Enhancement of in vivo 7T magnetic resonance neuroimaging via flexible metasurfaces

Abstract

Ultra-high-field (≥7T) MRI suffers from poor transmit (B₁⁺) homogeneity, particularly in neuroimaging. This study evaluated paired metasurfaces near the temporal lobes in five healthy volunteers versus dielectric pads and reference scans. Metasurfaces markedly improved SNR across the brain (FLAIR: ~24%, MPRAGE: ~97%) and in the temporal lobes (FLAIR: ~57%, MPRAGE: ~212%), also recovering signal in the cerebellum and neck. Simulations indicated a 23.5% increase in peak SAR compared to reference.

Ultra-high field strength (≥7T) MRI has allowed for improvements in signal-to-noise ratio (SNR), which consequently enables increases in spatial resolution and enhanced chemical shift sensitivity for improved spectral quality¹. With these systems becoming more common in both the research and clinical settings, optimally leveraging these improvements will improve diagnostic outcomes and further our understanding of disease mechanisms. However, a continuing technical challenge with ultra-high field strength systems is the reduced transmit (B₁⁺) field homogeneity manifested as central image brightening and signal drop off in certain regions^2,3. This problem stems from the Larmor frequency increasing as the MRI field strength increases, which corresponds to shorter radiofrequency (RF) wavelengths used during excitation. With these shorter RF wavelengths being approximately the same size as human anatomy at 7T, a standing wave effect can be observed, which produces regions of constructive and destructive phase interference at approximately quarter wavelength distances³. In neuroimaging applications this primarily occurs in the lateral and inferior regions, such as the temporal lobes and cerebellum.

While parallel transmission (pTx)^4,5 and post-processing^6,7 techniques have been effective at improving this, they come with their own challenges in the form a larger uncertainty with respect to RF power deposition in tissue (otherwise referred to as specific absorption rate (SAR)) and the lack of corrective effectiveness below a minimum B₁⁺ threshold⁶, respectively. Alternative approaches, based on aqueous mixtures of materials with a high dielectric permittivity (dielectric pads), have been shown to increase image quality. Dielectric pads support the formation of internal displacement currents which form secondary fields that are naturally in phase with the primary RF field and therefore allow them to be summed together directly³. However, dielectric pads face practical constraints due to the limited space available (due to the required 0.5–1 cm thickness needed for the dielectric pads to be effective) in the already tight-fitting RF coils, as well as other issues such as fixed shelf life and relatively toxic material usage^8–11.

More recently, metamaterials, integrating periodic conductive unit cells, have been explored as an alternative to conventional approaches such as pTx and dielectric padding. These have come in various forms, for example, secondary resonators^12–14, which function as additional passive elements that are tuned to the specific Larmor frequency and produce a secondary magnetic field. There also exists non-resonant 2D metamaterials (often referred to as metasurfaces) that also produce a secondary magnetic field in a similar manner to dielectric pads. These metasurfaces support the formation of conduction currents in the unit cell traces, which also form secondary fields¹⁵. However, unlike the displacement currents formed inside dielectric pads, metasurface conduction currents are 90° out-of-phase and therefore sum orthogonally to the primary RF field, potentially resulting in marginal field enhancing effects. To adjust for this phase misalignment an artificial 90° phase delay needs to be applied to the induced currents to bring the metasurface secondary fields in phase with the primary RF field¹⁵. Proper optimization of these surfaces can be accomplished by varying either the size of the unit cell^16,17 or by adjusting the surface reactive impedance, such as added distributed capacitance¹⁵. This is in contrast to RF shields which would require a -90° artificial phase delay to put the induced currents 180° out of phase, thereby producing an opposing field that would effectively prevent propagation of the RF field into the imaging sample^18,19. In addition, the extremely thin form factor and computer-based design of metasurfaces allow them to be easily and repeatably integrated into existing 7 T RF coils at a relatively low production cost.

