In vivo meso-scale whole-brain quantitative imaging with tailored MRF on the NexGen 7T scanner

Abstract

Purpose:

To push the speed and resolution limit of in-vivo quantitative imaging and enable estimation of quantitative tissue parameters of subtle brain structures that were previously difficult to assess.

Methods:

This study implemented an efficient quantitative imaging approach, 3D-SPI MRF, on the NexGen 7T scanner equipped with a high-performance head-only gradient and 96-channel receiver array. To address challenges associated with performing rapid mesoscale MRF on this system, acquisition and reconstruction mitigation methods were developed and incorporated into the MRF framework, including: i) flip-angle-aware dictionary fitting to account for both B₁⁺ inhomogeneity and voxel-specific RF frequency response, ii) gradient imperfection corrections via Skope measurements that incorporates a new per-TR trajectory rewinder compensation, iii) incorporation of rapid B₁⁺ and B₀ mappings into the MRF sequence, and iv) high-temporal motion navigation.

Results:

Whole-brain T₁ and T₂ maps were obtained at 560-μm isotropic resolution within 4 minutes, where ablation studies demonstrated the necessity of the various mitigation methods implemented in removing bias and artifacts. For comparison, MRF data were acquired using current state-of-the-art method but limited to typical whole-body gradient specifications to demonstrate that the proposed developments resulted in ~3× shorter scan time while producing more accurate parameter maps. Data were also acquired at ~3.8× smaller voxel size, 360-μm isotropic, using the developed technique, to achieve mesoscale multi-parameter quantitative mapping in vivo.

Conclusion:

Tailored 3D MRF acquisition and reconstruction were developed to enable fast and accurate T₁ and T₂ mapping across the whole-brain at mesoscale resolution on the NexGen 7T scanner.

1. Introduction

High-resolution quantitative imaging presents both a challenging and promising frontier in MRI research¹. Submillimeter resolution can enable the detection of subtle quantitative tissue parameter changes in small brain structures during early disease pathogenesis², as well as developmental and aging processes³. These quantitative changes could be used as biomarkers for many applications including myelination abnormality across cortical layers, micro plaques, early-stage fibrosis and iron deposition. However, such quantitative imaging is often limited by long scan time and limited SNR. Compared to traditional contrast-weighted images that can achieve submillimeter-level resolutions (e.g., 0.4-mm or higher⁴), this level of resolution in quantitative imaging remains difficult to perform, particularly across multiple tissue parameters.

As a rapid, multiparametric quantitative imaging technique, MR Fingerprinting (MRF⁵) holds the potential to push the limits of resolution in quantitative imaging when effectively implemented on high-performance, high-field scanners. In particular, 7T systems can provide substantially higher SNR, which is crucial for overcoming the signal limitations inherent to ultra-high-resolution imaging. Prior studies have leveraged the increased SNR at 7T for MRF applications targeting nuclei with inherently low signal, such as ²³Na⁶ and ³¹P⁷. Other investigations have proposed novel strategies to measure and correct the severe B₁+ inhomogeneity commonly observed at 7T^8,9 by optimizing MRF acquisitions that enjoy the benefit of high-field SNR avoiding the quantification biases associated with B₁+ inhomogeneity. Additionally, the use of shorter readout trajectories, such as radial^10,11 or Cartesian¹² sampling, has been explored to mitigate distortion and blurring induced by the severe B₀ inhomogeneity at 7T. Yet, whole-brain, ultra-high-resolution quantitative imaging at 7T with the use of MRF remains largely unexplored.

Other simultaneous multi-parametric approaches have also been developed at ultra-high field, including QRAGE¹³, MR-STAT¹⁴ and QUICS¹⁵. These studies demonstrate the technical feasibility of concurrent quantitative mapping at ultra high fields and highlight the potential of advanced reconstruction and modeling strategies to overcome the intrinsic field inhomogeneities of 7T. Several studies have leveraged parallel transmission coils available on 7T systems to achieve more accurate RF excitation, thereby improving the accuracy of quantitative imaging at 7T^16,17. In addition, methods such as MP2RAGE^18,19 and 3D-QALAS²⁰ adapted for 7T have pushed the spatial resolution of multi-parametric imaging to the 500–600 μm range, which is the highest reported to date. However, they typically require comparatively long acquisition time of 16–20 minutes for whole-brain coverage.

This study aims to implement a tailored MRF acquisition on a recently-developed Siemens NexGen 7T MAGNETOM Terra Impulse edition scanner²¹ to shorten the scan time required for submillimeter quantitative imaging and to push the spatial resolution limits. The NexGen 7T scanner features a head-only gradient system with PNS optimization^21–23 achieving a maximum gradient amplitude (G_max) of 200 mT/m and a maximum slew rate (SR_max) of 900 T/m/s along with a high-density 96-channel receiver coil that should enable faster k-space traversal, better parallel imaging performance, and higher SNR. The scanner has enabled high resolution fMRI with whole brain coverage²⁴, however, the full exploitation of MRF’s capability on this NexGen 7T scanner faces several challenges, as outlined below:

1. B₀ and B₁⁺ Field Inhomogeneities: in addition to the image blurring and distortions caused by significant B₀ inhomogeneities, it also impacts tissue quantification due to changes in the effective flip angle with the limited bandwidth of typical RF pulses used in MRF, especially when water-only excitation is used to avoid fat artifacts. Moreover, local B₀ inhomogeneities can cause the frequency of fat to deviate from its theoretical value, making spectral suppression of fat signal more challenging.

