SmartPulse, a machine learning approach for calibration‐free dynamic RF shimming: Preliminary study in a clinical environment

Artificial intelligence for neuro MRI acquisition: a review

OBJECT

To review recent advances of artificial intelligence (AI) in enhancing the efficiency and throughput of the MRI acquisition workflow in neuroimaging, including planning, sequence design, and correction of acquisition artifacts.

MATERIALS AND METHODS

A comprehensive analysis was conducted on recent AI-based methods in neuro MRI acquisition. The study focused on key technological advances, their impact on clinical practice, and potential risks associated with these methods.

RESULTS

The findings indicate that AI-based algorithms have a substantial positive impact on the MRI acquisition process, improving both efficiency and throughput. Specific algorithms were identified as particularly effective in optimizing acquisition steps, with reported improvements in workflow efficiency.

DISCUSSION

The review highlights the transformative potential of AI in neuro MRI acquisition, emphasizing the technological advances and clinical benefits. However, it also discusses potential risks and challenges, suggesting areas for future research to mitigate these concerns and further enhance AI integration in MRI acquisition.

Unsupervised deep learning with convolutional neural networks for static parallel transmit design: A retrospective study

PURPOSE

To mitigateB 1 + $$ {B}_1^{+} $$ inhomogeneity at 7T for multi-channel transmit arrays using unsupervised deep learning with convolutional neural networks (CNNs).

METHODS

Deep learning parallel transmit (pTx) pulse design has received attention, but such methods have relied on supervised training and did not use CNNs for multi-channelB 1 + $$ {B}_1^{+} $$ maps. In this work, we introduce an alternative approach that facilitates the use of CNNs with multi-channelB 1 + $$ {B}_1^{+} $$ maps while performing unsupervised training. The multi-channelB 1 + $$ {B}_1^{+} $$ maps are concatenated along the spatial dimension to enable shift-equivariant processing amenable to CNNs. Training is performed in an unsupervised manner using a physics-driven loss function that minimizes the discrepancy of the Bloch simulation with the target magnetization, which eliminates the calculation of reference transmit RF weights. The training database comprises 3824 2D sagittal, multi-channelB 1 + $$ {B}_1^{+} $$ maps of the healthy human brain from 143 subjects.B 1 + $$ {B}_1^{+} $$ data were acquired at 7T using an 8Tx/32Rx head coil. The proposed method is compared to the unregularized magnitude least-squares (MLS) solution for the target magnetization in static pTx design.

RESULTS

The proposed method outperformed the unregularized MLS solution for RMS error and coefficient-of-variation and had comparable energy consumption. Additionally, the proposed method did not show local phase singularities leading to distinct holes in the resulting magnetization unlike the unregularized MLS solution.

CONCLUSION

Proposed unsupervised deep learning with CNNs performs better than unregularized MLS in static pTx for speed and robustness.

Whole-liver flip-angle shimming at 7 T using parallel-transmit kT -point pulses and Fourier phase-encoded DREAM B 1 + mapping

PURPOSE

To obtain homogeneous signal throughout the human liver at 7 T. Flip angle (FA) shimming in 7T whole-liver imaging was performed through parallel-transmit kT -point pulses based on subject-specific multichannel absoluteB 1 + $$ {\mathrm{B}}_1^{+} $$ maps from Fourier phase-encoded dual refocusing echo acquisition mode (PE-DREAM).

METHODS

The optimal number of Fourier phase-encoding steps for PE-DREAMB 1 + $$ {\mathrm{B}}_1^{+} $$ mapping was determined for a 7T eight-channel parallel-transmission system. FA shimming experiments were performed in the liver of 7 healthy subjects with varying body mass index. In these subjects, firstB 0 $$ {\mathrm{B}}_0 $$ shimming and Fourier PE-DREAMB 1 + $$ {\mathrm{B}}_1^{+} $$ mapping were performed. Subsequently, three small-flip-angle 3D gradient-echo scans were acquired, comparing a circularly polarized (CP) mode, a phase shim, and a kT -point pulse. Resulting homogeneity was assessed and compared with estimated FA maps and distributions.