Metasurfaces have been used in vivo across a variety of applications as well as across a range of field strengths and anatomies including wrist imaging at 1.5T¹³, abdominal imaging at 3T¹⁵, as well as calf²⁰ and brain^21,22 imaging at 7T. Most recently, Koloskov et al.²³ demonstrated a metasurface design aimed at enhancing brain imaging at 7T, which showed comparable enhancement to dielectric pads. In this work we demonstrate that a metasurface previously developed by our group for calf image enhancement at 7T²⁰, also improves human brain images acquired at 7T using clinically relevant neuroimaging sequences such as gradient echo (GRE), fluid attenuated inversion recovery (FLAIR), and magnetization-prepared rapid gradient echo (MPRAGE) imaging.

Electromagnetic simulation results

Results, seen in Fig. 1, show the numerically calculated output for the B₁⁺ field_, electric field, and SAR distribution across all three experimental conditions (Fig. 1D). An increase in simulated B₁⁺ values was observed in the metasurface condition for both the left and right region of interest (ROI). Specifically, for the metasurfaces, B₁⁺ magnitude increased from the reference by 37.0% (left ROI) and 74.3% (right ROI) compared to the dielectric pads which increased from the reference by 8.6% (left ROI) and 10.4% (right ROI). The metasurface electric field results showed a similar trend, increasing from the reference by 62.7% (left ROI) and 44.4% (right ROI) compared to the dielectric pads which increased from the reference by 24.1% (left ROI) and 7.6% (right ROI). When comparing the peak spatial SAR (psSAR—maximum value averaged over any 1 g spatial volume) values, a 23.5% increase was observed when the metasurfaces were present, increasing from 0.204 W/kg to 0.252 W/kg. An increase of 3.2% was observed with the dielectric pads present, increasing from 0.204 W/kg to 0.211 W/kg. SAR efficiency values were also calculated in the same ROIs and were observed to increase, with metasurfaces increasing by 22.3% (left ROI) and 58.7% (right ROI) relative to the reference condition compared to the dielectric pads which only increased by 1.3% (left ROI) and 1.4% (right ROI). These simulated field results can be seen summarized as average and standard deviation values in Table 1.

Fig. 1. Human model EM simulations comparing metasurfaces to dielectric padding.

An image of the simulation environments showing the head model, RF volume coil, as well as the modeled A dielectric pads and B metasurfaces. C The simulated B₁⁺ field and electric field maps were quantified in the two separate white outlined ROIs on the left and right sides of the head model. D Numerically calculated field results for the reference, dielectric pads (position denoted by the white vertical lines), and metasurfaces (position denoted by the red vertical lines). Corresponding local SAR distribution can also be seen in the last row annotated with the corresponding psSAR values.

Table. 1.

Average and standard deviation values for the simulated B1+ and E field values in both the left and right ROIs across experimental conditions

	B1+ field (µT)
	Left ROI	Right ROI
Reference	0.095 ± 0.06	0.067 ± 0.06
Dielectric Pads	0.074 ± 0.06	0.076 ± 0.01
Metasurfaces	0.130 ± 0.05	0.118 ± 0.12

	E field (V/m)
	Left ROI	Right ROI
Reference	9.35 ± 0.86	8.59 ± 0.35
Dielectric Pads	11.61 ± 0.77	9.25 ± 0.29
Metasurfaces	15.22 ± 1.62	12.41 ± 0.77

	SAR Efficiency ($\mathrm{\mu }\mathrm{T}/\sqrt{\mathrm{W}/\mathrm{kg}}$)
	Left ROI	Right ROI
Reference	0.210 ± 0.013	0.148 ± 0.013
Dielectric Pads	0.213 ± 0.014	0.150 ± 0.003
Metasurfaces	0.259 ± 0.013	0.235 ± 0.028

Additionally, simulated SAR efficiency values have also been added from the same corresponding ROIs.

Image enhancement quantification results

Figure 2 shows GRE images in the axial orientation through the medial temporal lobes as well as in the sagittal orientation through the lateral temporal lobes for all five subjects across the three experimental conditions. Qualitatively, consistent and repeatable increases in the global and temporal lobe image quality can be observed with the metasurfaces present compared to both dielectric pads and the reference condition. Additionally, image quality in more inferior regions, such as the cerebellum and neck anatomy, are also improved.

