Multi‐Vendor and Multi‐Site Characterization of ¹³C Coils for Clinical Studies

ABSTRACT

Purpose

Due to the inherently low signal‐to‐noise ratio (SNR) domain for ¹³C imaging, optimal coil performance is essential to maximize data quality. Quality control (QC) for ¹³C coils employed in magnetic resonance spectroscopy (MRS) is of paramount importance. This study presents a gradient‐recalled echo (GRE) sequence alongside an image‐based analysis for SNR estimation and QC of ¹³C coils.

Methods

The proposed GRE sequence (slice selective RF pulse, slice thickness 100 mm, 16 averages, 64 × 64 matrix size, 400 × 400 mm FOV, 60 Hz/px BW, 50° FA, 80 ms TR, 14 ms TE, cartesian acquisition trajectory) was comprehensively evaluated on four clinical 3 T systems, across three sites and two vendors. Thirteen coils were evaluated, including multiple surface coils (transmit/receive and receive only) and volume coils for different body parts. The GRE sequence was compared with a previously described chemical shift imaging (CSI) technique for QC of ¹³C coils. The GRE sequence is widely available and built into all MR systems. In contrast, the CSI sequence is more specialized and is not available for all imaging systems.

Results

A Passing‐Bablok regression and a Spearman correlation of 0.78 demonstrated good agreement of the two methods. The GRE approach offers reduced acquisition time and simplified analysis with SNR trends demonstrating reproducible coil performance across all sites and vendors.

Conclusion

This standardization aims to strengthen multi‐center trials by ensuring data comparability and reliability, advancing the clinical translation of ¹³C MRS methodologies. The results and analysis pipeline is made publicly available to facilitate implementation and coil comparisons across institutions.

1. Introduction

Magnetic Resonance Imaging (MRI) has become an essential technology in modern hospitals. Magnetic Resonance Spectroscopy (MRS) using ¹H has been extensively used in research, and some extent clinical routine [1, 2]. ¹H MRS has seen extensive use in clinical research practice with proton density fat fraction (PDFF) representing an established example of a standardized clinical implemented ¹H biomarker [3]. Recently, the use of other nonzero‐spin isotopes (²H, ¹³C, ¹²⁹Xe) is being explored in clinical research, with ¹²⁹Xe already in clinical practice [4, 5]. The main challenge of ¹³C spectroscopic imaging is sensitivity [6, 7], for example, the natural abundance of ¹³C is only 1.1% and the gyromagnetic ratio is ¼ of that of ¹H. This can be overcome by hyperpolarization methods, which enable real‐time metabolic imaging with much greater sensitivity [6, 7]. The predominant technique for achieving hyperpolarization is dissolution dynamic nuclear polarization (d‐DNP), which has demonstrated an increase of the signal‐to‐noise ratio (SNR) exceeding 10 000 [8], with parahydrogen‐induced polarization (PHIP), emerging as a preclinical alternative [9].

Hyperpolarized ¹³C MRS with [1‐¹³C]‐labeled pyruvate provides in‐depth insight into pyruvate metabolic pathways at high temporal resolution, typically a few seconds. These pathways can be altered in multiple pathologies [6, 10, 11]. In 2013, the first clinical application in a prostate cancer patient, and [1‐¹³C]‐labeled pyruvate remains the most widely used probe [12, 13]. As of 2025, 15 sites have scanned approximately 1200 subjects across multiple diseases with hyperpolarized [1‐¹³C]pyruvate to demonstrate both early diagnosis and treatment response in cancer patients, as well as to investigate cardiac metabolism [14, 15, 16].

¹³C MRS requires electronics and coils that can transmit and receive ¹³C radiofrequencies (RF). Sequences applied for hyperpolarized imaging are often heavily accelerated, due to the short T₁ (approximately 30 s at 3 T in vivo) of [1‐¹³C] pyruvate in vivo [12, 17, 18] as well as the irreversible polarization loss that occurs with each RF excitation applied during imaging. Together, these constraints place strict demands on RF pulse design and sequence optimization [19].

Non‐hyperpolarized ¹³C MRS is also an active area of research, with applications including glycogen quantification and brain metabolism studies, each placing their own demands on coil design and performance [20, 21, 22, 23].

