Double‐tuned 31P/1H human head array with high performance at both frequencies for spectroscopic imaging at 9.4T

To develop a robust design of a human head double‐tuned 31P/1H array, which provides good performance at both 31P and 1H frequencies for MR spectroscopic imaging at 9.4T.


| INTRODUCTION
X-Nuclei MRI and MR spectroscopic imaging (MRSI) of the human brain provides valuable information about metabolic changes in many pathologies and is sensitive to detect abnormalities at an early disease stage. 1 However, imaging other than hydrogen nuclei, i.e., X-nuclei, such as 31 P, 13 C, or 23 Na, is often difficult due to a lower gyromagnetic ratio and, thus, a lower signal-to-noise ratio (SNR). In addition, the natural abundance of some X-nuclei (e.g., 13 C, 2 H) is in the low percentage range, which further decreases the detected signal. Therefore, an enhancement of SNR with an increase of the magnetic field strength, B 0 , one of the major advantages of ultra-high field (UHF, ≥7T) MRI, is very critical for X-nuclei imaging.
To provide high-resolution 1 H anatomical human head imaging and B 0 shimming, the radiofrequency (RF) coil must be double-tuned (DT). The latter implies that the same coil is capable of resonating at 2 substantially different frequencies, i.e., 1 H and X-nuclei. It is rather difficult to optimize the DT RF coil at both frequencies at the same time. Therefore, often the coil performance at the X-nuclei frequency is optimized while the 1 H-performance is not. 2,3 Good performance of the DT array at 1 H frequency, however, still is important for many applications. In fact, widespread translation of X-nuclei imaging into neuroscientific, physiological, and clinical studies is currently hindered by the need of changing the RF coil between X-nuclei data acquisition and conventional 1 H based study protocols to provide higher transmit (Tx) and receive (Rx) performance and better coverage for 1 H MRI. The ability to use the same RF coil for comprehensive anatomical, functional, and metabolic scan protocols, including 1 H MRI and X-nuclei MRI, without the necessity to move the subject out of the magnet and replace a DT coil with a single-tuned (ST) 1 H coil is an important step towards the establishment of X-nuclei imaging.
For a human head ST array, a nested double-layer combination of a Transmit-only array (or a volume coil) with a Receive-only array, i.e., a ToRo setup, [4][5][6][7] provides the most flexible design, which allows optimizing the Tx and Rx performance separately. In this design, a multi-element tight-fit Rx array is placed inside of a larger local Tx array or body transmit coil. At the same time, the ToRo-coil design requires additional electronics for detuning both Tx and Rx arrays, [4][5][6][7] which makes the design more complicated. Alternatively, the same elements of an array can be used during both transmission and reception. This design, a so-called transceiver (TxRx) array coil, 3,[8][9][10] is more simple to construct than a ToRo-setup but more difficult to optimize simultaneously for both transmission and reception. Also, any TxRx-design suggests a presence of additional Tx/Rx switches, which, however, do not have to be built into the RF coil itself.
There are several previously reported DT UHF head array coil designs. 3,[11][12][13][14][15][16][17] To the best of our knowledge, a 4-layer DT array coil consisting of 2 double-layer ToRo-setups (both X-nuclei and 1 H) has never yet been reported due to its high complexity. Commonly, DT human head array coils consist of 2 3,12,13,16,17 or 3 layers. 11,14,15 The double-layer design  contains 2 TxRx volume coils or TxRx arrays. A more complicated 3-layer design usually consists of an X-nuclei ToRo-setup  and either a local 1 H TxRx volume coil 11,15 or TxRx array. 14 This design mostly aims to optimize SNR at the lower X-nuclei frequency by increasing the number of elements (commonly surface loop) in the tight-fit X-nuclei Rx array. The efficiency of the X-nuclei Tx array (or a volume coil), and especially performance of the 1 H TxRx coil, which both made larger to fit the X-nuclei high loop count Rx array inside, are not optimized.