Fig. 2. Axial and sagittal GRE images with regions showing the temporal lobe(s) between the three experimental conditions tested across all five subjects imaged.

Dielectric pads and metasurfaces were placed on either side of the subject’s head near the temporal lobes. It can be seen qualitatively that the metasurfaces produced the highest amount of both global and temporal lobe(s) focused signal enhancement compared to both the dielectric pads and reference conditions.

Transmit efficiency maps can be seen in Fig. 3A, showing an increase in overall magnitude in the temporal lobes. It should be noted that small truncation artifacts can be observed in the posterior regions of brain when the metasurfaces are used. These results are supported by an increase in average transmit efficiency measurements which showed a 13.3% increase from 0.045 ± 0.012 µT/V in the reference configuration to 0.051 ± 0.013 µT/V when the metasurfaces were used, compared to a 4.4% increase to 0.047 ± 0.011 µT/V when using dielectric pads. These results can also be seen illustrated in Fig. 3B, which shows a subject wise bar plot of the average transmit efficiency for the reference, dielectric pad, and metasurface cases. It can be seen consistently, on a subject-by-subject level, that the metasurfaces produced higher magnitude transmit efficiency. It should be noted that, when measured, negligible changes in the average transmit reference voltage for each experimental condition were observed, with the reference condition yielding 236.4 ± 9.2 V, the dielectric pads yielding 233.4 ± 7.9 V, and the metasurfaces yielding 230.0 ± 12.0 V. A subject-by-subject breakdown of the transmit reference voltages used can be seen in Supplemental Table 1.

Fig. 3. Axial and sagittal transmit efficiency maps showing the temporal lobes across subjects between experimental conditions.

A Transmit efficiency maps reflecting the B₁⁺ field distribution in the experimental conditions studied. A consistent increase in the local transmit efficiency is observed across subjects when metasurfaces were used, showing larger enhancements compared to conventional dielectric pads. B A bar plot showing the average and standard deviation of the transmit efficiency within the displayed axial imaging slice for the reference (blue), dielectric pads (red), and metasurfaces (green) cases. Metasurfaces can be seen to increase overall transmit efficiency magnitude compared with the dielectric pad and reference conditions.

Relative receive field maps can be seen in Fig. 4 for two representative subjects in all three orientations both without and with the metasurfaces present. It can be qualitatively observed that relative receive sensitivity increased for both subjects primarily in the temporal lobes. Ratio maps seen for both subjects also show average quantitative increases of up to 72% across the whole brain image slice and 124% in a localized temporal lobe region.

Fig. 4. Representative relative receive field maps shown in the axial, coronal, and sagittal orientations across two subjects without and with metasurfaces present.

Ratio maps can also be seen for both subjects quantifying the amount of receive field gain achieved with the metasurfaces. It should be noted that due to experimental limitations, dielectric pads were not used here.

Sagittal and coronal FLAIR images as well as generated SNR maps for two representative subjects, seen in Fig. 5A, not only show an increase in whole brain and temporal lobe signal but also improved SNR in the cerebellum, superior neck, and mandible anatomy. Figure 5B shows quantified SNR bar plots for the sagittal FLAIR images within the whole brain (red outlined ROI in the inset image) and temporal lobe (blue outlined ROI in the inset image) for both representative subjects. Relative to the reference condition, average whole brain SNR increased across both subjects by 19.4% with the metasurfaces present compared to 10.7% with the dielectric pads present. Average temporal lobe SNR increased across both subjects by 51.2% with the metasurfaces present compared to 40.6% with the dielectric pads present. Figure 5C shows quantified SNR bar plots for the coronal FLAIR images within the whole brain (red outlined ROI in the inset image) and temporal lobes (blue outlined ROIs in the inset image) for both representative subjects. Relative to the reference condition, average whole brain SNR increased across both subjects by 28.1% with the metasurfaces present compared to 15.1% with the dielectric pads present. Average SNR in the temporal lobes increased across both subjects by 63.7% with the metasurfaces present compared to 31.2% with the dielectric pads present. Exact SNR values for these image sets can be seen in Supplemental Tables 2 and 3.