RESULTS

Fourier PE-DREAM with 13 phase-encoding steps resulted in a good tradeoff betweenB 1 + $$ {\mathrm{B}}_1^{+} $$ accuracy and scan time. Lower coefficient of variation values (average [min-max] across subjects) of the estimated FA in the volume of interest were observed using kT -points (7.4 [6.6%-8.0%]), compared with phase shimming (18.8 [12.9%-23.4%], p < 0.001) and CP (43.2 [39.4%-47.1%], p < 0.001). kT -points delivered whole-liver images with the nominal FA and the highest degree of homogeneity. CP and phase shimming resulted in either inaccurate or imprecise FA distributions. Here, locations having suboptimal FA in the estimated FA maps corresponded to liver areas suffering from inconsistent signal intensity and T1 -weighting in the gradient-echo scans.

CONCLUSION

Homogeneous whole-liver 3D gradient-echo acquisitions at 7 T can be obtained with eight-channel kT -point pulses calculated based on subject-specific multichannel absolute Fourier PE-DREAMB 1 + $$ {\mathrm{B}}_1^{+} $$ maps.

The effects of RF coils and SAR supervision strategies for clinically applicable nonselective parallel-transmit pulses at 7 T

PURPOSE

To investigate the effects of using different parallel-transmit (pTx) head coils and specific absorption rate (SAR) supervision strategies on pTx pulse design for ultrahigh-field MRI using a 3D-MPRAGE sequence.

METHODS

The PTx universal pulses (UPs) and fast online-customized (FOCUS) pulses were designed with pre-acquired data sets (B₀ , B₁ ⁺ maps, specific absorption rate [SAR] supervision data) from two different 8 transmit/32 receive head coils on two 7T whole-body MR systems. For one coil, the SAR supervision model consisted of per-channel RF power limits. In the other coil, SAR estimations were done with both per-channel RF power limits as well as virtual observation points (VOPs) derived from electromagnetic field (EMF) simulations using three virtual human body models at three different positions. All pulses were made for nonselective excitation and inversion and evaluated on 132 B₀ , B₁ ⁺ , and SAR supervision datasets obtained with one coil and 12 from the other. At both sites, 3 subjects were examined using MPRAGE sequences that used UP/FOCUS pulses generated for both coils.

RESULTS

For some subjects, the UPs underperformed when simulated on a different coil from which they were derived, whereas FOCUS pulses still showed acceptable performance in that case. FOCUS inversion pulses outperformed adiabatic pulses when scaled to the same local SAR level. For the self-built coil, the use of VOPs showed reliable overestimation compared with the ground-truth EMF simulations, predicting about 52% lower local SAR for inversion pulses compared with per-channel power limits.

CONCLUSION

FOCUS inversion pulses offer a low-SAR alternative to adiabatic pulses and benefit from using EMF-based VOPs for SAR estimation.

Ultra-high field MRI: parallel-transmit arrays and RF pulse design

This paper reviews the field of multiple or parallel radiofrequency (RF) transmission for magnetic resonance imaging (MRI). Currently the use of ultra-high field (UHF) MRI at 7 tesla and above is gaining popularity, yet faces challenges with non-uniformity of the RF field and higher RF power deposition. Since its introduction in the early 2000s, parallel transmission (pTx) has been recognized as a powerful tool for accelerating spatially selective RF pulses and combating the challenges associated with RF inhomogeneity at UHF. We provide a survey of the types of dedicated RF coils used commonly for pTx and the important modeling of the coil behavior by electromagnetic (EM) field simulations. We also discuss the additional safety considerations involved with pTx such as the specific absorption rate (SAR) and how to manage them. We then describe the application of pTx with RF pulse design, including a practical guide to popular methods. Finally, we conclude with a description of the current and future prospects for pTx, particularly its potential for routine clinical use.