It is also noteworthy that often increasing the number of smaller surface loops in a human head Rx array improves mostly the peripheral SNR, while the central SNR does not substantially increase, [18][19][20] or is even impaired due to insufficient loading. High peripheral SNR can be harmful for MRSI due to contamination of near-cortical voxels by strong fat ( 1 H) 21 or muscle ( 31 P) 22,23 signals. Low central SNR is a major limiting factor, because for most applications the metabolic information from either the whole brain or a combination of regional cortical and deeper structures is required. In addition, the strong signal intensity gradient from the periphery to the center of the brain makes it difficult to obtain quantitative X-nuclei images. A respective signal intensity correction is hard to achieve for X-nuclei MRI and MRSI due to the lack of a suitable reference signal. Summarizing all the above, a major idea for a new DT array design, which performs well at both frequencies and has a sufficiently low difference between central and peripheral SNR, includes limiting the number of elements in both 1 H and X-nuclei arrays to what is necessary not to compromise central SNR and Tx-performance. Thus, a double-layer DT coil consisting of only 2 TxRx arrays, which have a relatively low count of elements at each frequency, can offer a simple and robust design for X-nuclei MRI and MRSI studies.
In this work, we developed and constructed a novel 20-element DT 31 P/ 1 H 9.4T (399.72 MHz -1 H, 161.8 MHz -31 P) human head array, which provides for good coverage of the human brain, high central SNR, and efficient Tx-performance at both frequencies. The array consists of 16 TxRx surface loops (8 loops at each frequency) circumscribing the head. All 16 loops are placed on the surface of the same tight-fit array holder to make sure that both 31 P and 1 H surface loops are located at the same short distance from the sample. In addition, the array has 2 pairs ( 31 P and 1 H) of "vertical" loops placed at the superior location of the head. We performed a comprehensive array evaluation and comparison to the performance of a previously described 8-loop single-row (1 × 8) ST 1 H TxRx array of similar geometry 24 and 16-loop double-row (2 × 8) ST 1 H TxRx array. 10 We also compared the array 31 P SNR to that of the commercially available 7T 3-layer 31 P/ 1 H array head coil.

| Phased array design
As mentioned above, increasing the number of smaller surface loops in a human head Rx array can be harmful for X-nuclei MRI and MRSI. In addition, smaller loops can compromise the central SNR due to insufficient loading and, therefore, higher contribution from intrinsic coil losses. 25 To minimize these effects, we limited the number of loops in both the 1 H and 31 P portions of the DT array to 10 each. This relatively small loop count helps to substantially reduce SNR at the periphery of the head while keeping the central SNR high. [18][19][20] Figure 1 shows the electromagnetic (EM) simulation models of the 10-element 1 H and 31 P arrays loaded by phantoms. Larger sized 31 P loops help to maintain higher loading, which is very critical at the lower 31 P-frequency. Surface and vertical loops are constructed using 1.5-mm copper annealed wires. The choice of 1.5-mm wire is mostly determined by the mechanical stability and convenience in constructing the loops.
The 1 H array ( Figure 1A) consists of eight 11-cm long overlapped TxRx surface loops circumscribing the head and 2 TxRx "vertical" cross-loops placed at the superior location. Use of a single-row 8-loop array seems to be sufficient not to compromise central SNR [18][19][20] as well as B + 1 homogeneity. 3,24 Two 1 H vertical loops are slightly different in size and measured 90 mm × 40 mm and 110 × 40 mm. The 31 P array ( Figure 1B) also consists of 10 elements, i.e., eight 17-cm long TxRx surface loops gapped by 10 mm and circumscribing the head. In addition, the 31 P array had 2 Rx-only "vertical" loops at the superior location. Two 31 P Rx vertical cross-loops are larger than the corresponding 1 H loops and measure 110 mm × 50 mm and 130 × 50 mm. Two pairs of vertical loops ( Figure 1) have coplanar geometry with 1 H loops placed symmetrically inside corresponding 31 P loops.
