32-Element Receiver-Coil Array for Cardiac Imaging

Abstract

A lightweight 32-element MRI receiver-coil array was designed and built for cardiac imaging. It comprises an anterior array of 21 copper rings (75 mm diameter) and a posterior array of 11 rings (107 mm diameter) that are arranged in hexagonal lattices so as to decouple nearest neighbors, and curved around the left side of the torso. Imaging experiments on phantoms and human volunteers show that it yields superior performance relative to an eight-element cardiac array as well as a 32-element whole-torso array for both traditional nonaccelerated cardiac imaging and 3D parallel imaging with acceleration factors as high as 16.

MRI has been shown to be a promising diagnostic tool for ischemic heart disease because of its ability to reveal heart-wall motion (1) and strain (2,3), myocardial perfusion (4,5) and viability (6), and, to a certain extent, coronary anatomy (7–9). Its broader adoption into routine clinical use has been hampered, however, by the requirement for repeated breath-holding or respiratory navigation, together with the need for multiple oblique imaging slices or slabs that must be localized in advance, typically through a multistep process. In short, cardiac MRI examinations to date have proved to be relatively complicated and time-consuming.

The advent of parallel imaging techniques (10–15) provided a new means of accelerating MRI without putting additional strain on the gradients. Recently, volumetric scans have been accelerated by factors of 12–16 and higher through the use of parallel imaging on scanners with 32 receivers (16,17). In situations where a narrow time window is available for imaging, such as dynamic wash-in of contrast media, these high accelerations enable the replacement of limited targeted oblique imaging slabs with larger, non-oblique volumes, which greatly simplifies the imaging paradigm. This simplified dynamic volumetric imaging paradigm was recently extended to coronary MR angiography (CMRA), enabling whole-heart coverage with relatively high resolution in a single breath-hold (18), with the use of a 32-channel torso receiver array (16).

Because high accelerations can result in substantial loss of signal-to-noise ratio (SNR), in dynamic volumetric imaging it is critical for the imaging technique to have a relatively large baseline SNR, and for measures to be taken to minimize any noise amplification caused by the parallel encoding technique. The use of 3D imaging sequences, injected contrast agents and/or sequences with high intrinsic contrast (such as steady-state free precession (SSFP)), and relatively high field strengths can all help with the former requirement; for the latter, MR coil arrays should be designed to minimize noise amplification (as characterized (11) by the geometry factor g) over the volume of interest while preserving baseline SNR. Toward that end a 32-element array was designed that is targeted for cardiac MRI (19,20).

While spatially separating coils within an MRI receiver-coil array has been shown to improve SNR for parallel imaging under some circumstances (16,21), the challenge of intercoil coupling can become a significant issue, especially for high-density arrays with many elements. The direct SNR effects of signal and noise mixing due to inductive coupling may in principle be compensated for by appropriate image reconstruction (22), but practical considerations of coil tuning and matching also come to the fore as the number of elements increases. For this reason, several recent many-element parallel-imaging arrays (23–26) have followed an overlapped design to minimize coupling between nearest neighbors. Optimal overlapping was likewise taken as a design criterion for the array under consideration here.

In any cardiac array design, the coils located on the chest and left side are of primary importance because of the heart’s center-to-left anterior position within the chest cavity (27,28). In this study we first used simulations to test the hypothesis that designing for a higher number of smaller coils in this region would produce a higher overall SNR in the heart than using a uniform coil size and number on the anterior and posterior sides of the body. Based on the results of these simulations, the array with the higher predicted performance was then constructed. The constructed array’s effectiveness in conjunction with a 32-channel receiver system was then examined in phantom studies. Finally, its use was demonstrated for highly accelerated coronary artery imaging using large volume acquisitions that facilitate whole-heart coverage in a single breath-hold.