2. Gradient Imperfections: the ability to reach higher slew rates also comes with the risk of introducing additional image artifacts. Moreover, as part of this work, we observed that imperfect k-space trajectory rewinding in the MRF sequence can cause a reduction in signal refocusing, leading to a bias in T₂ quantification.

3. Motion Artifacts: subtle motions can lead to noticeable blurring in ultra-high-resolution imaging, even for a motion-robust sequence like MRF.

To address these challenges, this study integrates fast B₁⁺ and B₀ field mapping and high resolution motion-navigation modules into the 3D spiral-projection MRF sequence (3D-SPI MRF^25–27), along with a redesigned water-only excitation RF pulse for improved fat suppression. External field probes (Skope field camera²⁸) measurements were also employed to characterize gradient imperfections. Notably, this work is the first quantitative imaging study to highlight and correct the FA attenuation bias that is jointly caused by RF frequency response and local off-resonance shift, while also proposing the use of compensation gradients to improve spiral trajectory rewinding for better signal refocusing and thus more accurate quantitative estimation.

The technologies developed in this work were shown to enable accurate whole-brain MRF at 560-μm isotropic resolution in less than 4 minutes. Data were also acquired at 360-μm isotropic resolution with motion correction, to enable, for the first time, multi-parameter quantitative mapping of the human brain in vivo at this high a resolution.

2. Method

2.1. Sequence design

The FISP-based MRF sequence²⁹ used in this work is based on our recently developed 3D-spiral projection-imaging (3D-SPI) MRF sequence²⁵ that contains a 3D-SPI motion-navigator module³⁰. Using this motion-corrected 3D-SPI MRF sequence as a base, we tailored the acquisition to allow it to work well at high resolution with the high-performance and high-field system. A brief description of sequence is provided below, with detailed descriptions in our previous work²⁵. Following an inversion preparation using a 20-ms adiabatic tanh pulse³¹ with an inversion time (TI) set at 15 ms, there are 500 TRs acquired with a range of varying flip angles (FAs) spanning between 10 and 90 degrees, using different spiral projection interleaves and a constant TR of 12.5 ms. Once the data acquisition for a particular acquisition group across the 500 TRs is completed, there is a 1.2-s sequence deadtime to allow for M_z recovery, resulting in a total acquisition time of 8.7 s per group. The next acquisition group is then performed to gather additional spiral interleaves at each TR for a more comprehensive spatiotemporal k-space encoding. A constant low flip-angle (10 degree) FISP acquisition is also performed during the sequence deadtime across 100 TRs using 3D-SPI encoding to provide 2-mm navigator volume with each acquisition group, where the low-flip angle excitations used still allow for good M_z recovery.

Specific improvements to this sequence performed in this work include:

1. A rapid joint B₁⁺ and B₀ calibration scan at 1.6-mm resolution (Figure 1A) is incorporated into the MRF sequence in place of the conventional 8.7 s dummy scan used to reach steady-state M_z signal (typically the first acquisition group is discarded). This calibration scan is 15 s in total (9 s for B₁⁺ module and 6 s for B₀ module), thereby adding only 6.3 s of scan time to the MRF sequence.

   The B₁⁺ estimation module (Figure 2A) employs a 6-ms 240-degree Bloch-Siegert (B-S) adiabatic pulse³² following the excitation pulse, where the RF phase was alternated by 180 degrees between odd and even TRs (Figure 2B). Images from the odd and even TRs are then reconstructed separately, with the B₁⁺ maps derived from their phase difference. This B₁⁺ estimation module employs a 1.6-mm isotropic 3D-SPI spiral readout over 480 TRs with 18.5 ms per TR. The adiabatic B-S pulse design is used here to enable accurate B₁⁺ mapping over a large range of B₁⁺ inhomogeneities. To evaluate its performance, a Fermi-shaped B-S pulse is also implemented and acquired for comparison, as well as the standard Siemens B₁⁺ mapping sequence using TurboFLASH³³, which takes about 80 s for the same resolution.

   The sequence diagram for the B₀ estimation module is shown in Figure 2D. This module follows a similar acquisition scheme to the B₁⁺ estimation module, using the same k-space sampling trajectory, number of TRs, and reconstruction framework, but now without the B-S pulse. Different TEs (1.7ms and 2.7ms, respectively, resulting in a ΔTE of 1ms) are used across the odd and even TRs to generate phase differences needed to calculate the B₀ field map. To evaluate its performance, data were also acquired with the standard Siemens B₀ mapping sequence using dual-echo GRE for reference, which takes about 40 s for the same resolution. To ensure the same steady-state longitudinal magnetization M_z is reached at the end of acquisition like the regular MRF acquisition group, the same TR (12.5 ms) and flip angle (10 degree constantly) settings as those in the final 100 TRs of the MRF acquisition are used. This eliminates the need for a separate dummy acquisition group typically required in conventional 3D MRF, thereby saving the acquisition time associated with one entire group (8.7 s).