The choice of the vertical loop geometry is relatively simple. An increase of loops' size improves the penetration depth and coverage at the superior location of the head. Therefore, our selection procedure consists of an increase of the loop length and width within the limitation of the current mechanical design, which is determined by the size and geometry of the array holder, position of RF electronics inside the array (e.g., Tx/Rx switches), etc. All 16 TxRx surface loops ( 1 H and 31 P) are placed on the surface of the same array holder ( Figures 1C and 2A), which provides a similar tight fit for both 31 P and 1 H arrays. In addition, to minimize interaction between adjacent 31 P and 1 H surface loops, 1 H loops are shifted by half a loop size ( Figure 1C). The 1 H array is decoupled entirely by overlapping the adjacent loops. 24 The 31 P array is decoupled by transformer decoupling. 3 variable capacitors, and is intrinsically broadband. To decrease radiation losses 26 and minimize coupling between nonadjacent loops, 27 the array is shielded with the cylindrical shield placed at a 40-mm distance from the surface loops. The distance to the shield is chosen based on previously published data for UHF human head arrays. 3,10,24

| EM simulations
EM simulations of the transmit B + 1 , where B + 1 is the circularly polarized (CP) component of the RF magnetic field, and the local specific absorption rate (SAR) were performed using CST Studio Suite 2017 (CST, Darmstadt, Germany) and the time-domain solver based on the finiteintegration technique. The solver was stopped when the total energy in the system fell below −50 dB of the maximum monitored system energy. We also used approximately 40 million mesh cells with the smallest cell size of 0.8 mm.
The array was modeled following the array geometry described above (Figure 1), and consisted of 20 (16 surface and 4 vertical) loops. Ten and 8 fixed capacitors were distributed over the 1 H loop length for surface and vertical loops, respectively. Both vertical Rx and surface TxRx 31 P loops had 6 distributed capacitors. To account for additional losses, all fixed capacitors were modeled with equivalent resistors placed in series. Resistor values were obtained from component datasheets. Tuning and matching capacitors were substituted by ports and adjusted during co-simulations. 1 H surface loops were decoupled by overlapping. Overlap between adjacent loops was adjusted individually by changing the loop width. After adjustment, overlapping measured ~16 mm. For evaluation of the B + 1 and SAR distribution maps at 31 P frequency (162 MHz), 2 vertical Rx loops were actively detuned. As demonstrated previously, a presence of detuned Rx loops may alter the maximum local SAR value. 28 Three voxel models were used in simulations, i.e., the head and shoulder (HS) phantom ( Figure 1A), which was constructed to match tissue properties (ε = 58.6, σ = 0.64 S/m) at 400 MHz, 7,29 and 2 virtual family multi-tissue models, "Duke" and "Ella", 30 cropped at the level of the chest. For all 3 voxel models, we used a 2-mm resolution. For evaluation of the transmit performance of the 31 P array, we used an elliptical phantom (length, 17 cm; axes of ellipse, 18 cm × 15 cm; ε = 62.4; σ = 0.54 S/m at 160 MHz) shown in Figure 1B. B + 1 field profiles and local SAR 10g (averaged over 10 g of tissue) maps were calculated for 1 W of total (per 8 channels) power at the coil input and compared with experimentally measured data. Averaging of SAR was performed using the CST Legacy method.
Numerical optimization of the size of vertical loops within the entire design (including all surface loops) is a lengthy F I G U R E 2 A, Photo of the DT 31 P/ 1 H array with the cover removed. B, Photo of the superior part of the array with the cover removed to display "vertical" cross-loops. C, Photo of the array connected to the splitter. Tx and Rx ODU connectors are shown by arrows. D, Photo of the DT phased array loaded by the HS phantom process since the total model is rather large. However, since coupling between vertical and surface loops is sufficiently low, we simulated separately a pair of 1 H vertical loops with 20-mm larger width (i.e., width of 60 mm) to evaluate an increase in the B 1 field. As a result, larger loops generated a 7% higher magnetic field at the depth of 20 mm from the top of the head. There was practically no difference near the center of the head.

