Over the last four decades, there have been various evolutions in the design and development of coils, from volume coils to the recent introduction of wireless receive arrays. A recent aim has been to develop coils that can closely conform to the anatomy of interest to increase the acquired signal. This goal has given rise to designs ranging from adaptable transmit coils to on-body stretchable receive arrays made using fabric or elastomer substrates. This review covers the design, fabrication details, experimental setup, and MRI results of adaptable, wearable, and stretchable MRI coils. The active and passive automatic tuning and matching strategies are examined with respect to mitigating signal-to-noise ratio reduction when the coil form is altered. A brief discussion of wireless MRI coils, which provide a solution to overcome the cabling issues associated with MRI coil development, is also included. The adaptable, wearable, and stretchable coils and various coil tuning techniques represent innovative radiofrequency coil solutions that pave the way for next-generation MRI hardware development.

PURPOSE:Wearable coils fabricated using conductive threads have high resistance that limits SNR. The purpose of this work is to demonstrate the utility of conductive fabric as a coil conductor that can be fabricated using a cutting plotter. METHODS:A single-channel coil was developed by feeding a conductive fabric sheet into an automatic cutting plotter. The fabric coil was loaded on a spherical phantom to evaluate SNR and B1 + homogeneity and compared with a single-channel conductive thread coil and a rigid printed circuit board (PCB) coil. A 14-channel wearable neck array was developed for structural imaging of the cervical spine and 4D flow MRI of the carotid arteries. The SNR from structural images and velocity-to-noise ratio (VNR) from flow images were compared with a 16-channel commercial coil. RESULTS:The single-channel conductive fabric coil provided 6.7% and 125.9% SNR increase when compared to the rigid PCB and conductive thread coils across 10 scan repetitions. The B1 + field homogeneity was 96.4%, 1% higher than the rigid PCB and conductive thread coils. The wearable neck array demonstrated a 51.1% average SNR increase from the cervical spine images across three volunteers and a 12.0% VNR increase from the postprocessed 4D flow data when compared with the commercial 16-channel array. CONCLUSION:The possibility of developing wearable coils using conductive fabric to enhance SNR in structural images and VNR in 4D flow images is demonstrated. The conductive fabric technique enables fabrication on commercial garments resulting in form-fitting wearable coils.

Objectives: Conductors used for fabrication of coils on fabric substrates suffer from inherent losses because of the combination of the metallic core with other non-conductive materials, thus limiting the SNR achieved. The purpose of this work is to demonstrate the application of copper traces on a fabric substrate to reduce conductor losses experienced in wearable coils. Methods: A single channel coil was developed from a copper sheet using a cutting plotter. The coil was loaded onto a spherical phantom to evaluate the SNR when compared with standard rigid and flexible PCB based coils. A nine-channel wearable array was developed for structural and kinematic imaging of the shoulder at 3T. The SNR from the phantom and in vivo images was compared with a commercial flexible coil. Results: The single channel coil provided 2.1 times the SNR of the rigid PCB coil and 1.2 times the SNR of the flexible PCB coil. Image acquisition using the shoulder array can be accelerated twice or thrice in the left-right or superior-inferior directions. The shoulder array provided a 12.1% increase in SNR than the commercial array from phantom imaging. The wearable shoulder array provided an 10.5% increase in average SNR when compared to the commercial coil across 2D and 3D in vivo images. Clinical Impact: The application of copper traces directly on fabric provides a new outlook toward the development of wearable coils by eliminating inherent conductor losses to improve the image quality for musculoskeletal MRI.

Magnetic resonance imaging plays a vital role in advancing medical diagnostic capabilities. Flexible and stretchable RF coils provide better adaptability to patient anatomy and offers the potential for superior MRI image quality. However, these coils are prone to resonance frequency shifts caused by changes in inductance when stretched, compressed, or bent, which can degrade the signal-to-noise ratio and overall image quality. To address this challenge, we present a nonmagnetic system capable of detecting and retuning the coil's resonance frequency to the Larmor frequency. The system operates inside the MRI bore at frequencies of up to 600MHz, with retuning accuracy up to +/- 0.04MHz at 3T. During phantom scans with a stretched coil, our system demonstrated up to a 30% improvement in SNR. This system represents a significant step toward integrating automatic tuning and matching circuits in future MRI applications, ensuring reliable performance across diverse coil configurations, patient anatomies, and field strengths.