2. Compensation of imperfect gradient rewinder (Figure 1B): Gradient imperfections can cause incomplete gradient rewinding in the 3D-SPI MRF acquisition, where the amount of residual gradient area can vary across different interleaves and rotations of the 3D-SPI spiral readouts and hence across TRs. Based on Skope measurements (Supporting Figure S1A), the observed residual gradient moment can reach 50 μs·mT/m. We observed that the variations in this residual gradient moment across TRs could lead to imperfect signal refocusing in the FISP signal pathway, affecting the signal evolution and thus introducing large biases in T₂ quantitative estimation. To compensate for this, a TR-specific rewinder compensation gradient blip is added in each TR, where Skope field measurement confirms that this correction is able to reduce the rewinding imperfection by more than 10 fold to less than 5 μs·mT/m (Supporting Figure S1B).

3. Improved fat suppression: A water-only excitation RF pulse (2.86-ms duration, passband / stopband ripple = 0.46 / 0. 0001) was designed using the SLR algorithm³⁴ (Figure 3A) to provide improved water-only excitation with less fat contamination³⁵ at 7T. Compared to the hard pulse design previously employed in MRF at 3T^25,35 and to Gaussian pulse designs (the RF pulse duration was designed to ensure the presence of a zero-crossing around the fat resonance frequency), the frequency response of this SLR pulse (shown in Figure 3B) exhibited markedly lower excitation around the fat frequency (from −900Hz to −1100Hz for example), resulting in reduced fat signal contamination in the presence of local B₀ inhomogeneities. Furthermore, the broader Full-Width-at-Half-Maximum (FWHM) of its frequency response also reduces the flip-angle (FA) attenuation effect on the water signal caused by B₀ inhomogeneity. This FA attenuation effect and its correction will be described in Section 2.2 below (FA-corrected dictionary fitting).

4. A Skope trigger was added prior to the spiral readout. This allows the Skope field camera to measure system imperfections, such as the actual spiral trajectory, B₀-eddy-current induced phase shifts, and residual gradient moments, which are used for determining the amplitude of trajectory rewinder compensation gradients in ii), and for incorporating into the spiral reconstruction to improve reconstruction fidelity.

5. To provide higher resolution volume navigators needed for our target mesoscale imaging, the MRF motion navigator module^30,36 was modified to increase the volume navigator resolution from 4mm in our previous work to 2 mm here. This was achieved via faster k-space traversal with the head insert gradient as well as increasing the number of navigator acquisitions from 40 TRs to 100 TRs during the sequence deadtime after the MRF acquisition group to allow the encoding of the volume with 2-mm isotropic resolution to be completed during a single deadtime period and maintain motion navigation temporal resolution of 8.7s.

Figure 1.

(A) Global sequence diagram including an upfront calibration (B1+ and B0 estimation) module, and MRF acquisition and motion-navigator modules that get repeated across MRF acquisition groups.

(B) Sequence diagram of the modified MRF acquisition within a specific TR, where a redesigned SLR pulse is used to reduce fat signal contamination and a TR-specific rewinder compensation gradient is introduced to help keep gradient area constant for each TR, thus getting better signal refocus.

(C) Reconstruction pipeline, including i) Gradient imperfection correction, ii) Motion correction, iii) Subspace reconstruction with B0 correction, and iv) Dictionary fitting with FA correction.

Figure 2.

(A) Sequence diagram of the B₁⁺ estimation module.

(B) RF phases of the B-S adiabatic pulses alternated in sign between odd and even TRs.

(C) B₁⁺ and T₂ maps acquired from a uniform phantom, demonstrating that the use of B-S adiabatic pulses yields a more uniform T₂ map.

(D) B₀ estimation module using interleaved TE gaps (ΔTE = 1 ms) between odd and even TRs.

(E) Comparison of B₀ maps obtained using the proposed B₀ estimation module and the standard B₀ mapping sequence

Figure 3.

(A) The waveform of 4 different RF excitation pulses and (B) its corresponding frequency responses.

(C) Zoomed-in the first coefficient maps of subspace reconstruction (top row) and the T₁ maps (bottom row) using different RF excitation pulse.

Spiral trajectory design

In this work, spiral trajectories for both 560-μm and 360-μm isotropic resolutions were generated. For the 560-μm setting, two trajectories were designed using different SRmax of 500 T/m/s and 100 T/m/s, respectively. The latter was intended to mimic the performance of a whole-body gradient system (Siemens 7T MAGNETOM Terra) operating at ~85% of the peripheral nerve stimulation (PNS) threshold. Detailed design parameters are provided in Supporting Figure S2. This work employs a variable density spiral (VDS) design³⁷, where the undersampling factor at the outer k-space is twice that at the k-space center. The MRF data acquired using the 100 T/m/s trajectory, which features slower k-space traversal and smaller gradient-induced imperfections, will be used as a comparison baseline to assess the effectiveness of the proposed gradient imperfection correction methods applied to the faster 500 T/m/s design.

To ensure consistency in scan parameters, particularly the minimum TR used, we maintained the same total duration of ~8.8 ms for the spiral readout plus rewinder and spoiler gradients. Consequently, for the same 560-μm resolution, the trajectory with lower SRmax (100 T/m/s) requires longer durations for rewinder and spoiler gradients, resulting in a need to employ a slightly shorter spiral readout (6.24 ms vs. 7.20 ms for 100 T/m/s and 500 T/m/s, respectively).