| Array Construction
After EM modeling, we constructed the 20-loop DT 31 P/ 1 H array (Figure 2A,B). Figure 2C,D show the assembled array connected to the splitter and loaded by the HS phantom. The geometry of the array holder and the loops is the same as in EM modeling ( Figure 1). The size of the fiberglass array holder measures 230 mm in height and 200 mm in width. The array holder is also tapered to improve fitting to a human head and measures 155 mm in width (left-right) and 185 mm in height (anterior-posterior) as shown in Figures 1C and 2A. Eight 1 H surface loops circumscribing the head are decoupled by overlapping adjacent loops without any additional decoupling circuits. As shown previously, this method provides for very good decoupling and substantially simplifies the entire array design. 24 During the construction procedure, we started with the overlap distances used in EM simulations, i.e., ~16 mm, which was then adjusted manually to minimize coupling. Figure 3A shows the schematic of a single 1 H surface loop including the tuning and matching capacitors, DT cable trap, 3,11 and home-built Tx/Rx switch circuit (−0.25 dB of insertion loss, −40 dB of isolation). 31 Each surface loop has 12 fixed chip capacitors (100C series, American Technical Ceramics, Huntington Station, NY) distributed along the loop. Fixed capacitor values range from 5.6 pF to 6.8 pF, except for two 3.3-pF capacitors, which are connected in parallel to tuning and matching variable capacitors ( Figure 3A). Tx/Rx switches are connected to each loop and located inside the array holder to minimize losses. A low-noise preamplifier (WanTcom, Chanhassen, MN) is incorporated into each Tx/Rx switch circuit. All surface loops are individually tuned and matched using variable capacitors (Johanson Corp., Boonton, NJ). To prevent wave propagation along the cable, a shielded DT cable trap is introduced at the input of each surface loop. Figure 3B shows the schematic of a single 1 H TxRx vertical loop, which has 10 chip capacitors distributed along the loops and ranging from 6.2 pF to 8.2 pF. Each 1 H surface loop had a single 31 P trap constructed by connecting an inductor in parallel to one of the distributed capacitors ( Figure 3A,B). Each 31 P trap is individually tuned to minimize coupling between neighboring 1 H and 31 P loops at 31 P frequency (161.8 MHz). Figure 3C shows the schematic of a single 31 P TxRx surface loop, including the tuning and matching capacitors, DT cable trap, and home-built Tx/Rx switch circuit. Each loop has 8 fixed chip capacitors (100C series, American Technical Ceramics, Huntington Station, NY) distributed along the loop. Fixed capacitor values range from 12 pF to 15 pF, except for two 8.2 pF capacitors, which are connected in parallel to variable capacitors ( Figure 3C). Adjacent TxRx loops are decoupled using the transformer decoupling method. Since the coupling between closely located 17-cm long 31 P surface loops is quite large, 2 small transformers are placed between each pair of adjacent loops (Figures 1C and 2A). Each transformer is constructed using a 6-mm plastic cylinder and has 3 turns in both the primary and secondary windings. The schematic of a single vertical 31 P Rx loop is shown in Figure 3D. Both loops have 6 distributed capacitors ranging from 18 pF to 24 pF.