This work demonstrates a simple, easy to use, stretchable coil system for small animal MRI at 7T. Phantom images obtained using the stretchable coil were analysed for coil SNR. The SNR of the stretchable coil was compared with SNR of a standard commercial coil. Ex-vivo imaging of a mouse and rat brain was also performed. Current work involves minimizing SNR reduction due to coil stretch. Future work involves developing a 4-channel stretchable coil array for in-vivo imaging of a rat brain.

Magnetic resonance imaging (MRI) is an important tool for medical diagnostics. Stretchable and flexible RF coils for MRI offer improved adaptability to patient anatomy and the potential for enhanced image quality. These coils are susceptible to resonance frequency shifts due to variations in inductance when stretched, compressed, or bent, which can severely impact the signal-to-noise ratio and overall image quality. Addressing this challenge, we present a system to actively detect the resonance frequency of the coil. The system employs a superheterodyne receiver architecture and a broadband resistive return-loss bridge that does not rely on ferromagnetic components. Our system is capable of operating within the MRI bore at Larmor frequencies of up to 600 MHz with a resonance frequency measurement accuracy of 99.92% at 3 T. This system offers a potential solution for integrating automatic tuning and matching circuits in future MRI applications, ensuring reliable performance across various coil configurations and field strengths.

Objective: We propose a modular stretchable coil design using conductive threads and commercially available embroidery machines. The coil design increases customizability of coil arrays for individual patients and each body part. Methods: Eight rectangular coils were constructed with custom-fabricated stretchable tinsel copper threads incorporated onto textile. Tune, match, and detune circuits were incorporated on the coil. A hook-and-loop mechanism was used to attach and decouple the modular coils. Phantom and in vivo scans at various anatomical flexion angles were acquired to highlight performance, and a temperature test was performed to verify safety. Results: In vivo MRI experiments demonstrate high sensitivity and coverage of each anatomy. As the coils are stretched, the sensitive volume increases at a rate of 10.93 mL/cm 2 . The SNR reduction of a single coil was greater during compression than when stretched, but this did not affect image quality for the array. The modularity of the array allows for adaptability for any anatomy with simple on-demand adjustment to the number and position of coil elements. Conclusion: The images demonstrated high sensitivity and coverage of the stretchable array for various anatomies and flexion angles. Stretching the coils increases the sensitive volume, allowing for a larger region to be effectively imaged. The resonance shift and SNR decrease during stretch and compression support further investigation of methods to reduce frequency shift in stretchable coils. Significance: The proposed array design allows for highly stretchable, flexible, modular, and conformal patient-centered coils that allow for increased imaging quality, greater comfort, and rapid production.

Tendon biomechanical properties and fibril organization are altered in patients with diabetes compared to healthy individuals, yet few biomarkers have been associated with in vivo tendon properties. We investigated the relationships between in vivo imaging-based tendon properties, serum variables, and patient characteristics across healthy controls (n = 14, age: 45 +/- 5 years, body mass index [BMI]: 24 +/- 1, hemoglobin A1c [HbA1c]: 5.3 +/- 0.1%), prediabetes (n = 14, age: 54 +/- 5 years, BMI: 29 +/- 2; HbA1c: 5.7 +/- 0.1), and type 2 diabetes (n = 13, age: 55 +/- 3 years, BMI: 33 +/- 2, HbA1c: 6.7 +/- 0.3). We used ultrasound speckle-tracking and measurements from magnetic resonance imaging (MRI) to estimate the patellar tendon in vivo tangent modulus. Analysis of plasma c-peptide, interleukin-1 beta (IL-1 beta), IL-6, IL-8, tumor necrosis factor-alpha (TNF-alpha), adiponectin, leptin, insulin-like growth factor 1 (IGF-1), and C-reactive protein (CRP) was completed. We built regression models incorporating statistically significant covariates and indicators for the clinically defined groups. We found that tendon cross-sectional area normalized to body weight (BWN CSA) and modulus were lower in patients with type 2 diabetes than in healthy controls (p < 0.05). Our regression analysis revealed that a model that included BMI, leptin, high-density lipoprotein (HDL), low-density lipoprotein (LDL), age, and group explained similar to 70% of the variability in BWN CSA (R-2 = 0.70, p < 0.001). For modulus, including the main effects LDL, groups, HbA1c, age, BMI, cholesterol, IGF-1, c-peptide, leptin, and IL-6, accounted for similar to 54% of the variability in modulus (R-2 = 0.54, p < 0.05). While BWN CSA and modulus were lower in those with diabetes, group was a poor predicter of tendon properties when considering the selected covariates. These data highlight the multifactorial nature of tendon changes with diabetes and suggest that blood variables could be reliable predictors of tendon properties.