Dual‐tuned coils, ¹H/¹³C, are therefore frequently preferred as they enable anatomical imaging in conjunction with MRS, without modifying the patient setup. However, tuning to two different resonance frequencies presents challenges, as the transmit fields might interfere, resulting in additional SNR reduction [24]. Single‐tuned coils avoid this but require coil changes between ¹H and ¹³C acquisitions, compromising co‐registration.

¹³C MRS is still an application under development, and no established consensus exists regarding quality control and performance evaluation of dedicated ¹³C MR coils. Evaluation of both B₁ ⁺ homogeneity and B₁ ⁻ sensitivity is crucial due to the inherently low SNR in ¹³C MRS. Signal averaging, while potentially improving SNR, compromises temporal resolution, which is critical in capturing dynamic metabolic processes [25]. The balance between SNR and hardware choice represents a continuous optimization problem that is case‐dependent [26].

Given that multi‐center trials constitute a vital step toward clinical application, consensus is needed to ensure reliable data comparison [27, 28], with one proposal on clinical studies published in 2025 [16]. This paper describes how future studies should be conducted and how the quality control (QC) protocol can identify similarities and differences between coils in future multi‐center studies. QC procedures constitute a critical component of this standardization framework, ensuring the reproducibility and reliability of clinical investigations utilizing hyperpolarized ¹³C MRS [29, 30]. Implementing such procedures is challenging, as even established ¹H‐based quality control is not yet routine in clinical practice. Although ¹H‐MRI has a far higher SNR than ¹³C MRS, QC also remains essential in multi‐center ¹H studies to ensure data comparability [31].

We present a novel QC protocol for clinical ¹³C coils using a fundamental and cross‐platform compatible gradient‐recalled echo (GRE) sequence. The protocol is illustrated in Figure 1. The proposed GRE acquisition protocol was compared against the Chemical Shift Imaging (CSI)‐based methodology utilized by Sanchez‐Heredia et al. [32]. We aim to compare the GRE method to the CSI method, using CSI as reference [32]. Protocol validation was conducted across multiple 3 T systems: two Siemens Biograph mMR PET/MRI scanners at Rigshospitalet, Copenhagen, Denmark (designated RH1 and RH2, respectively) and one GE Discovery MR750 system at Aarhus University Hospital (AUH, Aarhus, Denmark), and one GE Premier XT scanner at Cambridge University (CAM, Cambridge, United Kingdom).

FIGURE 1.

An ethylene glycol phantom being scanned on a single site using our new GRE protocol with the locally available ¹³C coils. A selection of coils and SNR images is shown for all four scanners and the three sites included in this study. The signal in each image was determined to be the 500 pixels with the most signal, corresponding to the inside of the phantom.

2. Methods

2.1. QC Protocol

All acquisitions were done on custom ethylene glycol phantoms $(\ge$ 99.0% purity, Merk, Darmstadt, Germany) loaded with 17 g/L NaCl [32]. Two different cylinder‐shaped phantoms were manufactured to match coil geometry: small (Ø8 cm × L20 cm) and large (Ø24 cm × L40 cm). CAM used a separate phantom of similar size. Flip angle (FA) calibration was performed for each coil configuration prior to all experiments.

The Siemens systems performed FA calibration using a series of short RF pulses, with the signal fitted to a sinusoidal to identify the 90° FA [33, 34]. For the GE systems, the FA calibration was performed using the Bloch‐Seigert shift method [35].

All images acquired with multiple coil elements were combined using regular sum‐of‐square (SoS) [36].

The methodology proposed by Sanchez‐Heredia et al. [32, 37] established a framework for coil benchmarking utilizing a CSI sequence (9.6 min acquisition time) and a 0° FA free induction decay (FID) to estimate signal and noise, respectively. Sanchez‐Heredia et al. defined SNR as:

$$\mathit{SN}{R}_{\mathit{CS}{I}_{\mathit{dB}}}=10{\log }_{10}\left(\frac{\sqrt{2}\left|s\right|}{\sqrt{{\sigma }^2}}\right)$$ (1)

where the signal (s) represents the area under the spectral signal curve, and the noise (σ) is the covariance of the noise FID. This framework requires more calculation steps, especially for the use of multi‐element coils, as this requires the sensitivity of each element to be known.

