An 8-channel Tx/Rx dipole array combined with 16 Rx loops for high-resolution functional cardiac imaging at 7 T

A custom-built 8-channel Tx/24-channel Rx setup was fabricated (Fig. 1). The setup consists of 8 building blocks composed of 8 fractionated dipole antennas (300-mm length) that were used for both transmitting and receiving. In each building block, two loop coils (160/100 mm, long/short axis) were positioned between the dipole and the subject along the longitudinal direction with overlap for decoupling. These 16 loop coils were used for receiving only [9, 11]. The dimensions of the loops and the dipoles are based on simulations on ideal loop and dipole size, which were done in references [9, 25].

The loop coils were kept at a 6-mm spacing from the body and the dipole was kept at a 20-mm distance from the body by a polycarbonate housing of each building block. The two medial anterior elements were bent in the middle and positioned at a fixed angle to maintain full contact with the chest. Each loop coil was detuned at three positions by PIN diodes. Preamplifier decoupling ensured high impedance at the cable connection point during reception. A lattice balun was used for impedance matching of both the dipole antennas and the loop coils. More details on geometry and circuitry are presented in Fig. 1. Reflection and coupling levels (S11 and S12) were determined to ensure good array performance as well as Qunloaded-to-Qloaded ratios for the loop coil elements.

The 8-channel Tx/24-channel Rx setup was modeled using the Sim4Life environment (Zurich Medtech, Zurich, Switzerland). Simulations were carried out on human models Duke and Ella [body mass index (BMI) 23.1 and 22 kg/m²], using a resolution of 1.5 × 1.5 × 1.5 mm³ to render the coil geometry and the nearby tissue, resulting in 13.862 × 10⁶ cells for Duke and 13.746 × 10⁶ cells for Ella [26]. Simulations were done on a graphic processor unit (GPU, NVIDIA GeForce GTX TITAN Black, NVIDIA, Santa Clara, CA, USA). [26]. The worst-case SAR was calculated as the maximum sum of the modulus of all quality matrix (Q-matrix) entries [18]. This value was used to derive average power limits for imaging applications, based on a 10-g-averaged SAR limit of 20 W/kg in the trunk for the first level controlled mode [27]. To validate the simulations, single-channel B ₁ ⁺ maps have been acquired for all transmit channels on a phantom filled with ethylene glycol and saline (ε = 34, σ = 0.4 S/m, 0.4 g/L saline) [9]. The dual refocusing angle acquisition mode (DREAM) B ₁ ⁺ mapping method was used to acquire the B ₁ ⁺ maps [28]. The imaging parameters of this sequence were as following: 2D acquisition, echo time (TE) = 1.57 ms, repetition time (TR) = 10 ms, field of view (FOV) = 320 × 400 × 30 mm³, in-plane resolution = (5 × 5) mm², slice thickness = 30 mm, flip angle = 10°, pulse width = 0.17 ms, stimulated echo (STE) angle = 60°, pulse width = 1.01 ms, nominal B ₁ ⁺  = 16 μT, receive bandwidth = 4882.8 Hz, acquisition time = 11 s. A model of the phantom was imported from SolidWorks (SolidWorks, Dassault Systèmes SolidWorks Corp., MA, USA) into Sim4Life and used for simulations of the B ₁ ⁺ fields. An isotropic resolution of 1.5 × 1.5 × 1.5 mm³ was used for the full model, resulting in a total of 22.687 × 10⁶ cells.