We report on a 16-channel bilateral breast coil array for high-resolution MR imaging at 3T which can be used for both prone and supine breast MRI. This coil aims to improve signal-to-noise ratio (SNR) by positioning the coil array close to the breast. Sixteen 80-mm coil elements made on flexible printed circuit boards were placed on top of a 3D-printed plastic housing modeled to fit many cup sizes. Match, tune, and detune elements were incorporated onto the coil. Phantom and in vivo scans were performed to demonstrate the SNR profile and clinical efficacy of the coil. The in vivo images for both prone and supine positions show high sensitivity and coverage of the breast. The SNR profile evaluated from the phantom images was relatively uniform throughout the imaging volume. The increased sensitivity of the coils allows for improved accuracy of breast cancer diagnosis.

Radiofrequency coils are utilized during transmit and receive of MRI signals. Cable traps remove common-mode current from the coaxial cable shield, which helps improve the image quality and reduces risks of burns to the patient. Traditional cable traps use wounded coaxial cables that limit the flexibility in the design process. Floating cable traps were introduced which eliminated any physical connection between the trap and coaxial cable, allowing complete flexibility in design and placement. However, the design process of floating cable traps is iterative and may take several rounds of 3D modeling. This work seeks to optimize the design process through the use of parametric design methodologies. The proposed methodology allows for 3D printing the floating cable trap after inputting the design parameters. The cable trap was able to attenuate currents in the coaxial shields to −48 dB, highlighting its performance and design robustness.

Receive coils used in small animal MRI are rigid, inflexible surface loops that do not conform to the anat-omy being imaged. The recent trend toward design of stretchable coils that are tailored to fit any anatom-ical curvature has been focused on human imaging. This work demonstrates the application of stretchable coils for small animal imaging at 7T. A stretchable coil measuring 3.5 x 3.5 cm was developed for acquisition of rat brain and spine images. The SNR maps of the stretchable coil were compared with those of a traditional flexible PCB coil and a commercial surface coil. Stretch and conformance testing of the coil was performed. Ex vivo images of rat brain and spine from the stretchable a coil was acquired using T1 FLASH and T2 Turbo RARE sequences. The axial phantom SNR maps showed that the stretchable coil provided 48.5% and 42.8% higher SNR than the commercial coil for T1-w and T2-w images within the defined ROI. A 33% increase in average penetration depth was observed within the ROI using the stretch-able coil when compared to the commercial coil. The ex-vivo rat brain and spine images showed distin-guishable anatomical details. Stretching the coil reduced the resonant frequency with reduction in SNR, while the conformance to varying sample volumes increased the resonant frequency with decreased SNR. This study also features an open-source plug-and-play system with preamplifiers that can be used to interface surface coils with the 7T Bruker scanner.(C) 2023 Elsevier Inc. All rights reserved.

Implantable, bioresorbable drug delivery systems offer an alternative to current drug administration techniques; allowing for patient-tailored drug dosage, while also increasing patient compliance. Mechanistic mathematical modeling allows for the acceleration of the design of the release systems, and for prediction of physical anomalies that are not intuitive and may otherwise elude discovery. This study investigates short-term drug release as a function of water-mediated polymer phase inversion into a solid depot within hours to days, as well as long-term hydrolysis-mediated degradation and erosion of the implant over the next few weeks. Finite difference methods are used to model spatial and temporal changes in polymer phase inversion, solidification, and hydrolysis. Modeling reveals the impact of non-uniform drug distribution, production and transport of H+ ions, and localized polymer degradation on the diffusion of water, drug, and hydrolyzed polymer byproducts. Compared to experimental data, the computational model accurately predicts the drug release during the solidification of implants over days and drug release profiles over weeks from microspheres and implants. This work offers new insight into the impact of various parameters on drug release profiles, and is a new tool to accelerate the design process for release systems to meet a patient specific clinical need.