During transmission, 31 P vertical loops are actively detuned. Active detuning circuitry is constructed using 2 PIN diodes (MACOM, Lowell, MA) connected in series ( Figure 3D). Commonly active detuning circuits are constructed using 1 PIN diode. 7 The second PIN diode in our design provides additional blocking of 6 dB or more. PIN diodes are driven using MRI system direct current drivers, which provide 100 mA of current. Direct current voltage is delivered to the PIN diodes through the RF cable using an RF choke (RFC, Figure 3D). As a secondary safety feature, each 31 P vertical loop has a protective fuse. In addition, all 31 P Rx and TxRx loops have 2 1 H traps constructed by connecting an inductor in parallel to a 7.5 pF capacitor ( Figure 3C,D). Placement of 2 1 H traps increases Q U by ~20% and further improves decoupling of the 1 H surface loops. Each 1 H trap is individually tuned to maximize its impedance at 1 H frequency (399.72 MHz).
During transmission, both 1 H and 31 P arrays are driven in the quadrature CP mode, i.e., with a 45° phase shift between adjacent loops. For this purpose, we constructed 2 Wilkinson splitters both placed into the same box ( Figure 2D), i.e., one 9-way 1 H splitter and one 8-way 31 P splitter. Two 1 H TxRx vertical cross-loops ( Figure 1A) are driven in quadrature (90° phase shift between the loops) using an additional home-built 90° hybrid splitter. Also, to align the vertical loops' B 1 near the center of the array with that of corresponding surface loops, superior 1 H cross-loops are driven in phase with surface loops 1 and 7 (Figure 4). To simplify connection of the array to the splitters, all RF cables are combined into 2 bundles ( Figure 2C, ODU Tx), each with a modular connector (ODU GMBH, Muehldorf, Germany).

| Performance and safety evaluation
Before doing in vivo measurements, the phased array was evaluated on a bench and in the scanner and numerically simulated according to the safety procedure developed in our lab. 32 The human subjects participated in the study after | 1081 AVDIEVICH Et Al.
giving signed informed consent according to procedures approved by the local institutional review board committee.
Bench evaluation of the array includes measurements of the loaded and unloaded Q-factors (Q L and Q U ) and decoupling between each pair of the surface loops. Q-factors were measured from the absolute value of the S 11 reflection coefficient as twice the ratio of the resonance frequency over the 3-dB bandwidth 33 using a network analyzer (E5071C, Agilent Technology, Santa Clara, CA). Decoupling between TxRx loops (Tx-mode) was evaluated by measuring the S 12 transmission coefficient with 31 P Rx-only vertical loops actively detuned. In all these measurements, we used the HS phantom.
All data were acquired on a Siemens Magnetom (Erlangen, Germany) 9.4T human imaging system. 1  (a) and preamplifier (b). Active detuning circuit consists of capacitor C 1 , inductor L 1 and 2 PIN diodes D 1 , D 2 . C bl are 300-pF direct current blocking capacitors a corresponding noise scan, which was used to evaluate a noise correlation matrix. 35 SNR was then calculated as an optimal weighted root sum-of-squares combination taking into account a noise correlation matrix. 36 SNR maps were also corrected for the spatial flip angle variations using actual flip angle imaging B + 1 maps. Experimental 1 H B + 1 , and SNR maps were obtained using the HS phantom as well as in vivo. Experimental 1 H B + 1 maps were also compared to maps obtained using the 1 × 8 24 and 2 × 8 10 ST arrays constructed previously. The 1 × 8 array has very similar size and geometry to the 1 H portion of the DT array while the 2 × 8 array is longer and measures 17.5 cm. Both ST arrays lack a pair of TxRx cross-loops at the superior location of the head. 31 P B + 1 and SNR maps were evaluated using MRSI phantom data. For the SNR measurement, we used an elliptical phantom with a rounded top (length, 18 cm; axes of ellipse, 19 cm × 15 cm; ε = 62.4; σ = 0.54 S/m at 160 MHz). SNR maps were acquired with a 3D MRSI pulse sequence 37 (FOV, 220 × 220 × 220 mm 3 ; voxel size, 6.9 × 6.9 × 6.9 mm 3 ; elliptical k-space; TR = 215 ms; 1 average; flip angle, 20°; rectangular excitation pulse with 0.5 ms pulse duration, 5 kHz acquisition bandwidth; vector size, 1024; acquisition time, 61 min). A 31 P B + 1 map was acquired with a phase-sensitive sequence with spiral readout 14,38 (FOV, 250 × 250 × 240 mm 3 ; TR = 4 s; number of averages, 2; resolution, 4 × 4 × 5 mm 3 ; acquisition time, 3 h). 31 P SNR was calculated for each individual voxel as a ratio of the peak integral over the standard deviation of the noise in the spectrum.