Therefore, a higher slew rate enables faster traversal of k-space and a slightly longer spiral readout duration. As a result, each single spiral interleaf is substantially less undersampled. For example, at the center of k-space, the in-plane undersampling factor is 16 for the SRmax of 500 T/m/s, compared with 48 for the SRmax of 100 T/m/s. This indicates that, to cover the same k-space extent, the latter would require three times as many TRs.

To balance SNR and scan efficiency, 24 acquisition groups were scanned for the 560-μm acquisition with slew rate of 500 T/m/s, which was subsequently shown in the Results section to be sufficient for achieving high-quality reconstructions. With the lower slew rate of 100 T/m/s, achieving comparable k-space coverage requires increasing the number of acquisition groups to 72. This resulted in total scan times of 3.5 minutes and 10.5 minutes, respectively (8.7 s per acquisition group).

For the 360-μm design, 168 acquisition groups and 24-minute scan time are implemented. Given that the voxel size at 360-μm isotropic is 3.8 times smaller than at 560-μm isotropic resolution, achieving similar SNR at 360-μm resolution would theoretically require approximately 50 minutes, as SNR scales with the square of the scan time (i.e., 3.8² × 3.5 ≈ 50 minutes). Therefore, two repetitions (2 × 24 minutes) were acquired for 360-μm to achieve a comparable SNR.

2.2. Reconstruction pipeline

In this work, the subspace reconstruction with a locally low-rank (LLR) constraint^39–44 is employed, which can be described as follows:

$${\text{min}}_c{‖\mathit{\text{PFS}}\Phi c-y‖}_2^2+\lambda R_r(c)$$ [1]

where P is the sampling pattern of spiral-projection trajectory, F is the non-uniform Fourier transform (NUFFT), S are the coil sensitivity maps, y are the acquired k-space data, and λ is the regularization parameter for LLR constraint $R_r\left(c\right)$, and Φ is the subspace basis. $c$ are the reconstructed subspace coefficient maps. In this work, λ of 1e-5 was selected for the LLR constraint. Specifically, for each voxel r, a local block of size 8 × 8 × 8 centered around r was extracted using the operator R_r, and each block is reshaped to form a matrix whose columns correspond to the subspace coefficient images. SVD is then applied to each block, followed by soft-thresholding of the singular values according to λ. Further details and the rationale behind the selection of λ and the number of subspace coefficients can be found in Supporting Figure S3 and Figure S4.

To improve reconstruction quality, the following additional components are added to the reconstruction (as shown in Figure 1C):

1. Gradient imperfection correction: Skope field measurements were obtained in a phantom to provide 0th, 1st and 2nd order spatial field information. Significant deviation can be observed in the 0th order likely from B₀ eddy current^45,46 at high slew rates, as well as some minor 1st order deviations, with minimal 2nd order field variations observed. The reconstruction was modified to incorporate 0th and 1st order field information to improve performance. 1st order was incorporated by modifying the k-space trajectory used in the reconstruction, while 0th order was corrected via signal demodulation. For the 0th order, since the vendor’s provided k-space data has already undergone a default B₀ eddy current compensation (B₀ ECC) that is less accurate than what is measured via Skope, this effect has to be first removed prior to applying the updated B₀ eddy current correction. Supporting Figure S5A shows the simulated B₀-eddy-currents phase term, (obtained using Siemens sequence simulation software) and the measured one of one specific TR.

2. Motion correction: The 2-mm motion navigator images are registered using the AFNI toolbox^47–49 to estimate motion parameters. The motion correction strategy in this work is to apply rotation matrix on the k-space trajectory and additional phase modulation on the k-space data to reflect the translation motion³⁰.

3. B₀ correction: Severe B₀ Field inhomogeneities at 7T can lead to image blurring and distortion, even in a highly segmented spiral acquisition of a 6-ms duration. To correct B₀-induced distortion and blurring, the MFI approach⁵⁰ was employed. Due to the severe B₀ inhomogeneity at 7T, 11 demodulation frequencies ${\omega }_{1~11}$ = [−500:100:500] Hz were used to cover the typical ΔB₀ range in the human brain at 7T.

4. FA-corrected dictionary fitting: Through an acquisition of a B₁⁺ map and an expanded dictionary fitting, the effect of B₁⁺ inhomogeneity in creating a spatially varying FA can be accounted for in MRF tissue quantification, as we have performed in prior work^25,51. However, large B₀ inhomogeneity and limited RF excitation bandwidth can also cause a spatially varying FA, which also needs to be accounted for to achieve accurate tissue quantification. As shown in Figure 4C, the pronounced B₀ inhomogeneity at 7T, could reach up to ±300Hz, leading to significant FA attenuation due to the finite bandwidth of the applied RF pulse. Figure 4B shows the frequency response of the SLR pulse, which indicates a FA attenuation of 0.63 at a voxel where the local ΔB₀ is +300Hz, indicated by the red and blue arrows.

Figure 4.

(A) The waveform of the 2.86-ms SLR pulse and (B) its corresponding frequency response.

(C) The ΔB₀ map (Column 1) was used to determine the ΔB₀-induced FA attenuation map (Column 2) via the RF frequency response. Incorporated with the B₁⁺ map (Column 3), the apparent FA map (4th column) could be obtained.