Fig. 5. In vivo FLAIR image enhancement between experimental conditions.

A A comparison of FLAIR images and corresponding SNR maps in two representative subjects across the three experimental conditions in both the sagittal and coronal orientations. In both subjects more recovered signal can be observed in the medial temporal lobes and cerebellum when the metasurfaces are present in comparison to the dielectric padding. This is also reflected in the SNR maps with moderately higher SNR values both globally and in the same local areas such as the temporal lobes. B A bar plot showing the quantified reference, dielectric pad, and metasurface image SNR for both subjects individually in the sagittal orientation for a whole brain (red outline) and temporal lobe (blue outline) ROIs. C The same bar plot quantification performed for the coronally oriented images.

Figure 6A shows axial and sagittal MPRAGE images as well as generated SNR maps for the same two representative subjects, which also demonstrated similar SNR increases across the whole brain and in the temporal lobes. Figure 6B shows quantified SNR bar plots for the axial MPRAGE within the whole brain (red outlined ROI in the image inset) and temporal lobes (blue outlined ROIs in the image inset) for both representative subjects. Relative to the reference condition, average whole brain SNR increased across both subjects by 123.5% with the metasurfaces present compared to 46.3% with the dielectric pads present. Average temporal lobe SNR increased across both subjects by 246.2% with the metasurfaces present compared to 91.3% with the dielectric pads presents. Figure 6C shows quantified SNR bar plots for the sagittal MPRAGE images within the whole brain (red outlined ROI in the image inset) and temporal lobe (blue outlined ROI in the image inset) for both representative subjects. Relative to the reference condition, average whole brain SNR increased across both subjects by 69.5% with the metasurfaces present compared to 35.9% with the dielectric pads present. Average temporal lobe SNR increased across both subjects by 179.1% with the metasurfaces present compared to 94.0% with the dielectric pads presents. Exact SNR values for these image sets can be seen in Supplemental Tables 4 and 5.

Fig. 6. In vivo MPRAGE image enhancement between experimental conditions.

A A comparison of MPRAGE images and corresponding SNR maps in two representative subjects across the experimental conditions in both the axial and sagittal orientations. In both subjects, increased signal in the medial temporal lobes and occipital lobe can be observed when using the metasurfaces compared to both dielectric pad and reference cases. This is also reflected in the SNR maps which show moderately higher SNR values both globally and in local regions such as the temporal lobes. B A bar plot showing the quantified reference, dielectric pad, and metasurface image SNR for both subjects individually in the axial orientation for a whole brain (red outline) and temporal lobes (blue outlines) ROI(s). C The same bar plot quantification performed for the sagittal oriented images.

Interpretation of results

In the present work we used metasurfaces, previously developed by our group, for enhancing clinical anatomical brain images at 7T and establishing initial multi subject repeatability towards this application. Three sets of GRE brain images were acquired in five healthy volunteers, with two metasurfaces present, two dielectric pads present, as well as a reference configuration to serve as a comparison. Metasurfaces produced a 13.3% increase in transmit efficiency compared to the 4.4% observed with the dielectric pads. Metasurfaces also yielded larger increases in SNR seen in the SNR maps generated for the FLAIR and MPRAGE image sets compared to the dielectric pads. In some cases, such as in the FLAIR images, complete signal recovery was achieved in anatomical regions such as in the cerebellum, superior neck, and mandible anatomies, that was not achieved with the dielectric pads. Of note, the contrast recovery in the cerebellum using the metasurfaces would allow for more practical imaging studies of that region. Substantial magnitude increases in the receive field maps were also observed in similar regions as transmit efficiency enhancements.