A 7-T Philips Achieva multi-transmit system with 8 × 2 kW RF amplifiers (Philips Healthcare, Best, The Netherlands) was used to scan 8 healthy volunteers [7 males, 1 female, age 22–35, average BMI = 21.6 ± 1.14 kg/m², minimum BMI = 20.1 kg/m² and maximum BMI = 23.1 kg/m²]. The study was approved by the local medical ethics committee and all subjects signed informed consent prior to inclusion in the study. RF phase shimming was applied on three slices in the heart, in order to obtain the maximum average signal in the heart [29, 30]. For the phase shimming, low-flip-angle gradient echo images were acquired in three slices for every transmit channel, these images were combined to obtain maximum signal intensity averaged over three slices in the heart; the optimum phase settings were calculated with a numerical minimization in Matlab (Mathworks, Natick, MA, USA) [30]. The same phase settings were used for all three slices, while the amplitudes of the channels where all set equally. This procedure was done once for every volunteer, while the same phase settings were used for all views. Phase-only RF shimming was used for every acquisition in this work. The following imaging parameters were used for the low-flip-angle gradient echo images: 2D multislice acquisition (M2D), TE = 1.68 ms, TR = 24 ms, FOV = 309 × 522 × 60 mm³, in-plane resolution = (1.3 × 1.3) mm², slice thickness = 20 mm, flip angle = 3°, turbo field echo factor (TFE-factor) = 15, receive bandwidth = 498.7 Hz, pulse width = 0.20 ms, nominal B ₁ ⁺  = 4 μT and acquisition time = 102.7 s. Subsequently, a 10-slice cine planning sequence in the transverse orientation was acquired during 5 breath-holds (R = 2). The obtained images were used for planning of cine cardiac imaging. Pseudo 2-chamber, pseudo 4-chamber, short-axis and 4-chamber (p2Ch, p4Ch, SAX and 4Ch, respectively) view images were acquired during breath-holds. The following imaging parameters were used for the cine imaging: TE = 2.7 ms, TR = 4.2 ms, FOV = 280 × 420 × 8 mm³, in-plane resolution = (1.3 × 1.3) mm², slice thickness = 8 mm, flip angle = 9°, TFE-factor) = 10, receive bandwidth = 998.2 Hz, pulse width = 0.61 ms, nominal B ₁ ⁺  = 4 μT and acquisition time = 10 heartbeats/10 s on average, with 30 cardiac phases. Retrospective gating with electrocardiographic (ECG) pads and breath-hold triggering were used for motion compensation. For one volunteer, the 4Ch view acquisition was repeated at increasing resolutions (1.3 × 1.3 × 8 mm³, 1.1 × 1.1 × 2.5 mm³ and 0.75 × 0.75 × 1.75 mm³), using a sensitivity-encoding (SENSE) acceleration factor of R2 in the anterior–posterior (AP) direction. Acquisition times increased from 10 to 12 and 17 heartbeats, respectively.

To obtain a measure of the overall image quality, the cine images of all eight volunteers were rated on a four-point scale by two experienced readers. Overall image quality, artifacts and noise where taken into account in this rating [23], where higher scores represent better image quality. The rating scale and scoring criteria are shown in detail in Table 1. Inter-observer agreement percentages and Cohen's kappa were used to calculate inter-observer variability for all ratings [31].

In order to make a quantitative comparison to other literature, the imaging parameters used in [13] have been reproduced to the best of our ability. The following imaging parameters were used for the SNR analysis: TE = 3.8 ms, TR = 6.0 ms, FOV = 280 × 420 × 2.5 mm³, in-plane resolution = (1.1 × 1.1) mm², slice thickness = 2.5 mm, flip angle = 20°, TFE-factor = 9, receive bandwidth = 998.2 Hz, pulse width = 3.6 ms, nominal B ₁ ⁺ 1.5 μT, acquisition time = 20 heartbeats/20 s on average, 28 cardiac phases, acceleration factor R = 2 and applied along the AP direction. This was done on three additional volunteers (2 males, 1 female, age 25–36, average BMI = 21.9 ± 1.12 kg/m², minimum BMI = 20.01 kg/m², maximum BMI = 23.03 kg/m²) in the 4Ch view and the SAX view. Phase shimming was done for all acquisitions on three transverse slices in the heart. To assess the SNR performance of the coil array, as well as the separate contributions of the loop and dipole elements, SNR-scaled images were reconstructed according to the method described by Robson et al. [32]. The mean SNR in the heart, and the CNR between the myocardium and the blood, defined as (SNR_blood–SNR_myo), was calculated for all three volunteers, according to [13].

To assess the acceleration performance of the array, the cine acquisitions were repeated on the same three volunteers, using sensitivity encoding (SENSE) with acceleration factors ranging from R = 2 to R = 6. Phase encoding was applied along the left–right (LR) direction for the 4Ch view images and along the feet/head (AP) direction for the SAX view images. Geometrical factor (g-factor) maps were reconstructed on the scanner using reconstruction software available on the Philips system (delayed reconstruction). The mean g-factor in the heart was calculated for all three volunteers.

The elements were tuned and matched to obtain matching and decoupling values of −12 dB on the torso. Bench measurements show a Q _unloaded-to-Q _loaded ratio of 140:11 for the loop elements. PIN diodes were used to detune the loops during RF transmission. Decoupling between the loops and the dipoles was improved from −12 to −18 dB or less after detuning of the receive loop. The geometry and circuitry of a single loop/dipole element is shown in Fig. 1. Figure 2 shows a schematic overview of the imaging setup on a volunteer.