Parallel transmission 2D RARE imaging at 7T with transmit field inhomogeneity mitigation and local SAR control

PURPOSE

We develop and test a parallel transmit (pTx) pulse design framework to mitigate transmit field inhomogeneity with control of local specific absorption rate (SAR) in 2D rapid acquisition with relaxation enhancement (RARE) imaging at 7T.

METHODS

We design large flip angle RF pulses with explicit local SAR constraints by numerical simulation of the Bloch equations. Parallel computation and analytical expressions for the Jacobian and the Hessian matrices are employed to reduce pulse design time. The refocusing-excitation "spokes" pulse pairs are designed to satisfy the Carr-Purcell-Meiboom-Gill (CPMG) condition using a combined magnitude least squares-least squares approach.

RESULTS

In a simulated dataset, the proposed approach reduced peak local SAR by up to 56% for the same level of refocusing uniformity error and reduced refocusing uniformity error by up to 59% (from 32% to 7%) for the same level of peak local SAR compared to the circularly polarized birdcage mode of the pTx array. Using explicit local SAR constraints also reduced peak local SAR by up to 46% compared to an RF peak power constrained design. The excitation and refocusing uniformity error were reduced from 20%-33% to 4%-6% in single slice phantom experiments. Phantom experiments demonstrated good agreement between the simulated excitation and refocusing uniformity profiles and experimental image shading.

CONCLUSION

PTx-designed excitation and refocusing CPMG pulse pairs can mitigate transmit field inhomogeneity in the 2D RARE sequence. Moreover, local SAR can be decreased significantly using pTx, potentially leading to better slice coverage, enabling larger flip angles or faster imaging.

Universal pulses for homogeneous excitation using single channel coils

PURPOSE

Universal Pulses (UPs) are excitation pulses that reduce the flip angle inhomogeneity in high field MRI systems without subject-specific optimization, originally developed for parallel transmit (PTX) systems at 7 T. We investigated the potential benefits of UPs for single channel (SC) transmit systems at 3 T, which are widely used for clinical and research imaging, and for which flip angle inhomogeneity can still be problematic.

METHODS

SC-UPs were designed using a spiral nonselective k-space trajectory for brain imaging at 3 T using transmit field maps (B1+) and off-resonance maps (B0) acquired on two different scanner types: a 'standard' single channel transmit system and a system with a PTX body coil. The effect of training group size was investigated using data (200 subjects) from the standard system. The PTX system was used to compare SC-UPs to PTX-UPs (15 subjects). In two additional subjects, prospective imaging using SC-UP was studied.

RESULTS

Average flip angle homogeneity error fell from 9.5 ± 0.5 % for 'default' excitation to 3.0 ± 0.6 % using SC-UPs trained over 50 subjects. Performance of the UPs was found to steadily improve as training group size increased, but stabilized after ~15 subjects. On the PTX-enabled system, SC-UPs again outperformed default excitation in simulations (4.8 ± 0.6 % error versus 10.6 ± 0.8 % respectively) though greater homogenization could be achieved with PTX-UPs (3.9 ± 0.6 %) and personalized pulses (SC-PP 3.6 ± 1.0 %, PTX-PP 2.9 ± 0.6 %). MP-RAGE imaging using SC-UP resulted in greater separation between grey and white matter signal intensities than default excitation.

CONCLUSIONS

SC-UPs can improve excitation homogeneity in standard 3 T systems without further calibration and could be used instead of a default excitation pulse for nonselective neuroimaging at 3 T.