Through this work and other studies, it has been demonstrated that metasurfaces have an advantage over other correction methods. PTx methods have been shown in multiple cases to aid in correcting B₁⁺ inhomogeneities^5,24, however, they involve an increased uncertainty with respect to local SAR hotspot formation. One notable pTx approach was by Herrler et al.²⁵ who showed the use of a combination of universal pulses and individualized optimization to accommodate for inhomogeneous B₁ field for ultra-high field imaging under strict SAR limits. However, this method is still subject to the same SAR restrictions which may limit its use for more SAR intensive sequences. Additionally, this method must be recalibrated for use with different pTx coils while the metasurfaces are effectively universal for most 7T single channel volume transmit coils. Other correction techniques include the k_T-points method which involves efficiently navigating through k-space in order to skip spatial frequencies that do not need to be acquired in the context of smoothly varying B₁ inhomogeneities²⁶. However, limitations of this method include potentially higher SAR, the reliance on small flip angle approximations, as well as k_T points being a 3D acquisition method, which makes it challenging to apply to 2D slice selection. When comparing metasurfaces to dielectric pads, the metasurfaces were not only able to outperform them in image enhancement but were also much thinner and lighter in weight which makes them more compatible with current 7T head coils. Additionally, metasurfaces in principle have an indefinite shelf life which, coupled with their low cost and being easily sanitized for multiple patient use, make them a more sustainable solution. However, it should be noted that dielectric pads are a well-established passive correction method with a more predictable safety profile. So, while the metasurfaces may perform better in certain aspects, more work is needed in safety characterization. Finally, while initial metasurface design and testing can be a tedious process, once complete, their design files can be easily disseminated to individual groups for independent reproduction and use through standard manufacturers.

Koloskov et al.²³ recently reported a flexible metasurface design for brain imaging at 7T. Their results showed relative increases in B₁⁺ magnitude in the temporal lobes and cerebellum that were comparable to dielectric padding. In the present work, our design showed similar to larger improvements in image enhancement in the temporal lobes and cerebellum, which was also observed to extend into the superior neck anatomy in comparison to both reference and dielectric padding conditions. One potential factor in the difference in performance between these studies could be due to the optimization methods used. The metasurface reported by Koloskov et al.²³ was designed via electromagnetic simulations for optimal capacitance values in separate left/right metasurfaces of 4.8 pF and 7 pF. The design presented here was optimized via exhaustive physical testing to determine which impedance configuration produced the greatest image enhancement, which yielded a value of 11 pF. This demonstrates the high response sensitivity of these devices to small shifts in additional applied impedance. This observation also shows that while simulations have tremendous value, the translation of computationally derived designs to real world applications may be offset by small modeling errors.

Several limitations were present in this work, including the lack of modeling of the Nova head coil 32-channel receive array in the EM simulation environment. This prevented evaluation of the interaction between the full RF coil environment, the metasurface, and the human head model. Additionally, the imposed 50% SAR limit (normal mode operation) was also an acquisition limitation. This restriction was set as a safety precaution not only based on the simulation results presented here, which showed a maximum ~23.5% increase in peak spatial SAR, but also on phantom temperature experiments conducted in our previous work²⁰. Specifically in the previous work, when a single metasurface was used with a phantom over a 20-minute 100% SAR scan period (three replicates), the total temperature rise increased from 0.99 ± 0.03 °C to 1.52 ± 0.08 °C (53.5%). However, when the same experiment was performed using a two metasurface configuration the temperature rise only increased to 1.11 ± 0.02 °C (12.1%). This observed heating was measured using fiber optic probes placed immediately adjacent to the metasurface, at increasing surface depths, and was not observed in any of the non-local fiber optic probes. This difference in heating between single and dual metasurface conditions was partly the reason why two metasurfaces were used here, with one on either side of the subject’s head, rather than a single metasurface placed near the cerebellum, as some other groups have done^12,23,27. This imposed SAR limit precluded us from evaluating the effects of metasurfaces in certain SAR intensive sequences, such as glutamate-weighted chemical exchange saturation transfer (CEST) imaging²⁸. It has been shown in previous literature that dielectric pads are able to improve SAR by allowing the scanner to self-calibrate towards a lower transmit reference power to achieve the same effective B₁ field^9,29,30. Since the metasurfaces function in a highly similar manner to dielectric pads, they should also have the same effect, in which the transmit reference voltage would self-calibrate to a lower input power and therefore actually improve SAR. However, in practice the observed transmit reference voltage values showed marginal to no change between experimental conditions. One potential explanation for these observed results may come from the fact that when the scanner is self-calibrating to achieve an optimal 90° flip angle, it does so over the entire brain region rather than a localized region. Since the metasurface effects are relatively localized, a fairer comparison would be to perform the same experiment on a smaller voxel region and observe the transmit reference power necessary to produce the optimal flip angle. Our group has performed similar spectroscopy-based experiments in the past that match the hypothesized results³¹. Additionally, in the presented simulations the calculated SAR efficiency was observed to substantially increase with the metasurfaces present, indicating improved safety. However, a further and more dedicated evaluation of this effect is necessary.