MATERIALS AND METHODS

Simulations

Simulations were performed using the Biot-Savart equation in MATLAB® (The MathWorks, Inc, Natick, MA, USA). Dielectric effects were ignored, consistent with operation at field strengths of around 1.5 T. While quarter-wavelengths in the human torso are on the order of 15 cm at this field strength, and an argument could be made for using full-wave simulation techniques, the simulations were intended as approximate guides to be tested with experiments, and it was judged that the modeling subtleties introduced by the use of finite-difference time-domain methods (including the need to model drive ports, body geometry, etc.) could cause inaccuracies commensurate with any small errors caused by the use of quasistatic methods. Vector potentials and magnetic fields were calculated for each coil of the array and used to determine the baseline SNR, mutual inductance, noise-resistance matrix, g-factor, parallel-imaging encoding-matrix condition number (a measure of the stability of the reconstruction), and accelerated SNR over coronal planes at various depths. Arrays were designed to maximize the average SNR over a 24 × 16 × 16 cm³ (R/L × S/I × A/P) volume starting 2 cm below the chest wall. Two different array configurations, illustrated in Fig. 1, were compared. Each consisted of anterior and posterior hexagonal lattices of circular coils that were overlapped to give zero mutual inductance between nearest neighbors, with the anterior and posterior arrays separated by 26 cm. In the first (nonuniform) array (Fig. 1a), 21 coils of 75 mm diameter were aligned in four rows on the anterior side, and 11 coils of 107 mm diameter were aligned in three rows on the posterior. For the second (uniform) array (Fig. 1b), the anterior and posterior sides each had four rows of four coils, all of which were 92 mm in diameter.

FIG. 1.

Array geometry for nonuniform (a) and uniform (b) arrays.

The results of the simulation are given in Fig. 2. Here the accelerations were assumed to be fourfold in x (R/L direction) and twofold in z (S/I direction). (The choice of these target accelerations reflects the reduced encoding capability along the S/I direction arising from the lesser S/I extent of typical cardiac imaging volumes in the experimental studies described below.) Figure 2a and b show results for the nonuniform array of Fig. 1a, while Fig. 2c and d show comparable results for the uniform array of Fig. 1b. These images display coronal planes through the anterior-heart (Fig. 2a and c) and posterior-heart (Fig. 2b and d) regions. Average values over the central 24 cm × 16 cm region (demarcated by the dashed line in the top-left plot) are reported above each plot. It can be seen that the nonuniform array has a better average baseline SNR and g-factor than the uniform array in the anterior heart, and slightly worse values in the posterior heart. For a 24 × 16 × 16 cm³ (R/L × S/I × A/P) volume starting 2 cm below the anterior array, and roughly encompassing the heart, the normalized average baseline SNRs were 100 and 83.2, and the eightfold-accelerated SNRs were 29.4 and 20.6 for the nonuniform and uniform arrays, respectively. The average g-factors over the same volume were 1.50 and 1.57, respectively. On the basis of these results, the nonuniform design was chosen for construction.

FIG. 2.

Simulations of SNR, g-factor, and conditioning for the two cardiac array configurations of Fig. 1, for fourfold (x) by twofold (z) acceleration. Results for (a and b) nonuniform (see Fig. 1a) and (c and d) uniform (see Fig. 1b) density arrays, viewed over a 32-cm FOV in a coronal plane through anterior-heart (a and c) and posterior-heart (b and d) regions. Average values over the central 24 cm × 16 cm region (shown as dashed line on top-left plot) are reported above each plot.

Coil Array Construction

A custom 32-channel MRI system (26) based on GE 1.5 T Excite eight-channel scanners (GE Healthcare Technologies, Waukesha, WI, USA), and including 32 low-input-impedance GE Excite preamplifiers mounted on the end of the patient table, was available for imaging experiments. A 32-element cardiac receiver-coil array was constructed on two lightweight Lexan formers, with 21 75-mm-diameter circular rings mounted on the anterior former, and 11 107-mm-diameter rings on the posterior former. The rings were placed in a hexagonal lattice and were overlapped so as to decouple nearest neighbors, as shown in Fig. 1a (3D modeling of the array was used to determine optimal overlaps between nearest neighbors). The two formers were curved to conform to an average chest and back, and the rings were laid out to cover the left side of the torso (Fig. 3). Coils were omitted from the superior left side of the anterior array to make room for the upper arm (top right side of Fig. 1a), resulting in an asymmetric shape. The rings were cut from a 0.5-mm-thick copper sheet using a computerized water jet. They were then electropolished and grouped into three layers separated by 1-mm-thick Lexan sheets. Circular grooves cut into the Lexan sheets held the rings in place. The sheets for the anterior portion were bent around a 21-cm radius and were riveted together. Transmit blocking circuits and baluns were added to each coil. Cables were elevated off the array by 2 cm to minimize cable–coil interactions.