(D) The T₁ (top row) and T₂ (bottom row) maps without any correction (Column 1), with only B₀ MFI correction (Column 2), with both B₀ & B₁+ correction (Column 3), and with B₀ & B₁⁺ & frequency response correction (Column 4).

To correct this FA attenuation effect, this work proposed to use the frequency response of the excitation RF pulse as a look-up table for deriving the FA attenuation map based on the ΔB₀ map. Subsequently, this FA attenuation map (the second column in Figure 4C) was then multiplied with the B₁⁺ map (the third column) to obtain the actual FA factor map (the fourth column). Based on the FA factor maps, the T₁ and T₂ maps were then obtained by matching the reconstructed signal evolutions with the MRF dictionary entries corrected by the corresponding FA factor at each voxel.

After obtaining quantitative maps, a series of contrast-weighted images, including T₁-MPRAGE, T₂-weighted (T₂W), T₂ Fluid Attenuated Inversion Recovery (FLAIR) and Double Inversion Recovery (DIR), were synthesized using a Bloch simulation⁵². It’s worth highlighting that this study employed a local head-only transmission coil. Thus, the spatially-non-selective RF excitation trains in the MRF sequence operated in a similar fashion to slab-selective excitations used in MR angiography (MRA) sequence. As a result, synthesized MRA-like images could be obtained by applying a maximum intensity projection (MIP) to the PD maps.

All computations were conducted on an Ubuntu 20.04 server, equipped with 32 Core i7 Intel Xeon 2.8 GHz CPUs, an Nvidia A6000 GPU, and a memory of 1-TB RAM. The primary softwares utilized for these computations were MATLAB R2022a (The MathWorks, Inc., Natick, MA) and Python 3.8 (Python Software Foundation, Wilmington, DE).

2.3. In-vivo validation

Four healthy volunteers were scanned with approval from the Institutional Review Board. These scans were conducted on a Siemens NexGen 7T MAGNETOM Terra Impulse edition scanner (Erlangen, Germany) using a 96-channel-receive plus 16-channel-transmit coil and a 64-channel-receive plus 8-channel-transmit coil⁵³ (MrCoilTech, Glasgow, UK). Additionally, to validate the stability of the gradient rewinder compensation based on Skope field camera measurements, five scan sessions were conducted over the same six-month period.

3. Results

3.1. Methodological validation

Since this study involves multiple correction methods, this section shows the results of ablation studies for each correction method.

3.1.1. Field inhomogeneity correction

Figure 4D shows the T₁ and T₂ mapping results, subsequent to field inhomogeneity corrections, including B₀, B₁⁺, and frequency response corrections. For the T₁ map, the zoom-in region in the frontal lobe demonstrates the ability of B₀ MFI correction in mitigating image blurring. As expected, B₁⁺ and frequency response corrections have minimal effect on T₁ quantification in this MRF sequence, since it utilizes an adiabatic inversion pulse to create a spatially uniform inversion flip angle. The effects of these corrections are much more significant for T₂ quantification where it can be seen that the frequency response correction is critical in removing large biases.

Figure 2C shows T₂ mapping results from the proposed MRF sequence on a uniform phantom and in an in vivo case, where comparison of B₁⁺ corrections using B₁⁺ maps obtained via i) proposed fast B₁⁺ mapping with adiabatic B-S pulse, ii) proposed mapping but with Fermi-shaped B-S pulse and iii) with Siemens’s TurboFLASH mapping sequence is provided. The use of adiabatic B-S pulse in our proposed fast mapping produces T₂ maps that are most spatially uniform, without observable large spatial biases seen with the other approaches. Figure 2E shows the consistency between the B₀ maps obtained using the proposed B₀ estimation module and those from the standard Siemens B₀ sequence of a much longer scan time.

3.1.2. Fat ring suppression

Figure 3C shows the 1^st coefficient maps from the subspace reconstruction and the T₁ maps obtained from MRF acquisitions with the following RF excitation pulse types: a 2.94-ms hard pulse, 3-ms Gaussian pulse, 5.12-ms Gaussian pulse, and 2.86-ms SLR pulse. As indicated by the red arrows, the 2.94-ms hard pulse exhibits a pronounced fat ring artifact even though it was designed to minimize excitation around −1030 Hz, namely the predominant fat frequency at 7T. This is because the presence of significant B₀ inhomogeneity affects the local resonance frequency of fat causing fat in some areas to still be excited by this RF pulse. The 3-ms Gaussian pulse mitigates some of these fat artifacts (yellow arrows), while the 5.12-ms Gaussian and 2.86-ms SLR pulses further diminish these artifacts, but with SLR pulse being significantly shorter in duration to allow for shorter TR and faster imaging. Importantly, the SLR pulse offers a broader FWHM within the ΔB₀ range, which helps reduce the FA attenuation effect at off-resonance for the water signal.