| RESULTS
First, we evaluated the array on a bench, which included measurements of Q-factors and the S 12 matrix between TxRx loops. The ratio of Q U /Q L of the surface loops in the 1 H array loaded by the HS phantom measures from ~9 (anterior) to 6 (posterior). Q U measures ~240. Q U /Q L of the 1 H vertical TxRx loops was lower and measures ~3.5. Q U measures ~340. The ratio of Q U /Q L of the surface loops in the 31 P array loaded by the HS phantom measures from ~4.5 (anterior) to 3.5 (posterior). Q U /Q L of the 31 P vertical Rx loops measures ~2. Q U of all 31 P loops measures ~280. Figure 4 shows S 12 matrices measured for both the 1 H and 31 P arrays loaded by the HS phantom. The 1 H array is well decoupled by overlapping the loops with the strongest coupling (less than −18 dB) measured between closest nonadjacent loops, i.e., 1 and 3, 2 and 4, etc.
As we demonstrated previously, at 9.4T (400 MHz) overlapping provides an excellent way of decoupling surface loops for the 1 × 8 tight-fit array due to diminishing of the resistive coupling. 24 Decoupling of the 31 P array was worse due to lower loading but still sufficiently good.

F I G U R E 4 8 × 8 S 12 matrices measured for 1 H (A) and 31 P (B)
TxRx surface loops of the DT array loaded by the HS phantom. For S 12 matrix measurements, 31  All adjacent TxRx loops are decoupled better than −18 dB with an average S 12 value of −19.2 dB. Closest nonadjacent loops (i.e., 1 and 3, 2 and 4, etc) show the strongest coupling, which ranged from −13 dB to −16 dB with an average S 12 value of −14.6 dB. Isolation between 1 H and 31 P neighboring loops measured −30 dB or better at both frequencies.
Together with the bench evaluation, we performed the safety evaluation of the new DT array according to the procedure developed in our institution. 32 As an example, Figure 5 shows the B + 1 and SAR 10g distributions obtained for the 1 H and 31 P arrays both loaded by the Duke voxel model. Maximum SAR 10g values were calculated for the CP mode at both frequencies, and various loading conditions mimicking variation in the head size. In these simulations, we first evaluated the maximum local SAR 10g for perfectly tuned and matched 31 P and 1 H arrays for both voxel models. In addition, we calculated the maximum SAR 10g for each model using the array tuned and matched on the other model. For example, an array adjusted on the Duke model was loaded by the Ella model and vice versa. At the end, the maximum local SAR 10g value at each frequency was chosen from 4 values. Final maximum SAR 10g values measured 0.76 W/kg and 0.64 W/kg at 1 H and 31 P frequencies, respectively. In both cases, worse maximum SAR 10g values were obtained for the Ella voxel model when tuning and matching were performed on the Duke model. We also calculated a change in the maximum SAR of the 1 H array due to presence of the vertical loops. The maximum local SAR increased by 8% and 2% for Duke and Ella models, respectively.