Objective: We propose a 16-channel bilateral breast coil array prototype for high-resolution MR imaging at $\mathbf{3T}$ . This coil aims to improve signal-to-noise ratio (SNR) by positioning the coil array close to the breast. Methods: Sixteen 8O-mm coil elements made on flexible printed circuit boards were placed on top of a $\mathbf{3D}$ -printed plastic housing modeled to fit most cup sizes. Match, tune, and detune elements were incorporated onto the coil. Phantom and in-vivo scans were performed to demonstrate the SNR profile and clinical efficacy of the coil. Conclusion: The in-vivo images show high sensitivity and coverage of the breast, and the SNR profile evaluated from the phantom images was relatively uniform throughout the imaging volume.

Reports estimate between 1.6-3.8 million sports-related concussions occur annually, with 30% occurring in youth male American football athletes. Many studies report neurophysiological changes in these athletes, but the exact reasons for these changes remain elusive. Investigation of injury mechanics highlights a need to address how player position might impact these changes. Here, 55 high school American football athletes (20 linemen; 35 non-linemen) underwent magnetic resonance spectroscopy four times over the course of a football season—once prior to the season (Pre), twice during (In1, In2), and once following (Post) to quantify metabolites (N-acetyl aspartate, choline, creatine, myo-inositol, and glutamate/glutamine) in the dorsolateral prefrontal cortex (DLPFC) and primary motor cortex (M1). Head acceleration events (HAEs) were monitored at each practice and game. Spectroscopic and HAE data were analyzed by imaging session and player position. Linear regression analyses were conducted between metabolite levels and HAEs, and metabolite levels in football athletes were compared with age-and gender-matched non-contact athletes. Across-season (i.e., between Pre and In1, In2, Post), different DLPFC and M1 metabolites decreased (p < 0.05) according to player position (i.e., linemen vs. non-linemen). The majority of regression results involved DLPFC metabolites in linemen, where metabolite levels were higher from Pre to Post, with increasing HAE load. Comparisons with control athletes revealed higher metabolite levels in football athletes both before and after the season. This study highlights the importance of player position when conducting analyses on American football athletes and demonstrates elevated DLPFC and M1 brain metabolites in football athletes compared with control athletes at both Pre and Post, suggesting potential HAE-related neurocompensatory mechanisms.

Great advances have been made towards patient-centric, lightweight, and flexible coils that provide greater conformability across patient sizes and anatomies. Innovations in flexible designs have been made possible through miniaturization of electronics and ultra-flexible conductors. To accommodate a variety of anatomical structures, including joints at various degrees of flexion, stretchable coils have been demonstrated. For the coil conductor, these prototypes have employed liquid metal [1] , elastomers [2] , and coated threads [3] , [4] . Our group has prototyped stretchable and flexible RF receive arrays using stitched, conductive thread. These designs provide two main advantages over other stretchable designs. Firstly, the conductive thread may be used in an automated, professional embroidery machine, facilitating rapid and consistent manufacturing of the loops. The thread is cut-resistant, solderable, and has a 12× greater break strength than 30 AWG copper. Evaluation of single loops demonstrated greater SNR over flexible PCB loops spaced 4.2 cm above the phantom, simulating a volume coil, while exhibiting ~14% reduction in SNR with similar, unstretched placement; however, the stitched design overcomes conductor-associated challenges when stretched and wrapped around curved anatomies, resulting in SNR increases compared to solely flexible counterparts. Examples of the use of this coil for wrist and breast phantom imaging can be seen in Figure 1 . This segues to the second advantage, which is the multipurpose application facilitated by the thread durability. Unlike some liquid metal applications, the distribution of the fibers and conductive material remains consistent, meaning that the SNR of the coil is almost unchanged before and after being stretched. When fully stretched, i.e., approximately a 20% increase in loop base loop diameter of 71 mm, SNR measurements showed ~30% decrease; however, this fully stretched condition would not be expected in clinical applications, as stretching of all elements to this taut degree should not be necessary if the coil array were of adequate size. For joints at varying degrees of flexion, it would be expected that elements curved around the anatomy would be stretched, but peripheral loops would be relatively undistorted. The breathable, stretchable fabrics produced no proton signal during phantom or in vivo scans.