FIG. 3.

Position of anterior and posterior formers for cardiac array (thick lines), and coronal slices (thin lines) for g-factor measurements.

Array tuning was performed in the following manner. First, each coil was separately tuned and matched to 50 ohms when loaded on a sample torso. As the coils were brought into the anterior or posterior array, their peaks underwent broadening (by as much as 100%) and shifting (on the order of 20–50 kHz), due in part to next-nearest-neighbor coupling within the array, but more importantly to the presence of coils, matching circuits, baluns, cables, etc., within a relatively limited space. Matching and tuning capacitors were then simultaneously adjusted to restore the loading to 50 ohms while ensuring that the peak in the impedance curve was centered at the Larmor frequency. This retuning was performed on each coil in the array, one after another, with two iterations around the array.

The constructed array was tested using a Hewlett-Packard© 3577A network analyzer. Each coil was plugged into one of the ports on the analyzer, and measurements of complex coil impedance were made across a frequency range of typically ± 5 MHz centered on the Larmor frequency. Loading factors (measured by taking the ratio of unloaded to loaded impedance) averaged roughly 4 for the posterior array and 2 for the anterior array. Coupling tests were performed in the unloaded state on nearest and next-nearest neighbors, by making two coils active at a time and measuring the ratio of peak separation to peak frequency for the split resonance. For the posterior array, nearest-neighbor coupling was no higher than 0.1% for R/L neighbors and 0.5% for S/I neighbors, with coupling of 2.7% for next-nearest neighbors. For the anterior array, nearest-neighbor coupling was 0.3% for R/L neighbors and 0.9% for S/I neighbors, with coupling of 1.6% for next-nearest neighbors. Coils were also analyzed using a Hewlett-Packard© 8560E spectrum analyzer and separate transmit and receive probes inductively coupled to the coil. The frequency spectrum of each coil was monitored as a DC voltage was applied to its RF-transmit-blocking diode to ensure that the coil was adequately deactivated during transmit. Total weights were 1.6 kg for the anterior array and 2.0 kg for the posterior array. The posterior and anterior arrays are displayed in construction in Fig. 4, and on a volunteer in Fig. 5.

FIG. 4.

Posterior (a) and anterior (b) portions of the array, under construction.

FIG. 5.

The 32-element cardiac array on a volunteer.

Phantom Experiments

A 52 × 41 × 30 cm³ elliptical loading phantom (Fig. 6a) containing CuSO₄ was titrated with NaCl until it loaded the coils by the same amount as a sample human torso. The 32-element cardiac array was positioned around the phantom, and a coronal multislice 2D gradient-echo data set was acquired (FOV = 44 cm, slice thickness = 3 mm, TR = 34 ms, TE = 3.4 ms, flip angle = 30°, matrix = 256 × 192) to illustrate the sensitivity patterns of the individual coils in the array. Nonaccelerated imaging of the phantom was then performed using the same pulse sequence in an axial orientation (FOV = 40 cm, slice thickness = 3 mm, TR = 34 ms, TE = 3.5 ms, flip angle = 30°, matrix = 256 × 192) to produce a baseline SNR comparison between the 32-element cardiac array and two other arrays. One of these was a 32-element torso array (26) comprising two clamshell formers, each of which had a hexagonal lattice of 16 active optimally overlapped 10.6-cm-diameter octagonal coil elements. The other was an eight-element product cardiac array (anterior and posterior 2 × 2 grids, each covering 49 cm × 42 cm; part #100391, USA Instruments, Aurora, OH, USA). The center of each array was aligned with the S-I center position of the phantom. Six equally spaced axial slices were imaged, spanning a range of 21 cm in the S/I direction. Thirty identical data sets were acquired from each array. For each slice the 30 images were averaged together to form a “signal” image, and their standard deviation (SD) was calculated on a pixel-by-pixel basis to produce a “noise” image. The ratio of these two images was then taken to create an SNR map for each of the three arrays. To smooth noise in the maps, a median filter was applied with a neighborhood size of 5 pixels in both dimensions. The SNR map for the 32-coil cardiac array was then divided by the map from each of the other two arrays to gauge relative performance. The background regions were masked out, including the annular region within the phantom, between the inner cylinder and outer ellipse. Average values for the relative SNR were then calculated over volumes corresponding to anterior, mid, and posterior regions of the heart.