Neuroimaging at 7 Tesla: a pictorial narrative review

Neuroimaging using the 7-Tesla (7T) human magnetic resonance (MR) system is rapidly gaining popularity after being approved for clinical use in the European Union and the USA. This trend is the same for functional MR imaging (MRI). The primary advantages of 7T over lower magnetic fields are its higher signal-to-noise and contrast-to-noise ratios, which provide high-resolution acquisitions and better contrast, making it easier to detect lesions and structural changes in brain disorders. Another advantage is the capability to measure a greater number of neurochemicals by virtue of the increased spectral resolution. Many structural and functional studies using 7T have been conducted to visualize details in the white matter and layers of the cortex and hippocampus, the subnucleus or regions of the putamen, the globus pallidus, thalamus and substantia nigra, and in small structures, such as the subthalamic nucleus, habenula, perforating arteries, and the perivascular space, that are difficult to observe at lower magnetic field strengths. The target disorders for 7T neuroimaging range from tumoral diseases to vascular, neurodegenerative, and psychiatric disorders, including Alzheimer's disease, Parkinson's disease, multiple sclerosis, epilepsy, major depressive disorder, and schizophrenia. MR spectroscopy has also been used for research because of its increased chemical shift that separates overlapping peaks and resolves neurochemicals more effectively at 7T than a lower magnetic field. This paper presents a narrative review of these topics and an illustrative presentation of images obtained at 7T. We expect 7T neuroimaging to provide a new imaging biomarker of various brain disorders.

Rigid motion-resolved B1+ prediction using deep learning for real-time parallel-transmission pulse design

PURPOSE

Tailored parallel-transmit (pTx) pulses produce uniform excitation profiles at 7 T, but are sensitive to head motion. A potential solution is real-time pulse redesign. A deep learning framework is proposed to estimate pTx B 1 + distributions following within-slice motion, which can then be used for tailored pTx pulse redesign.

METHODS

Using simulated data, conditional generative adversarial networks were trained to predict B 1 + distributions in the head following a displacement. Predictions were made for two virtual body models that were not included in training. Predicted maps were compared with ground-truth (simulated, following motion) B1 maps. Tailored pTx pulses were designed using B1 maps at the original position (simulated, no motion) and evaluated using simulated B1 maps at displaced position (ground-truth maps) to quantify motion-related excitation error. A second pulse was designed using predicted maps (also evaluated on ground-truth maps) to investigate improvement offered by the proposed method.

RESULTS

Predicted B 1 + maps corresponded well with ground-truth maps. Error in predicted maps was lower than motion-related error in 99% and 67% of magnitude and phase evaluations, respectively. Worst-case flip-angle normalized RMS error due to motion (76% of target flip angle) was reduced by 59% when pulses were redesigned using predicted maps.

CONCLUSION

We propose a framework for predicting B 1 + maps online with deep neural networks. Predicted maps can then be used for real-time tailored pulse redesign, helping to overcome head motion-related error in pTx.

Mitigating transmit‐B 1 artifacts by predicting parallel transmission images with deep learning: A feasibility study using high‐resolution whole‐brain diffusion at 7 Tesla

Purpose

To propose a novel deep learning (DL) approach to transmit‐B₁ (B₁⁺)‐artifact mitigation without direct use of parallel transmission (pTx), by predicting pTx images from single‐channel transmission (sTx) images.

Methods

A deep encoder–decoder convolutional neural network was constructed and trained to learn the mapping from sTx to pTx images. The feasibility was demonstrated using 7 T Human‐Connectome Project (HCP)‐style diffusion MRI. The training dataset comprised images acquired on 5 healthy subjects using commercial Nova RF coils. Relevant hyperparameters were tuned with a nested cross‐validation, and the generalization performance evaluated using a regular cross‐validation.

Results

Our DL method effectively improved the image quality for sTx images by restoring the signal dropout, with quality measures (including normalized root‐mean‐square error, peak SNR, and structural similarity index measure) improved in most brain regions. The improved image quality was translated into improved performances for diffusion tensor imaging analysis; our method improved accuracy for fractional anisotropy and mean diffusivity estimations, reduced the angular errors of principal eigenvectors, and improved the fiber orientation delineation relative to sTx images. Moreover, the final DL model trained on data of all 5 subjects was successfully used to predict pTx images for unseen new subjects (randomly selected from the 7 T HCP database), effectively recovering the signal dropout and improving color‐coded fractional anisotropy maps with largely reduced noise levels.