In the next step, we evaluated Tx and Rx performance of the DT array loaded by phantoms described above. Figure 6 shows simulated and experimentally measured 1 H and 31 P B + 1 maps. Simulated and experimental data match each other well. For both arrays, simulated B + 1 values are ~10% higher than experimentally measured, which commonly happens due to difficulties taking into account losses in all components of the RF coil. 40 This, however, only leads to a small overestimation of the maximum local SAR and does not cause any safety issues. The experimentally measured 31 P B + 1 value averaged over the entire phantom is 19.4 ± 4.2 µT/√kW. The B + 1 value near the center is 30.3 µT/√kW. Figure 7 shows SNR maps obtained using the 1 H and 31 P arrays loaded by phantoms. Figure 7B,C compare 31 P SNR obtained with and without Rx vertical cross-loops. The Supporting Information Figure S1A, which is available online, shows the 31 P SNR map for the central coronal slice. It is noteworthy that the addition of vertical cross-loops improves the 1 H B + 1 distribution ( Figure 6A,B) and SNR of both arrays (Figure 7) at the superior location of the head. Transversal 31 P SNR maps ( Figure 7B) shows left/ right asymmetry. The asymmetry most likely occurred due to residual coupling between closest nonadjacent loops (e.g., 1 and 3, 2 and 4, etc), which is seen in the S 12 matrix ( Figure 4B) and is difficult to eliminate due to the distant location.
After performing coil test measurements using phantoms and evaluating coil safety, 32 we tested the array in vivo. Figure 8 shows in vivo results obtained using the new DT array at 1 H-frequency (399.72 MHz). In addition, Figure 8 also demonstrates experimental data obtained using the 1 H 2 × 8 and 1 × 8 ST arrays. Similar to the phantom data, the addition of the cross-loops substantially improves the longitudinal brain coverage as compared to the 1 × 8 ST array. It is also of importance that the Tx efficiency of the DT array is very similar to that of the ST arrays. Averaged  Figure  1B. Central sagittal B + 1 (C) and SAR 10g (D) maps obtained for the model of the 1 H array shown in Figure 1A. All B + 1 and SAR 10g maps were obtained for arrays driven in the CP mode and loaded by the Duke voxel model F I G U R E 6 A, Experimentally measured and simulated central sagittal, coronal, and transversal B + 1 maps obtained using the 1 H part of the 31 Figure 8E,F compare SNR obtained by DT and ST 1 H arrays. The DT array delivers much higher SNR at the superior location than the 1 × 8 ST array. Both arrays have very similar SNR near the center. The 2 × 8 ST array provides ~35% higher SNR than the DT array both at the center and periphery. Figure 9 demonstrates PCr 31 P MRSI images obtained using the new DT array. Shown are transversal and sagittal slices, which exhibit good coverage over the entire brain.
As an example, we show 2 voxels, 1 from the periphery and 1 from the central area of the brain, to demonstrate that we achieve a good signal in the center relative to the periphery with the constructed 31 P/ 1 H array. The central coronal slice is shown in Supporting Information Figure S1B. Finally, Figure 10 shows a comparison of the new 9.4T DT array's Rx performance to that of the commercial 7T DT array using the same human head shaped 31 P phantom. 37 In this example, we also show SNR at 2 voxels located near the center and periphery. While both arrays have very similar peripheral SNR (less than 3% difference), the 9.4T array has 3.9 times higher SNR near the center. The ratio of the peripheral to the central SNR measured 10.8 and 2.7 for the commercial array and our arrays, respectively.

| DISCUSSION
We developed, constructed, and evaluated the new DT 31 P/ 1 H (161.8 MHz/ 399.72 MHz) array design for brain studies at 9.4T. Commonly, to simplify decoupling and placement of multiple elements of a DT array, 1 H loops are moved away from the subject and positioned in a second layer. 3,12,13 Alternatively, they are made substantially smaller than X-nuclei loops and are located inside of them. [14][15][16] Both methods decrease loading of 1 H loops and, thus, compromise the Tx-performance and SNR of the DT array at 1 H frequency. 24 It is also noteworthy that, at UHF, simple increasing of the length of a single-row 1 H 1 × 8 loop array is not sufficient to provide for whole-brain coverage. 3,[7][8][9] Good longitudinal coverage of the whole brain has been demonstrated by using multi-row (≥2) arrays, e.g., 2 × 8 arrays, 3,7-9 which provide the capability of 3D RF shimming. 3,[7][8][9]40 However, the combination of a 2 × 8 1 H TxRx array with any 31 P array has never been reported due to the complexity of the design.