Figure 3 shows maximum intensity projections of 10-g-averaged SAR (SAR_10g) and 1-g-averaged SAR (SAR_1g) distributions for human models Duke and Ella. Phase-only shimming was used to optimize for maximum average B ₁ ⁺ in the heart and results are normalized to an average input power of 1.0 W per channel (8.0 W total). It was demonstrated that SAR_10g does not exceed 2.0 W/kg for both models using phase-shimmed transmit phases. The worst-case SAR_10g is calculated to be 4.05 W/kg for the Ella model, and 2.96 W/kg for the Duke model, using an input power of 1.0 W per channel (8.0 W total input power). Considering a maximum allowed SAR_10g of 20 W/kg in the first level controlled mode, the maximum average power limit is 20 W/4.05 = 4.92 W per channel based on these two models. The transmit phases for which the worst-case SAR values are obtained do not correspond to the transmit phases that maximize B ₁ ⁺ in the heart, so this value represents a conservative estimate of the required power constraint. However, with only two subjects investigated, inter-subject variability may still increase the maximum SAR value. A recent study has investigated these opposing effects for prostate imaging at 7 T [18, 33]. Based on these results, and including a safety margin of 20%, the average power limit in these experiments was set to 4.0 W per channel. The maximum worst-case SAR_1g that is calculated in simulations is 4.85 W/kg for a total input power of 1.0 W per channel. If the average power limit of 4.0 W/channel is applied, the local SAR limits are also not exceeded when using this small averaging volume.

Figure 4 shows the simulation and measurement results on a phantom. The overall field patterns and the magnitude of the B ₁ ⁺ fields correspond qualitatively; however, some differences between simulations and measurements are to be noted. For all the single-transmit channels, minor differences between the simulated and measured field patterns can be discerned. In the shimmed combination, the field distribution in the center of the antenna corresponds well, but the measured transmit field at the bottom of the phantom has a lower intensity than the simulated field. The absence of highly intense peaks in the measured B ₁ ⁺ fields indicate that the loops are detuned well, and do not influence the transmit fields of the dipole antennas.

Figure 5 shows the imaging results of functional imaging for all eight subjects. Below each image, the average rating of overall image quality is reported. The image quality ratings are shown in Table 2 where 0 corresponds to poor image quality and 3 represents excellent image quality. A document with all the separate ratings is added as supplementary material. The overall image quality is rated between good and excellent (overall score 2.41). No remarkable noise is present in any of the images (overall score 3). Some of the images show artifacts (overall score 2.28), which can be recognized in Fig. 5 for volunteer 1 and 4. Inter-observer agreement was calculated for every view and criterion. The overall inter-observer agreement was 74%. Inter-observer variability was calculated using Cohen’s kappa, which is κ = 0.56 for all observations. This corresponds to a moderate agreement between both raters.

Figure 6 shows the four-chamber view results of volunteer 6, using an acceleration factor of R = 2 and different resolutions. In general, as spatial resolution increases, SNR is reduced. However, due to the higher spatial resolution, more details can be recognized in the image, especially in the region close to the cardiac walls. The maximal resolution that was reached in this experiment on one volunteer is 0.75 × 0.75 × 1.75 mm³, which corresponds to a cubic resolution of 0.98 mm³. At this resolution, noise becomes more clearly present in the image, especially at locations far away from the transmit/receive elements.

Figure 7 shows SNR maps of a single volunteer, acquired in the 4Ch view and SAX view. The separate contributions of the loop and dipole elements are demonstrated in this image. It can be seen in both views that combining the loop and dipole elements leads to a remarkable enhancement of the SNR in the heart. Table 3 shows the SNR and CNR values in the heart for the different views, averaged over the three volunteers. When using all receive elements, the average SNR in the heart is 10.9 and 11.9 for the SAX view and the 4Ch view, respectively. The CNR_blood/myo for these views is 10.7 and 12.3. The SNR increases over two-fold when all elements are used, compared to the situation where only loops coils are used for receive. Compared to a situation where only dipoles are used in receive mode, the SNR increases 50% when all elements are used.

Figure 8 shows reconstructed images and g-factor maps for a single volunteer, acquired in the SAX view and the 4ch view. As the acceleration factor is increased, the g-factor in the heart clearly increases, and noise becomes more present in the reconstructed images. However, even for an acceleration factor of R = 6, the anatomy of interest is still clearly visible. The average g-factor in the heart increases from 1.02 to 1.08 in the SAX view and from 1.04 to 1.27 when moving from an acceleration factor of R = 2 to R = 6. The maximum g-factor obtained in the SAX view is 2.00 at R = 6; in the 4Ch view, a g-factor of 3.1 is reached at R = 6 (see Table 4).

The worst-case SAR calculated here corresponds to the phase setting that yields the highest possible SAR value in the entire subject. This phase setting will not correspond to a realistic phase setting that is used for scanning. The safety simulations indicate that there is a large difference between the worst-case SAR and a SAR that corresponds to a realistic shim setting (The SAR corresponding to a realistic shim setting is 33 and 51% lower than the worst-case SAR, respectively, for Duke and Ella). While it is not likely that the worst-case SAR is obtained during a scan, this value was nevertheless used to set the upper limit of the total average power [33]. In other work [13], the total average power limits are derived based on peak SAR calculated in human models for fixed phase settings. These fixed phase settings are also used in the scan. Using this method leads to a total average power limit of 65 W, more than two-fold higher than the 8 × 4 = 32 W used in this work. This difference in the average power limit does not necessarily represent a difference in the efficiency of these two setups, but a difference in choices regarding transmit phase optimization and safety assessment.