Truncation effects stemming from asymmetric unit cell loading is another known limitation of the current metasurface design. An idealized theoretical description of the metasurface would assume that there is an infinite plane of identical unit cells in which each unit cell is surrounded by an equal number of identical unit cells. However, in practice metasurfaces have finite dimensions and therefore, this description breaks down. Due to this effect, portions of the metasurface near the edges may not improve image quality to the same degree and under certain circumstances, when placed incorrectly near a patient’s head, can further reduce image signal. Brizi et al.³² characterized this problem in the context of wireless power transfer and showed that by varying the capacitance across the metasurface, these truncation effects could be compensated for. Care should be taken when manually placing the metasurfaces on either side of the subject’s head to avoid these potential artifacts.

In conclusion, we have shown a metasurface that is effective at enhancing in vivo neuroimaging at 7 T using clinically relevant sequences demonstrating clinical translatability of this technology. Future work will focus on performing more detailed volumetric temperature mapping with the metasurfaces, refining the metasurface design to yield further performance improvements for proton and other nuclei of interest, performing further validation studies, and assessing the integration of metasurfaces with other sequences, such as metabolic imaging (i.e., glutamate weighted CEST)³³ and spectroscopic techniques^34,35.

Methods

Metasurfaces and dielectric pads

A pair of two-layer printed circuit boards (PCBs) 18 cm × 18 cm flexible metasurfaces, illustrated in Fig. 7A, were used to enhance images acquired at 7T. The metasurfaces utilized an inductive mesh grid design which incorporated parallel plate capacitors at the unit cell mid points. The design parameters included an 11.25 mm unit cell period, a 1 mm trace width, and a 3 mm × 3 mm capacitor plate size manufactured (PCBWay, Shenzhen, China) on a 25-µm thick flexible polyimide substrate sourced from DuPont (Kapton), with a relative permittivity (ε_r) of 3.5 and loss tangent of 0.003 at 1 GHz. A zoomed in annotated unit cell region can be seen in Fig. 7B. In this design, the conductive mesh grid is identical on the top and bottom layers of the PCB layout. The metasurfaces used here are identical to those used in our previous in vivo calf and metabolic brain imaging work^20,21. As an added standard comparison, two conventional 0.8 cm thick aqueous calcium titanate dielectric pads (7TNS, Multiwave Imaging, Marseille, France; ε_r = 110) with the same length and width as the metasurfaces were also separately used³⁶.

Fig. 7. Metasurface device and in vivo placement.

A The prototype metasurface with integrated parallel plate capacitors shown and B a close-up view of an individual unit cell annotated with the relevant dimensions. C An image showing the physical placement of two metasurfaces within the 7T head coil near the temporal lobes on either side of a subject’s head.

Electromagnetic simulations

Numerical simulations were executed using a full wave finite-difference time domain solver (Sim4Life version 8.0, ZMT, Switzerland) to calculate the B₁⁺ field, electric field, and SAR distributions. The simulation environment consisted of tuned lumped element resonant RF coil model tuned to 300 MHz. A human head model (Duke v3.0, IT’IS Foundation) was used for three simulation conditions: a reference, dual dielectric pads, and dual metasurfaces.