Purpose With increased interest in parallel transmission in ultrahigh-field MRI, methods are needed to correctly calculate the S-parameters and complex field maps of the parallel transmission coil. We present S-parameters paired with spatial field optimization to fully simulate a double-row 16-element transceiver array for brain MRI at 7 T. Methods We implemented a closed-form equation of the coil S-parameters and overall spatial B1+ field. We minimized a cost function, consisting of coil S-parameters and the B1+ homogeneity in brain tissue, by optimizing transceiver components, including matching, decoupling circuits, and lumped capacitors. With this, we are able to compare the in silico results determined with and without B1+ homogeneity weighting. Using the known voltage range from the host console, we reconstructed the B1+ maps of the array and performed RF shimming with four realistic head models. Results As performed with B1+ homogeneity weighting, the optimized coil circuit components were highly consistent over the four heads, producing well-tuned, matched, and decoupled coils. The mean peak forward powers and B1+ statistics for the head models are consistent with in vivo human results (N = 8). There are systematic differences in the transceiver components as optimized with or without B1+ homogeneity weighting, resulting in an improvement of 28.4 +/- 7.5% in B1+ homogeneity with a small 1.9 +/- 1.5% decline in power efficiency. Conclusion This co-simulation methodology accurately simulates the transceiver, predicting consistent S-parameters, component values, and B1+ field. The RF shimming of the calculated field maps match the in vivo performance.

The discovery that the stiffness of the tumor microenvironment (TME) changes during cancer progression motivated the development of cell culture involving extracellular mechanostimuli, with the intent of identifying mechanotransduction mechanisms that influence cell phenotypes. Collagen I is a main extracellular matrix (ECM) component used to study mechanotransduction in three-dimensional (3D) cell culture. There are also models with interstitial fluid stress that have been mostly focusing on the migration of invasive cells. We argue that a major step for the culture of tumors is to integrate increased ECM stiffness and fluid movement characteristic of the TME. Mechanotransduction is based on the principles of tensegrity and dynamic reciprocity, which requires measuring not only biochemical changes, but also physical changes in cytoplasmic and nuclear compartments. Most techniques available for cellular rheology were developed for a 2D, flat cell culture world, hence hampering studies requiring proper cellular architecture that, itself, depends on 3D tissue organization. New and adapted measuring techniques for 3D cell culture will be worthwhile to study the apparent increase in physical plasticity of cancer cells with disease progression. Finally, evidence of the physical heterogeneity of the TME, in terms of ECM composition and stiffness and of fluid flow, calls for the investigation of its impact on the cellular heterogeneity proposed to control tumor phenotypes. Reproducing, measuring and controlling TME heterogeneity should stimulate collaborative efforts between biologists and engineers. Studying cancers in well-tuned 3D cell culture platforms is paramount to bring mechanomedicine into the realm of oncology.

Human brains develop across the life span and largely vary in morphology. Adolescent collision-sport athletes undergo repetitive head impacts over years of practices and competitions, and therefore may exhibit a neuroanatomical trajectory different from healthy adolescents in general. However, an unbiased brain atlas targeting these individuals does not exist. Although standardized brain atlases facilitate spatial normalization and voxel-wise analysis at the group level, when the underlying neuroanatomy does not represent the study population, greater biases and errors can be introduced during spatial normalization, confounding subsequent voxel-wise analysis and statistical findings. In this work, targeting early-to-middle adolescent (EMA, ages 13–19) collision-sport athletes, we developed population-specific brain atlases that include templates (T1-weighted and diffusion tensor magnetic resonance imaging) and semantic labels (cortical and white matter parcellations). Compared to standardized adult or age-appropriate templates, our templates better characterized the neuroanatomy of the EMA collision-sport athletes, reduced biases introduced during spatial normalization, and exhibited higher sensitivity in diffusion tensor imaging analysis. In summary, these results suggest the population-specific brain atlases are more appropriate towards reproducible and meaningful statistical results, which better clarify mechanisms of traumatic brain injury and monitor brain health for EMA collision-sport athletes.