At the same time, the high Tx-performance and wholebrain coverage of the DT array at 1 H frequency are important for many applications. Our new DT array design provides for a solution to these issues by using 2 new features as compared to previously reported designs. First, we placed TxRx surface loops of both 31 P and 1 H parts of the array on the surface of the same holder and at the same distance to a head. Second, we added 2 pairs of 1 H vertical cross-loops at the superior location of the head. First, these modifications allowed us not to compromise the Tx-efficiency of the 1 H portion of the array as compared to the ST arrays of similar size. The F I G U R E 8 A, Central sagittal and transversal in vivo human brain GRE images obtained using the 1  new DT array showed very similar B + 1 value averaged over the 120-mm transversal slab ( Figure 8C) as compared to both the 1 × 8 and 2 × 8 ST arrays. Such a tight-fit single-layer arrangement is feasible only because of the relatively small number of elements. With a higher element count, 1 H loops have to be placed in the second layer at a larger distance to the sample, which inevitably decreases the 1 H Tx-efficiency. 3 Second, an addition of 2 pairs of 1 H vertical cross-loops provides a significant improvement of the longitudinal coverage compared to the 1 H ST 1 × 8 array. As seen from Figure 8, vertical cross-loops substantially improve both the 1 H B + 1 distribution and SNR at the superior location of the head. While the coverage (both Tx and Rx) is still worse compared with the 1 H ST 2 × 8 surface loop array, it is much better than that of the 1 H ST 1 × 8 array. As previously shown, combining a pair of cross-loops with a 2 × 8 TxRx surface loop array further improves the longitudinal coverage of the 1 H array. 41 However, such a design, i.e., the 18 TxRx loop array, would be even harder to combine with a multi-channel 31 P array within a single layer. Thus, the new design provides for a solution having a reasonable longitudinal coverage using a relatively small loop count.
Regarding performance at the 31 P frequency, as seen from Figure 7, the addition of 31 P vertical cross-loops improves SNR at the superior location of the head. Also, data shown in Figure 9 demonstrate that we achieved very good coverage over the whole brain using the new DT array with a reasonable measurement time of 30 min and a nominal voxel size of 1.75 mL and an improved 1 H performance. The representative central voxel shows good SNR compared to the peripheral location. It is also noteworthy that our low loop count DT array provides a substantially more uniform SNR distribution than that of the commercial high loop count array. As seen from Figure 10, the peripheral SNR of the 9.4T DT array is only 2.7 times higher than the central SNR. For comparison, the 32-channel commercial array shows a significantly greater (i.e., ~4 times) difference, which is harmful for MRSI experiments. It is also important that the 9.4T array demonstrates almost 4 times higher SNR near the center of the phantom while showing very similar SNR at the periphery.
According to recent theoretical 42 and experimental 43 studies, the central human head SNR has a nearly quadratic dependence on the constant magnetic field B 0 , while the peripheral SNR grows rather linearly. Therefore, such a large difference in the central SNR cannot be explained by the difference in B 0 values, i.e., 7T versus 9.4T, which provides only factor of 2 difference. The residual difference occurs due to a larger count of smaller loops in the 31 P commercial Rx array, i.e., 32 versus 20. Smaller loops are less loaded and, therefore, do not provide for the sample noise domination. 24 As a result, the central SNR is compromised as compared to that of an array with larger loops. On the other hand, because the 7T 31 P resonance frequency is very close to that of the 3T 1 H resonance frequency, a reasonable comparison can be made with previously reported 32-loop 3T Rx arrays, which demonstrate about 6 times difference between peripheral and central SNR. 18,44,45 The residual difference between central and peripheral SNR could also occur due to the presence of the 1 H coil or some differences in array geometries.