Dielectric pads were modeled as an 18 cm × 18 cm × 0.8 cm curved slab with corresponding properties to aqueous calcium titanate (σ = 0.09 S/m, ε_r = 110)^2,37. The dielectric pad models were positioned on either side of the head model near the temporal lobes, as seen in Fig. 1A. Metasurfaces with the same curvature as the dielectric pads were modeled in the simulation environment using the PCB design. In place of the parallel plate capacitors, 454 lumped element capacitors were set to 11 pF. Metasurface conductor traces were assigned as perfect electric conductor and the PCB substrate material as polyimide (σ = 60 µS/m, ε_r = 3.50). As with the dielectric pad models, the curved metasurfaces were also positioned on either side of the head model in the same orientation, as seen in Fig. 1B.

Simulations were performed with a spatial resolution of 0.8 mm over the combined metasurface/dielectric pad and human head models, for a total of approximately 173 MCells. Simulations were set to reach a −50 dB convergence and executed using a dedicated GPU (GeForce RTX, 4080, NVIDIA, USA), taking approximately 3 h to complete. Simulation post-processing consisted of adjusting the RF coil to quadrature drive with a normalized input power of 1 W. The primary extracted quantities from all simulation conditions consisted of B₁⁺ field, electric field, and 1 g-average SAR (SAR_1g). For SAR calculations, tissue volume averaging was chosen to be performed over 1g masses (rather than 10 g) due to its ability to capture smaller scale SAR maxima that may be present due to the smaller RF wavelengths used at 7T³⁸. Additionally, psSAR was also measured as a SAR quantification method, while two ROI regions were used to quantify the average and standard deviation of the B₁⁺ and electric fields (Fig. 1C). Using the quantified ROI B₁⁺ values and the psSAR values from the EM simulations, SAR efficiency was also calculated using the same 1 W normalized input power, as shown in Eq. 1.

$$\mathit{SAR}\mathit{Efficiency}=\frac{B_1^{+}}{\sqrt{\mathit{psSAR}}}$$ 1

MR acquisition and post-processing

All imaging data were acquired on a whole body 7T MRI system (MAGNETOM Terra, Siemens Healthcare, Erlangen, Germany) using a single channel transmit/32-channel receive array head coil (Nova Medical, Wilmington, MA, USA). In vivo imaging was performed on five healthy volunteers (4 males, 1 female) between the dates of 01/12/2024 and 04/04/2024, with written informed consent under an approved University of Pennsylvania regulatory board protocol. All research involving human research participants, material, or data was performed in accordance with the Declaration of Helsinki. Experimental conditions included a reference control acquisition, an acquisition with two dielectric pads placed on either side of the subject’s head, and an identical acquisition separately using two metasurfaces placed on either side of the subject’s head, as shown in Fig. 7C. Based on our previous phantom safety assessments²⁰ and in vivo use²¹, all scans were run at a conservative 50% SAR limit (normal mode of operation on Siemens user interface) to mitigate any concerns of potential heating effects. Initial whole brain GRE images were acquired on all five subjects with an isotropic resolution of 1.5 mm³ and were acquired using the following parameters: TR = 15 ms, TE = 1.53 ms, flip angle = 15°, matrix size = 160 × 120, and number of slices = 128. Whole brain B₁⁺ maps were also acquired on all five subjects using a modified turbo flash sequence with an isotropic resolution of 2.5 mm³ as described by Volz et al³⁹. utilizing nonselective preparation pulses with the following parameters: shot TR = 6000 ms, imaging TR = 3.50 ms, TE = 1.30 ms, flip angles = 15°, 20°, 30°, 40°, 60°, 80°, shots = 5, matrix size = 96 × 72, and number of slices = 168. Relative B₁⁺ maps were produced via fitting to the six acquired flip angles and were then converted into transmit efficiency maps by normalizing them to the corresponding transmit reference voltage.