Much effort has already been spent on matching SAR simulations with measurements as closely as possible [17, 18]. Future work into worst-case SAR estimations will very likely focus on deriving realistic drive settings for multiple human models and matching those to scan results, with the use of bi-directional couplers [33]. These methods have the potential to increase the average power limits, leading to faster cardiac exams at 7 T. Work focused on using nonlinear optimization of the transmit phases and amplitudes to reduce the maximum local SAR, or the use of SAR constraint RF pulses has not yet been applied in this work, and also have the potential to improve transmit efficiency [15, 17, 34‐36].

Electromagnetic simulations were validated by comparing simulated to measured B ₁ ⁺ maps. Field measurements or MR thermometry are possible alternatives for safety validation, but are not treated within the scope of this paper. A comparison between simulated and measured transmit fields on a phantom correspond qualitatively. However, when comparing the results in detail, differences between the simulations and measurements are clearly visible. The qualitative correspondence between the results indicates that the antennas perform as modeled in the simulation, and it indicates that the loops are detuned properly. The results could be improved upon by exactly matching the simulation and measurement geometry by doing a computed tomography scan of the measurement setup and importing this in Sim4Life. Exactly matching the simulated and measured coupling parameters by using circuit co-simulations would be another way of improving the correspondence between simulations and measurements. This will be necessary for deriving SAR limits that closely match simulations for realistic drive settings, but it is not treated within the scope of this work.

High-quality functional images were acquired for a total of eight volunteers. Image quality of the cine images has been scored by two experienced readers, leading to an average score of 2.4 (between good and excellent) for overall image quality. The rating shows that the diagnostic quality of the cine images is not at all affected by noise. The image quality scores are impacted by artifacts, which are most generally caused by a non-uniform excitation field. This is clearly present in volunteers 1 (SAX view) and 4 (p4Ch view), which is represented by the lower rating of the images. Although optimized RF transmit phase settings are calculated for every volunteer, it is clear that RF shimming methods that are used in this work do no suffice for every situation. The shimming method that is currently used maximizes the average signal in the heart, but does not necessarily provide a uniform signal. Additional RF calibrations can be used to enhance homogeneity of the signal, but this can also increase acquisition time and examination complexity [14, 16]. Future work will explore more advanced techniques to acquire homogeneous excitation fields combined with rapid calibration scans and procedures.

For SNR comparison, cine images have been acquired using scan parameters reproduced from literature to the best of our ability [13]. The flip angle of 30° could not be reproduced within a reasonable scan time. This is caused by different choices in setting the average power limits, as mentioned earlier in this discussion. The TR and TE that are reported by Oezerdem et al. (2.17 and 4.17 ms, respectively) was also not reproducible with our system and the other sequence parameters. The overall SNR values that we report at R = 2 are lower than for the two coil setups mentioned by Oezerdem et al. (11.9 compared to 29 for the 4Ch view, 10.9 compared to 29 for the SAX view). The CNR_blood/myo values that we report are comparable (10.3 compared to 11 for the 4Ch view, 9.9 compared to 9 for the SAX view). Because of the differences in acquisition methods, the results of this comparison should be interpreted with caution. An interesting comparison would be to scan both coil setups at the same imaging site; however, this will not be treated within the scope of this paper.

The average g-factor that we report at R = 6 is lower than the g-factors mentioned in literature (1.08 compared to 1.5 for the SAX view, 1.27 compared to 1.6 for the 4Ch view) [13]. As a result of the difference in MRI system vendor, SENSE is used in our work, while generalized autocalibrating partial parallel acquisition (GRAPPA) is used in [13]. Results show that the average g-factor remains low even up to an acceleration factor of 6. Although such accelerations may provide an insufficient SNR for many applications, these results show that in terms of encoding power of the array, it is possible.

The SNR maps presented here show that a setup which combines loops and dipoles in receive mode increases SNR in the heart by more than 50% compared to a setup where only loops or dipoles are used. This strong increase in SNR in not only caused by the receive sensitivity of the combined setup, but also relates to the improved acceleration that is gained my moving from 16 or 8 to 24 channels. Most setups that are used specifically for cardiac imaging make use of loop coils, while some setups use dipole antennas [9, 12, 13, 37‐39]. This work shows that these kinds of setups can be improved by combining both elements.