The VOP and k-means compression algorithms in the Note have an error that will lead to some SAR underestimation in pTx RF pulse design. The spectral decomposition method1, 2 was applied to determine the representative SAR matrix for each VOP or k-means cluster. This method should loop every SAR matrix in the cluster and update the representative SAR matrix, but it failed to save the previously computed results and might lead to SAR underestimation in some cases. To ensure these SAR compression models have absolutely no SAR underestimation, we fixed this error and posted the latest code changes to GitHub (https://github.com/VincentMao/MR_VOP_kMeans, Tag: v2.0). The function "FindPSD" (ie, the spectral decomposition method; lines 64-91 in "Clustering_VOP_10g.m", lines 60-87 in "Clustering_VOP_brain.m", and lines 1-23 in "FindPSD.m") was changed to save the previously computed results to avoid causing the SAR underestimation. This implementation error does not affect the overall algorithm design for both VOP and k-means SAR compression, but affects the overestimation ratio and compression ratio reported in the Note. The conclusion in the Note regarding improved performance of k-means over VOP method remains the same, principally for SAR matrices with higher norm and eigenvalue. However, the performance of the two methods is very close in regions with lower SAR. The authors regret this mistake and apologize for any inconvenience this may have caused. There are two corrections in this erratum. First, to avoid causing any confusion, we replaced the term "Lee's method" with "VOP compression method" because our implementation modifies the algorithm reported by Lee et al.3 The VOP compression method refers to the method we defined in section 2.2.1. Second, we updated the figures and tables affected by the latest code changes. In the majority of reported cases (particularly in SAR matrices with higher norm and eigenvalues), the k-means clustering method can generate a narrower overestimation bound than the VOP compression method. However, in other cases the k-means clustering method has a similar performance to the VOP compression method. We modified the SAR compression algorithm in Lee et al3 so it adopts the overestimation bound defined in this Note. The overestimation factor in this Note allows for direct comparison between VOP and k-means methods, but it is different than the overestimation factor defined in Lee et al.3 To avoid causing any confusion about the model performance across different methods and metrics, we include one example illustrating VOP (Lee) vs. VOP (Mao) vs. k-means (Mao) in the Appendix of this erratum, Figure A1. Prior to the code correction, the VOP compression model generated 150 representative SAR matrices with 5.0% overestimation, 59 representative SAR matrices with 10.0% overestimation; the k-means method generated 150 representative SAR matrices with 4.46% overestimation, 59 representative SAR matrices with 7.91% overestimation. However, the designed RF pulses could not guarantee that all representative SAR matrices satisfy the peak local SAR constraints in some cases, particularly in the relatively low-SAR environment. After the correction, the VOP compression model generated 360 representative SAR matrices with 5.0% overestimation rate, 73 representative SAR matrices with 10.0% overestimation rate. The k-means compression model generated 360 representative SAR matrices with 4.65% overestimation, 73 representative SAR matrices with 8.90% overestimation in the whole head. Methods section, "2.3 Simulation", P3 (remove the words "(ie, Lee's method)"): In each 2D/3D volume, VOP compression was implemented twice to compare with the k-means compression using either the same overestimation rate or the same number of clusters. Results section, "3.1 SAR Compression", P1 (change the words "Lee's method" to "VOP compression method"): The k-means clustering method either generated a smaller overestimation rate or a higher compression ratio compared to the VOP compression method. Results section, "3.1 SAR Compression", P2 (changes in numbers, change the words "Lee's method" to "VOP method"): Figure 2A,B illustrate that if under the same overestimation rate (ie, 4.49%), the VOP method generated 91 clusters while the k-means method only generated 81 clusters. If under the same compression ratio (ie, compressing ~1.1 × 104 voxels to 37 clusters), then the VOP method yielded 10.0% overestimation while the k-means method yielded 8.54%. Unlike the greedy clustering strategy in the VOP compression method, the k-means clustering method tends to cluster the voxels with similar SAR behaviors. Results section, "3.1 SAR Compression", P3 (change the words "Lee's method" to "VOP compression method"): Further, Supporting Information Figure S2A compares the number of clusters and overestimation rates in the 79th axial slice using the VOP compression method and the k-means method. The two lines