Conclusion

The proposed DL method has potential to provide images with reduced B1⁺ artifacts in healthy subjects even when pTx resources are inaccessible on the user side.

B1 field map synthesis with generative deep learning used in the design of parallel-transmit RF pulses for ultra-high field MRI

Spoke trajectory parallel transmit (pTX) excitation in ultra-high field MRI enables B₁⁺ inhomogeneities arising from the shortened RF wavelength in biological tissue to be mitigated. To this end, current RF excitation pulse design algorithms either employ the acquisition of field maps with subsequent non-linear optimization or a universal approach applying robust pre-computed pulses. We suggest and evaluate an intermediate method that uses a subset of acquired field maps combined with generative machine learning models to reduce the pulse calibration time while offering more tailored excitation than robust pulses (RP). The possibility of employing image-to-image translation and semantic image synthesis machine learning models based on generative adversarial networks (GANs) to deduce the missing field maps is examined. Additionally, an RF pulse design that employs a predictive machine learning model to find solutions for the non-linear (two-spokes) pulse design problem is investigated. As a proof of concept, we present simulation results obtained with the suggested machine learning approaches that were trained on a limited data-set, acquired in vivo. The achieved excitation homogeneity based on a subset of half of the B₁⁺ maps acquired in the calibration scans and half of the B₁⁺ maps synthesized with GANs is comparable with state of the art pulse design methods when using the full set of calibration data while halving the total calibration time. By employing RP dictionaries or machine-learning RF pulse predictions, the total calibration time can be reduced significantly as these methods take only seconds or milliseconds per slice, respectively.

Standardized universal pulse: A fast RF calibration approach to improve flip angle accuracy in parallel transmission

PURPOSE

In parallel transmission (pTX), subject-tailored RF pulses allow achieving excellent flip angle (FA) accuracy but often require computationally extensive online optimizations, precise characterization of the static field ( Δ B 0 ), and the transmit RF field ( B 1 + ) distributions. This costs time and requires expertise from the MR user. Universal pulses (UPs) have been proposed to reduce this burden, yet, with a penalty in FA accuracy. This study introduces the concept of standardized universal pulses (SUPs), where pulses are designed offline and adjusted to the subject through a fast online calibration scan.

METHODS

A SUP is designed offline using a so-called standardized database, wherein each B 1 + map has been normalized to a reference transmit RF field distribution. When scanning a new subject, a 3-slice B 1 + acquisition (scan time < 10  s) is performed and used to adjust the SUP to the subject through a linear transform. SUP performance was assessed at 7T with simulations by computing the FA-normalized root mean square error (FA-NRMSE) and the FA pattern stability as measured by the average and coefficient of variation of the FA across 15 control subjects, along with in vivo experiments using an MP2RAGE sequence implementing the SUP variant for the FLASH readout.

RESULTS

Adjusted SUP improved the FA-NRMSE (8.8 % for UP vs. 7.1 % for adjusted SUP). Experimentally in vivo, this translated in an improved signal homogeneity and more accurate T 1 quantification using MP2RAGE.

CONCLUSION

The proposed SUP approach improves excitation accuracy (FA-NRMSE) while preserving the same offline pulse design principle as offered by UPs.

Calibration-free regional RF shims for MRS

PURPOSE

Achieving a desired RF transmit field ( B 1 + ) in small regions of interest is critical for single-voxel MRS at ultrahigh field. Radio-frequency (RF) shimming, using parallel transmission, requires B 1 + mapping and optimization, which limits its ease of use. This work aimed to generate calibration-free RF shims for predefined target regions of interest, which can be applied to any participant, to produce a desired absolute magnitude B 1 + (| B 1 + |).

METHODS

The RF shims were found offline by joint optimization on a database comprising B 1 + maps from 11 subjects, considering regions of interest in occipital cortex, hippocampus and posterior cingulate, as well as whole brain. The | B 1 + | achieved was compared with a tailored shimming approach, and MR spectra were acquired using tailored and calibration-free shims in 4 participants. Global and local 10g specific-absorption-rate deposition were estimated using Duke and Ella dielectric models.