3.1.3. Gradient imperfection correction

Figure 5 shows the T₂ maps at 560-μm resolution acquired via an MRF acquisition using SR_max of 500 T/m/s where significant gradient imperfections exist. As indicated by the green arrows, the use of gradient rewinder compensation improves the uniformity of the T₂ maps. The pink arrows indicate improved image sharpness that was achieved by incorporating the measured gradient trajectory into the reconstruction. T₁ and T₂ maps obtained using a low SR_max of 100 T/m/s with much reduced gradient imperfections are shown as reference. Figure S5B shows a zoomed-in view of the T₁ and T₂ maps reconstructed under three different conditions: 1) without any B₀ ECC, 2) using the simulated B₀ ECC (vendor’s default setting, obtained using Siemens sequence simulation software), and 3) using the measured B₀ ECC. The measured B₀ ECC shows a better visualization of PVS (perivascular spaces, red arrows), which is less discernible in images reconstructed using the simulated B₀ ECC.

Figure 5.

(A) T₂ maps at 560-μm isotropic resolution using MRF with SR_max = 500 T/m/s: Column 1) without any correction, Column 2) with rewinder compensation, Column 3) with rewinder compensation and measured trajectory correction, compare to Column 4) reference acquisition with SR_max = 100 T/m/s.

(B) and (C) show the zoom-in T₂ maps corresponding to the red and yellow boxes in (A), respectively.

Supporting Figure S1C shows the residual gradient areas across five scan sessions when using the rewinder compensation gradient settings measured from the first scan session. The sum-of-squared residual gradient area of all TRs are listed, which indicates that the gradient compensation could consistently remove 85% ~ 95% of the residual gradient moment compared to the one without implementing rewinder compensation gradient. Figure S1D shows both the measured and designed trajectories. When compared to the first session, the subsequent four sessions exhibit notable consistency. This demonstrated good stability of the gradient hardware, indicating that a single Skope calibration measurement can yield compensation parameters that remain valid for at least several months.

3.2. Comparison to baseline

Figure 6 shows results exhibiting how the proposed refinements and modifications to the MRF acquisition and reconstruction have enabled 560-μm whole-brain T₁ and T₂ mapping to be performed 3 times faster with significantly less biases compared to current state-of-the-art method. As a baseline, the first column in Figure 6B and 6C shows T₁ and T₂ maps in sagittal and coronal views from an MRF acquisition performed using conventional body gradient specifications (slew rate limited to ~100 T/m/s due to PNS constraint) along with Siemens’s standard B₁⁺ and B₀ mapping sequences, resulting in a total scan time of 12 minutes and 26 seconds. Such an acquisition should not contain significant gradient imperfections, but the resulting T₂ map still contains large spatial biases from i) not employing the proposed RF frequency response correction (red arrow), and ii) imperfect B₁⁺ correction from biases in the B₁⁺ map from the standard B₁⁺ mapping sequence (yellow arrows). The second column of Figure 6B and 6C shows the results from the same setup, but with three times the reduction in MRF k-space encoding, thereby facilitating a three-fold reduction in scan time, with additional aliasing artifacts observed (pink arrow). The third column shows that such aliasing artifacts can be overcome via faster k-space traversal with a higher slew rate of 500 T/m/s, but at a cost of added T₂ biases and blurring from gradient imperfections and incomplete gradient rewinding (green arrows). The last column shows results obtained using an MRF acquisition at 500 T/m/s with gradient rewinder compensation and fast B₁⁺ and B₀ mapping, along with all the proposed reconstruction modifications. Here, high quality T₁ and T₂ maps can be obtained in just 3 minutes 44 seconds. Additional slices and orientations are presented in Supporting Figure S6.

Figure 6.

(A) The proposed method enables quantitative imaging at 560-μm resolution with a total acquisition time of less than 4 minutes on the NexGen 7T with Impulse gradient, which is more than three times faster than conventional gradient systems combined with product field mapping sequences.

(B) Without the proposed correction methods, even a 12-minute acquisition leads to noticeable T₂ bias, caused by either uncorrected RF frequency response (red arrows), or inaccurate B₁⁺ estimation (yellow arrows). Additionally, simply using a higher acceleration factor to match the same acquisition time introduces severe undersampling artifacts (pink arrow), while increasing slew rate without employing the proposed compensation methods leads to significant T₂ shading caused by imperfect signal refocusing.

(C) Coronal view further reveals T₂ bias caused by inaccurate B₁⁺ estimation, as indicated by the yellow arrow.

3.3. 360-μm MRF

3.3.1. motion correction

Figure 7 shows comparisons of quantitative maps, with and without applying motion correction, for both one-repetition and two-repetitions scans at 360-μm isotropic resolution across 24 and 48 minutes respectively. It can be observed that even for a dedicated healthy volunteer, the prolonged scan time inevitably leads to observable motion. The maximum translation is only around 0.5 mm (Figure 7A), which may not cause noticeable issues at lower resolution, however, this translation corresponds to a displacement of approximately 1.4 pixels at the 360-μm resolution. The reconstructed quantitative maps with and without motion correction (Figure 7B) validated that even such small motion can lead to observable blurring, making subtle structures (PVS as indicated by red arrows) invisible. Notably, the motion-induced blurring is more pronounced for the two-repetitions scan, highlighting the necessity of motion correction in this case.

Figure 7.

(A) Motion estimation (rotation and translation) of a 2-repetition scan (48 minutes) using the 2-mm navigators every 8.7 s.

(B) One-repetition and two-repetitions quantitative maps at 360-μm isotropic resolution, reconstructed with and without motion correction (MOCO). With motion correction, the subtle structures, such as PVS (red arrows), become more clearly visible.

