. J. Neuroradiol., February 1, 2007; 28(2): 293 - 298. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