RESULTS

There was no difference in the mean | B 1 + | produced using calibration-free versus tailored RF shimming in the occipital cortex (p = .15), hippocampus (p = .5), or posterior cingulate (p = .98), although differences were observed in the RMS error | B 1 + |. Spectra acquired using calibration-free shims had similar SNR and low residual water signal. Under identical power settings, specific-absorption-rate deposition was lower compared with operating in quadrature mode. For example, the total head specific absorption rate was around 35% less for the occipital cortex.

CONCLUSION

This work demonstrates that static RF shims, optimized offline for small regions, avoid the need for B 1 + mapping and optimization for each region of interest and participant. Furthermore, power settings may be increased when using calibration-free shims, to better take advantage of RF shimming.

DeepControl: 2DRF pulses facilitating B 1 + inhomogeneity and B0 off-resonance compensation in vivo at 7 T

PURPOSE

Rapid 2DRF pulse design with subject-specific B 1 + inhomogeneity and B₀ off-resonance compensation at 7 T predicted from convolutional neural networks is presented.

METHODS

The convolution neural network was trained on half a million single-channel transmit 2DRF pulses optimized with an optimal control method using artificial 2D targets, B 1 + and B₀ maps. Predicted pulses were tested in a phantom and in vivo at 7 T with measured B 1 + and B₀ maps from a high-resolution gradient echo sequence.

RESULTS

Pulse prediction by the trained convolutional neural network was done on the fly during the MR session in approximately 9 ms for multiple hand-drawn regions of interest and the measured B 1 + and B₀ maps. Compensation of B 1 + inhomogeneity and B₀ off-resonances has been confirmed in the phantom and in vivo experiments. The reconstructed image data agree well with the simulations using the acquired B 1 + and B₀ maps, and the 2DRF pulse predicted by the convolutional neural networks is as good as the conventional RF pulse obtained by optimal control.

CONCLUSION

The proposed convolutional neural network-based 2DRF pulse design method predicts 2DRF pulses with an excellent excitation pattern and compensated B 1 + and B₀ variations at 7 T. The rapid 2DRF pulse prediction (9 ms) enables subject-specific high-quality 2DRF pulses without the need to run lengthy optimizations.

Fast online-customized (FOCUS) parallel transmission pulses: A combination of universal pulses and individual optimization

PURPOSE

To mitigate spatial flip angle (FA) variations under strict specific absorption rate (SAR) constraints for ultra-high field MRI using a combination of universal parallel transmit (pTx) pulses and fast subject-specific optimization.

METHODS

Data sets consisting of B₀ , B 1 + maps, and virtual observation point (VOP) data were acquired from 72 subjects (study groups of 48/12 healthy Europeans/Asians and 12 Europeans with pathological or incidental findings) using an 8Tx/32Rx head coil on a 7T whole-body MR system. Combined optimization values (COV) were defined as combination of spiral-nonselective (SPINS) trajectory parameters and an energy regularization weight. A set of COV was optimized universally by simulating the individual RF pulse optimizations of 12 training data sets (healthy Europeans). Subsequently, corresponding universal pulses (UPs) were calculated. Using COV and UPs, individually optimized pulses (IOPs) were calculated during the sequence preparation phase (maximum 15 s). Two different UPs and IOPs were evaluated by calculating their normalized root-mean-square error (NRMSE) of the FA and SAR in simulations of all data sets. Seven additional subjects were examined using an MPRAGE sequence that uses the designed pTx excitation pulses and a conventional adiabatic inversion.

RESULTS

All pTx pulses resulted in decreased mean NRMSE compared to a circularly polarized (CP) pulse (CP = ~28%, UPs = ~17%, and IOPs = ~12%). UPs and IOPs improved homogeneity for all subjects. Differences in NRMSE between study groups were much lower than differences between different pulse types.

CONCLUSION

UPs can be used to generate fast online-customized (FOCUS) pulses gaining lower NRMSE and/or lower SAR values.