FIG. 6.

a: Elliptical loading phantom with the 32-element cardiac array. b and c: Coronal slices showing sensitivity patterns of the (b) 11 individual coils in the posterior portion of array, and (c) 21 coils in the anterior portion. Images bordered in gray are sum-of-squares composites from each set of coils.

In Vivo Experiments

A fat-saturated, ECG-gated 3D-SSFP pulse sequence was customized to synchronize the prospectively ECG-gated data acquisition for all 32 channels. CMRA was conducted in four volunteer sessions using normal volunteers. 3D SSFP was performed with FOV = 41 cm, data matrix = 256 × 256, TE = 1.9 ms, and TR = 3.7 ms. Data acquisition was completed in a single heartbeat for each acquired slice partition. Net acceleration factors of 8 (4 × 2, slice thickness = 2 mm) to 16 (4 × 4, slice thickness = 1 mm) were applied for 2D accelerations distributed along both phase-encoding directions. For single breath-held whole-heart coverage CMRA, large 3D axial slabs consisting of up to 120 slice partitions covering an S/I range of 12 cm were acquired (interpolated voxel size of 0.8 × 0.8 × 1.0 mm³). Images were reconstructed using the generalized encoding matrix (GEM) approach (29). For comparison the traditional targeted slab method was also applied (30). Experimental g-factor maps were calculated using measured coil sensitivities and receiver noise matrix (11).

RESULTS

The sensitivity patterns from the 11 individual coils comprising the posterior portion and from the 21 coils comprising the anterior portion of the 32-element cardiac array are shown in Fig. 6b and c, respectively. For comparison, a composite image of the same coronal slice is shown (outlined in gray) in the figures to illustrate phantom geometry. Figure 7a shows an axial image obtained from a slice near the S/I center of the 32-element cardiac array. The corresponding SNR map is illustrated in Fig. 7b. Figure 7c and d show the ratio of this SNR map with SNR maps calculated for the eight-element cardiac array and 32-element torso array, respectively. The ratio of SNRs in Fig. 7c varied from 1.06 in the center of the phantom to 20 near the left anterior surface (top right on the image). In a region near the far right surface of the phantom the ratio dropped to a minimum of 0.25. As shown in Fig. 7d, the ratio of SNRs ranged from 0.87 in the center of the phantom to 18 on the anterior left surface, and dropped to a minimum of 0.11 on the far right side. The average relative SNR for the anterior, middle, and posterior heart regions (outlined by the red rectangles in Fig. 7a, and extending 21 cm in the S/I direction) are given in Table 1. The 32-element cardiac array enabled 2D parallel imaging using net acceleration factors ranging from R = 8–16, which in turn provided imaging volumes covering the entire heart in a single breath-hold. Reformats of an eightfold-accelerated 3D axial cardiac data set (fourfold × twofold acceleration along the two phase-encode directions) acquired in 30 heartbeats are shown in Fig. 8, through the short axis (Fig. 8a) and four-chamber long axis (Fig. 8c), with corresponding measured g-factor maps in Fig. 8b and d, respectively. The minimum, mean, and maximum g-factors were found to be 1.1, 1.7, and 4.9 (Fig. 8b), and 1.1, 1.7, and 4.4 (Fig. 8d). For reformatted coronal planes through the anterior, mid, and posterior heart (Fig. 3), the minimum, mean, and maximum) g-factors were 1.1, 1.9, and 4.3; 1.1, 1.6, and 3.2; and 1.1, 1.7, and 4.0, respectively. These three coronal planes were at distances from the anterior array of roughly 6 cm, 9 cm, and 12 cm, respectively, with distances from the posterior array of roughly 19 cm, 16 cm, and 13 cm, respectively.

FIG. 7.

a: One of 30 identical axial images of the loading phantom, acquired with the 32-element cardiac array. b: SNR image generated by dividing the pixel-by-pixel average of the 30 images by their SD. c and d: Maps showing the ratio of the SNR image of (b) with those generated from (c) the eight-coil cardiac array and (d) the 32-coil torso array. Red rectangles in a indicate regions corresponding to anterior, middle, and posterior heart, for average SNR comparisons.

Table 1.