illustrate that the k-means method always generated fewer clusters than the VOP compression method when they have the same overestimation bound. There is a difference in the definition of overestimation bound between Lee et al3 and Mao et al4; accordingly, a direct VOP (Lee) vs. VOP (Mao) vs. k-means (Mao) comparison using the same metric is not straightforward. We applied Lee's SAR compression algorithm3, 5 in select cases, and a representative example is shown in Figure A1. Lee's method and our compression models roughly exhibited the same performance. To illustrate an example in a higher-SAR environment, Figure A2 shows the SAR/excitation accuracy (ie, NRMSE) trade-offs in another slice. The SAR matrices in this slice had higher norms and eigenvalues. This figure shows that although the global SAR stayed almost the same, the peak local SAR was suppressed more if using the k-means model. The code and data that support the findings of this study are openly available in Github at https://github.com/VincentMao/MR_VOP_kMeans. FIGURE S1 The surface of the 3D numerical Duke model (from Virtual Family) placed in a 16-channel transceiver coil array. The eight coils in the lower ring of the loop array were used for the simulation. An example x–y plane of the fields of 8 transmit channels, and Ex, Ey, Ez fields in transmit channel 1 are presented. The applied transmit sensitivity maps (i.e., field maps) were normalized into [0,1] scale. An ellipse ROI mask was applied to choose the interested region (xradius = 10 cm, yradius = 9.5 cm) FIGURE S2 A, Plot of the maximum overestimation rates (R) across all clusters along with the total number of clusters in the 79th slice of the Duke model, using two different SAR compression models. B, 10 000 random unit excitation pulses for local SAR prediction between the actual SAR model and the VOP compression model (R = 10.0%) in the 79th slice of the Duke model. C, 10 000 random unit excitation pulses for local SAR prediction between the actual SAR model and the k-means compression model (R = 8.4%) in the 79th slice of the Duke model FIGURE S3 Box-whisker plot of the overestimation rates (R) across all VOPs in each of the VOP assignments in the 79th axial slice of the Duke head. The red solid line refers to the overestimation bound applied in each VOP assignment FIGURE S4 Box-whisker plot of the overestimation rates (R) across all clusters in each of the k-means clustering results in the 79th axial slice of the Duke head. The blue dash line refers to the overestimation bound found in each k-means clustering result FIGURE S5 Analysis of 30 000 random unit excitation pulses for local SAR prediction in the actual SAR model and the compressed SAR model in the 90th axial slice of the Duke model. A, The compressed SAR model (R = 10.0%, 37 VOPs) created by the VOP compression model. B, The compressed SAR model (R = 8.54%, 37 k-means clusters) created by the k-means compression model FIGURE S6 (A) The Duke head was pre-loaded in the FDTD mesh, the red dashed ROI is a smoothed ellipse (xradius = 10 cm, yradius = 9.5 cm) in each slice. (B) The desired excitation pattern in this simulation was a smoothed elliptic cylinder. (C) The mesh plot of the desired excitation, with peak scaled to π/2 FIGURE S7 10-g SAR distribution maps in the 79th axial slice of the Duke head, across three different pTx approaches. A, The pTx approach with the VOP compression model (R = 5.0%, 360 clusters). B, The pTx approach with the k-means compression model (R = 4.65%, 360 clusters). C, The conventional pTx approach with only the RF power regularization term. The lambda parameter associated with the RF power was adjusted in the pTx approach with power control but constant in the pTx approach with SAR compression models (ie, λ = 5e-4) FIGURE S8 10-g SAR distribution maps in the 90th axial slice of the Duke head, across three different pTx approaches. A, The pTx approach with the VOP compression model (R = 5.0%, 360 clusters). B, The pTx approach with the k-means compression model (R = 4.65%, 360 clusters). C, The conventional pTx approach with only the RF power regularization term. The lambda parameter associated with the RF power was adjusted in the pTx approach with power control but constant in the pTx approach with SAR compression models (ie, λ = 5e-4) FIGURE S9 Simulated excitation pattern profiles at y = 0 cm in the 79th axial slice of the Duke head, created by the desired excitation pattern (black solid line) and the designed RF pulses (blue dash line) using different pTx approaches. A, The pTx approach with the VOP compression model (R = 5.0%, 360 clusters). B, The pTx approach with the k-means compression model (R = 4.65%, 360 clusters). C, The conventional pTx approach with only the RF power regularization term. The lambda parameter associated with the RF power was only constant in the pTx approach with SAR compression models (ie, λ = 5e-4) Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.