3.3.2. Quantitative maps and synthesis contrasts

Figure 8 presents three orthogonal views of the T₁ and T₂ maps at 360-μm of a healthy volunteer, reconstructed using all the correction methods described above. The hippocampus, highlighted within the red boxes, is used to demonstrate the image quality of these high-resolution quantitative images, where the complex internal structures of the hippocampus are clearly visible. Black arrows indicate other subtle structures, such as the optic radiations (Figure 8A), cerebellar vermis (Figure 8B), and pontes grisei (caudato-lenticular gray matter bridges, Figure 8C). Additionally, as indicated by the yellow arrows, some nucleus exhibit distinct differences from their surrounding regions in the T₂ maps, whereas these differences are not as easily distinguishable in the T₁ maps. In addition, Supporting Video S1 provides dynamic visualizations of T₁, T₂, and PD maps across three orthogonal views for all four subjects. The corresponding quantitative values of T₁ and T₂ for WM and GM are reported in Supporting Table S1, along with reference values from previous studies^{14,15,18,19,53,54}.

Figure 8.

Whole-brain T₁ and T₂ maps at 360-μm isotropic resolution. The red box highlights the hippocampus, where the red arrow indicates the exceptional image quality and structural detail achieved in the ultra-high-resolution quantitative maps. Black arrows, from top to bottom, point to the optic radiations, cerebellar vermis, and pontes grisei, while the yellow arrows point to deep-brain nucleis.

Figure 9 shows the synthesized T₁ MPRAGE, T₂W, T₂ FLAIR, and DIR images at 360-μm isotropic resolution, where high quality results with detailed structural information can be observed. Furthermore, we also observed that, with the head-only RF transmission coil, simply applying MIP to the proton density map can achieve a MRA-like image, which provides additional information beyond conventional T₁ and T₂ contrasts.

Figure 9.

Contrast-weighted images are synthesized from the quantitative maps using the proposed method, including T₁ MPRAGE, T₂W, DIR, FLAIR and MRA-like.

Figure 10A and 10B show the T₁ and T₂ maps of the slice containing the substantia nigra (SN) and subthalamic nucleus (STN). It can be observed that the T₁ values of SN and STN (critical for the early diagnosis and intervention of Parkinson disease) are very similar to those of the surrounding white matter, making them difficult to distinguish on T₁ maps. However, their T₂ values show a significant difference against the surrounding white matter, where in the zoom-in sub-figure with compressed color scaling, it can be seen that STN exhibits a slightly higher T₂ value than SN. As a result, pure T₁-weighted imaging, such as T₁ MPRAGE (Figure 10C), fails to differentiate these two structures, whereas incorporating T₂ weighting (e.g., through introducing T₂ preparation, Figure 10D) allows for clearer distinction and segmentation of these structures.

Figure 10.

T₁ (A) and T₂ (B) maps using the proposed MRF acquisition. The STN (red arrow) exhibits slightly higher T₂ values than the SN (black arrow), while their T₁ values remain similar to surrounding white matter. A synthesized T₁-MPRAGE image (C) demonstrates poor contrast between STN and SN. Incorporating T₂ weighting (e.g., by adding a T₂ preparation module) could improve differentiation between these two nuclei (D).

Discussion

In this study, we successfully developed a tailored MRF acquisition and reconstruction framework optimized for a high-performance, head-only 7T scanner. Leveraging the system’s superior gradient performance, we achieved whole-brain quantitative imaging at 560-μm isotropic resolution with a scan time under 4 minutes to allow it to be easily included in many neuroscientific and clinical research studies. Furthermore, this work marks the first demonstration of a whole-brain MRF scan at 360-μm isotropic resolution. These results push the spatial resolution boundaries of in-vivo human brain imaging and provide a practical pathway for ultra-high-resolution quantitative mapping. In addition, using the example of brain nucleus SN and STN, we demonstrated that the combined use of T₁ and T₂ can facilitate the differentiation of these brain structures. This suggests that performing multi-parametric quantitative imaging can help optimize conventional contrast-weighted sequences to maximize contrast for specific structures of interest, thereby enabling more accurate structural delineation and segmentation.

Ablation studies highlighted the necessity of several key modifications to both the acquisition and reconstruction processes in order to achieve reliable high-resolution imaging at 7T and high-performance gradient system. The two major innovations of correcting FA attenuation and gradient rewinder compensation prove important for obtaining high-quality T₂ maps. While this correction was developed for 7T, retrospective analysis suggests it could also be beneficial at 3T, where moderate FA attenuation effects are still present in certain regions, such as the front lobe of brain and in body applications with severe B₀ inhomogeneity. Additionally, we implemented gradient rewinder compensation to mitigate signal dephasing caused by imperfect gradient moment nulling in FISP-based sequences. This correction, implemented with a small blip gradient derived from Skope field camera measurements, effectively preserves the stimulated echo and improves the accuracy of T₂ estimation. The same principle could be extended to other sequences such as TSE, SPACE, or GRASE where echo refocusing is critical. The compensation gradient could be integrated directly into the gradient crusher or rewinder to avoid incurring an increased in echo-spacing.