Receive field ratio maps were generated by first acquiring a second set of GRE images with an isotropic resolution of 2 mm³ using the following parameters: TR = 10 s, TE = 5.47 ms, flip angle = 90°, matrix size = 120 × 90, and number of slices = 80. These GRE images were post-processed into receive field maps by first describing the GRE image signal ($\mathit{GR}{E}_{\mathit{Sig}}$) as primarily dependent on the flip angle ($\alpha$), TR, and B₁^-, as seen in Eq. 2. When the TR is sufficiently long (TR » T₁) this relationship can be simplified to form Eq. 3. It can be seen in Eq. 4 that $\alpha$ is proportional to the nominal flip angle (${\alpha }_{\mathit{nom}}$) and the B₁⁺ field. When Eq. 4 is substituted into Eq. 3, relative B₁^- can be determined as the ratio of $\mathit{GR}{E}_{\mathit{Sig}}$ and the sine of ${\alpha }_{\mathit{nom}}$ and B₁⁺, seen in Eq. 5^40,41.

$${\mathit{GRE}}_{\mathit{Sig}}\propto \frac{M_0\sin \left(\alpha \right)\left(1-e^{-\frac{\mathit{TR}}{T_1}}\right)B_1^{-}}{1-\cos \left(\alpha \right)e^{-\frac{\mathit{TR}}{T_1}}}$$ 2

$$\mathit{GR}{E}_{\mathit{Sig}}\propto M_0\sin \left(\alpha \right)B_1^{-}$$ 3

$$\alpha ={\alpha }_{\mathit{nom}}{B}_1^{+}$$ 4

$$M_0{B}_1^{-}\propto \frac{\mathit{GR}{E}_{\mathit{Sig}}}{\sin \left({\alpha }_{\mathit{nom}}{B}_1^{+}\right)}$$ 5

Whole brain FLAIR images were acquired using a modified SPACE sequence⁴² on two subjects in the sagittal and coronal orientations with the following parameters: TR = 7600 ms, TI = 2400 ms, TE = 476 ms, flip angle = 120°, turbo factor = 230, in-plane resolution = 0.75 mm × 0.75 mm, slice thickness = 2 mm, matrix size of 320 × 320, and number of slices = 208. Whole brain axial MPRAGE⁴³ images were acquired on the same two subjects with the following parameters: TR = 4000 ms, TE = 4.21 ms, TI = 1100 ms, flip angle = 7°, in-plane resolution = 0.9 mm × 0.9 mm, slice thickness = 1 mm, matrix size = 256 × 192, and number of slices = 160. Finally, for the FLAIR and MPRAGE images acquired on two subjects, a second acquisition was performed with the transmit reference voltage set to 0 V. A standard deviation filter, with a 5 × 5-pixel moving window, was then applied to generate noise images⁴⁴. Mean signal images were generated by convolving the same size 5 × 5-pixel window across the image. SNR maps were calculated as the ratio of the mean signal image and the noise image, which was performed for both imaging sequences. All images were equally windowed and leveled across experimental conditions and subjects for each individual sequence. Image registration between experimental conditions was performed using the FMRIB Software Library (FSL) toolbox⁴⁵.

Average and standard deviation SNR values were then quantified from the SNR maps across the combined cerebrum and cerebellum regions within the presented image slice for each experimental condition. Additionally, a separate SNR quantification was also performed for the temporal lobe region(s). Here, only one ROI was used for sagittal image acquisitions, while two regions were used for coronal and axial acquisitions.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Author contributions

P.S.J. developed the metasurface, collected the in vivo data, and wrote the original manuscript. N.W. provided imaging sequences and aided in developing the metasurfaces. W.B. provided insight into experimental design and metasurface physics. J.D. and M.E. provided manuscript feedback. R.R. provided funding and aided experimental design. All authors read and approved the final manuscript.

Data availability

The datasets generated and/or analyzed during the current study are not publicly available due to sensitivity and participant privacy, but are available from the corresponding author upon reasonable request.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s44303-026-00162-x.