Ratio of SNR of 32-Element Cardiac Array Relative to Two Other Arrays, Measured in a Loading Phantom, in Three Regions Corresponding to Locations in the Heart^*

Average relative SNR	Anterior heart region	Middle heart region	Posterior heart region
Compared to 8-element cardiac array	2.04	1.47	1.14
Compared to 32-element torso array	1.41	1.07	0.99

^*Anterior, middle, and posterior heart regions are defined as outlined by the red rectangles in Fig. 7a, and extending 21-cm in S/I direction.

FIG. 8.

Reformatted images (a and c) and corresponding g maps (b and d) of eightfold-accelerated 3D axial data set from the array.

Figure 9 shows coronary artery images reformatted from eightfold-accelerated 3D SSFP data sets acquired in 30 heartbeats. The origin and proximal and more-distal segments of the LAD and RCA are clearly visible for the highly accelerated large-volume approach (Fig. 9a and d). The corresponding g-factor maps (Fig. 9b and e) indicate relatively low noise amplification overall. The image quality derived from a nonaccelerated conventional restricted targeted slab acquired in 24 heartbeats is shown for comparison (Fig. 9c and f).

FIG. 9.

Reformatted MR angiograms of the left main artery, LAD (a), and RCA (d) using eightfold (4 × 2) accelerated parallel MRI of the whole heart, with corresponding g maps (b and e). MR angiograms acquired with the conventional thin targeted slab approach (c and f) are shown for comparison.

DISCUSSION

We designed and built a lightweight 32-element MRI receiver-coil array for cardiac imaging. It comprises an anterior array of 21 copper rings (75 mm diameter) and a posterior array of 11 rings (107 mm diameter) that are arranged in hexagonal lattices so as to decouple nearest neighbors, and are curved around the left side of the torso. The use of electropolished rings mounted in thin Lexan sheets is a departure from our previous practice of using etched printed circuit boards (16,26). Imaging experiments on loading phantoms show that it yields substantially improved baseline SNR in the heart region relative to both an eight-element cardiac array and a 32-element torso array. CMRA with acceleration factors as high as 8 showed comparable image quality relative to targeted slab imaging.

A higher number of smaller array elements were allocated to the anterior surface of the torso because of the anterior position of the heart in the chest cavity. The larger posterior array elements allowed roughly the same surface coverage and sensitivity at depth with fewer elements. Simulations predicted improved baseline SNRs and g-factors in the heart with this arrangement relative to a uniform array. In principle, one might expect higher 2D accelerations to be possible near the front of the heart than the back because it is closer to the anterior array, and because of the larger number of coils on that array. We found, however, that experimental g-factors were similar in the posterior, mid, and anterior regions of the heart. For coronal planes through the mid to posterior heart, the average measured g-factors (1.6–1.7) were similar to the predicted g-factors (1.5). In the anterior heart, however, the measured values (1.9) varied somewhat from the predicted ones (~1.3). This may reflect the fact that the simulations assumed that the arrays were laid out flat on the anterior and posterior surfaces of the torso. For the imaging volumes and slice thickness used in this CMRA study, the design of the 32-element cardiac-optimized array can alleviate noise amplification to some extent for acceleration factors up to R = 12, but electrodynamic constraints dictate that a fairly rapid degeneration of SNR at accelerations larger than R = 12 is inevitable.

The feasibility of single breath-hold whole-heart CMRA using a cardiac-optimized 32-coil array has been demonstrated. The highly accelerated whole-heart coverage paradigm presented here promises to extend the capabilities of breath-hold CMRA from multiple targeted slabs to single large volumes. This supports the visualization of tortuous segments of the coronaries and reduces the demand for precise localization. Despite the clear SNR advantage of the cardiac-optimized 32-coil array, however, SNR remains a challenge because very large accelerations together with thinner slice partitions for increased spatial resolution are explored. The use of higher magnetic field strengths would improve the baseline SNR. They also promise to reduce noise amplification in parallel imaging and thus to mitigate SNR limitations in highly accelerated cardiovascular applications.

Finally, the application of highly accelerated parallel acquisitions to single-breath-hold whole-heart imaging promises to benefit not only CMRA, but also a range of other cardiovascular applications, including assessment of cardiac and large-vessel anatomy, and myocardial perfusion and viability. Moreover, this new paradigm offers the prospect of improved operator convenience and patient comfort in all of these areas.