Despite the multitude of the corrections/mitigation approaches involved, the proposed MRF pipeline has been designed for seamless usability. B₁⁺ and B₀ mapping modules were embedded directly within the MRF sequence to minimize inter-scan motion and streamline workflow. Given that the B₁⁺ and B₀ fields are inherently smooth, ultra-high-resolution field maps offer limited practical value. Moreover, a joint estimation of B₁⁺ and B₀ fields together with tissue parameters in the MRF framework is feasible, but would require modifications to our MRF acquisition, to increase its sensitivity to these fields and enable them to be estimated accurately. This would significantly lengthen the overall acquisition. Therefore, we adopted dedicated, rapid, low-resolution B₁⁺ and B₀ estimation modules instead. Furthermore, motion correction is integrated into the acquisition, ensuring motion robustness in routine applications, including studies involving motion-prone populations. Skope-based measurements and corresponding correction strategies demonstrated good stability, indicating that a single phantom-based calibration session may remain valid over extended periods without the need for frequent recalibration. These characteristics should make the fast 560-μm MRF protocol a promising tool that can be widely deployed in neuroscience research, enabling reproducible and high-resolution brain mapping in a time-efficient manner.

The need for accurate B₁⁺ mapping to achieve high fidelity T₁ and T₂ mapping was demonstrated, where we achieved this using a rapid B₁⁺ mapping sequence that makes use of B-S pulse. Here, the use of adiabatic design rather than conventional Fermi design of B-S pulse enables short duration pulses to be employed within the SAR constraint to achieve high accuracy in B₁⁺ mapping. This work has demonstrated the effectiveness of the proposed B₁⁺ correction. However, in areas where B₁⁺ is very low, such as in the lower part of the brain, the MRF’s FA train is at markedly lower flip angles, making T₂ quantification more difficult even when the actual flip angle is accounted for. On the other hand, a 20-ms adiabatic tanh pulse was employed for inversion preparation with the assumption that it would be relatively insensitive to B₁⁺ variations. However, as in extreme cases (for example in cerebellum), insufficient inversion efficiency due to B₁⁺ may lead to incomplete Mz inversion and faster-than-expected recovery, resulting in underestimation of the T₁ values. Future work would explore the use of pTx pulse design^8,55–57, such as through k-t pulse^58–60 or GRAPE pulse⁶¹ design, to provide more uniform excitation. In particular, an interleaved binomial kT-point design⁶² for water-selective imaging could be a good fit here. Considering that conventional pTx techniques often require complex B₁⁺ field calibration procedures and per subject pulse calculation, adoption of plug-and-play⁶³ or semi–plug-and-play⁶⁴ pTx solutions could also help avoid adding complexity to our MRF acquisition.

Despite applying both B₁⁺ and RF frequency response corrections, accurate T₁ and T₂ estimations remain challenging in regions with substantial blood flow, such as in the venous structures, due to spin history effects⁶⁵. Addressing this issue may require physical modeling of flow-related signal evolution. Furthermore, inaccuracies in the estimated T₁ and T₂ values can affect the PD quantification, since PD is typically computed after the dictionary matching process of MRF by dividing the sum of squared magnitudes of the acquired signal evolutions over the 500 TRs at each voxel by that of the corresponding matched dictionary entries⁵. If the estimated T₁ and T₂ are incorrect, the associated matched dictionary entry will also be incorrect, leading to a mismatch in the expected signal evolution and an altered sum of squared values, thereby introducing additional bias into the PD estimation. This bias in PD can subsequently impact synthesized MRA-like images obtained via MIP. Therefore, the MRA-like images presented in this study should not be interpreted as precise vascular measurements but rather as a proof-of-concept exploration of additional information that may be extracted from the MRF datasets. On the other hand, due to the influence of the receiver coil sensitivity profile, the PD maps often exhibit peripheral-to-central signal shading, which may in turn compromise the accuracy of synthesized MRF-like images. This issue could potentially be mitigated through the application of appropriate correction techniques⁶⁶. The data acquisition strategy developed in this study enables rapid and robust k-space encoding. However, at ultra-high-resolution such as 360-μm isotropic, the MRF acquisition can become more motion limited which we have dealt with, and SNR-limited rather than image-encoding speed limited. Therefore, an important future direction is to integrate the proposed acquisition approach with advanced denoising or regularization techniques for ultra-high resolution. Unrolled deep learning-based reconstruction⁶⁷ represents a promising avenue to pursue. However, further development is required, particularly in addressing the challenges associated with training such large-scale reconstruction problems. Methods such as GLEAM⁶⁸ and related frameworks may provide useful solutions in this context.

Conclusion

This study presents a fast and robust MRF framework tailored for a head-only high-performance NexGen 7T scanner, enabling whole-brain quantitative imaging at 560-μm isotropic resolution in under 4 minutes and achieving the first whole-brain in-vivo 360-μm multi-parametric maps. Key innovations include embedded fast B₀ and B₁⁺ mapping, gradient rewinder compensation, RF frequency response correction, and redesigned water-only RF pulse. These methods address challenges of field inhomogeneities, gradient imperfections, and low SNR at ultra-high resolution. Field estimation and motion estimation are seamlessly integrated into the MRF framework, ensuring workflow efficiency. This method not only provides a rapid, high-resolution quantitative imaging approach that enables the application of meso-scale quantitative imaging in clinical and neuroscience research, but also pushes higher the current resolution limits of multi-parametric quantitative MRI, offering new possibilities for studying quantitative changes of fine-scale structure in the human brain.