The present study proposes a simpler, streamlined method for SNR estimation based on a GRE sequence and image domain analysis, in contrast to spectral analysis utilized in the CSI framework. The noise is extracted from the GRE images itself. According to the NEMA MS3‐2008 standard, at least 1000 pixels should be extracted from the image, and the noise should be Rayleigh distributed [38]. The pixels were chosen as the five outer rows and columns in the image, 1280 pixels in total, with no visible artifacts. The Rayleigh distribution was validated through Akaike Information Criterion (AIC) by comparing a Rayleigh, Rician and Gaussian distribution. The signal from within the phantom was extracted as a user‐defined number of pixels with the highest signal. The SNR could then be calculated as follows:

$$\mathit{SN}{R}_{\mathit{GR}{E}_{\mathit{dB}}}=10{\log }_{10}\left(\frac{\bar{s}}{\sigma (b)}\right)$$ (2)

where $\bar{\mathrm{s}}$ represents the mean signal and σ(b) represents the standard deviation of the noise, estimated from the background. For ease of visualization, so‐called SNR images were also computed as the entire image divided by the noise standard deviation. Data analysis was performed using a Python script (version 3.9, Beaverton, OR, USA), available on GitHub [39].

2.2. Image Acquisition

The CSI data was acquired with a slice selective RF pulse with a single slice of 20 mm, 1 average, 24 × 24 matrix size, 360 × 360 mm field of view (FOV), 5000 Hz bandwidth (BW), 70° FA, 1000 ms repetition time (TR), minimized echo time (TE) at 2.3–2.7 ms (at RH and AUH), 1024 FID points, cartesian trajectory. All data were 2D Fourier transformed and FFT‐shifted.

The GRE data were acquired with a slice selective RF pulse with a single slice with a thickness of 100 mm, signal averages of 1, 4, 8, 16, and 32, a 64 × 64 matrix size, 400 × 400 mm (FOV), 60 Hz/px BW, 50° FA, 80 ms TR, 14 ms TE; acquisition trajectory was cartesian—resulting in an 83 s acquisition time for 16 averages. The BW and FA were chosen for SNR maximization, whereas TR was a scan time compromise. TE was minimized to minimize artifacts by the ethylene glycol triplet. SNR measurements were acquired in both axial and sagittal scan planes, and the values presented represent the mean of these two acquisitions for each coil configuration. A total of 13 different coils were analyzed, across four systems, with a total of 16 different scans, each in two imaging planes. Both sequences were implemented on all sites without any waveform modulation.

Validating the scaling of SNR versus the number of averages was conducted to verify that the GRE sequence scaled with a factor of $\surd$N, with N representing # of averages. To determine the number of averages of the GRE sequence, an efficiency analysis was conducted focusing on SNR gain per average of sequence, with efficiency defined as SNR/averages.

Following the comparative study between RH and AUH, the GRE sequence was implemented at CAM employing 16 averages, which represent the recommended implementation. Specifications for all coils tested in this study are presented in Table 1.

TABLE 1.

Overview and characteristics of the included coils.

Coil #	Coil type	Coil specs	Manufacturer	MR system	Location
1	Abdominal Birdcage	1H/13C (4 Tx/Rx)	RAPID Gmbh	Siemens 1 + 2	Copenhagen, DK
2	Breast Array	1H/13C (4 Tx/Rx)	RAPID Gmbh	Siemens 1 + 2	Copenhagen, DK
3	Head Birdcage	13C (1 Tx/Rx)	RAPID Gmbh	Siemens 1 + 2	Copenhagen, DK
4	Flex	1H/13C (1 Tx/Rx)	RAPID Gmbh	Siemens 2	Copenhagen, DK
5	Flex	1H/13C (16 Rx)	JD Coils	GE	Aarhus, DK
6	Flex	1H/13C (8 Rx)	JD Coils	GE	Aarhus, DK
7	Heart Array	1H/13C (16 Rx)	RAPID Gmbh	GE	Aarhus, DK
8	Helmholtz loop	1H/13C (1 Rx)	PulseTeq Limited	GE	Aarhus, DK
9	Patch Coil	1H/13C (8 Rx)	RAPID Gmbh	GE	Aarhus, DK
10	Loop/Ring	1H/13C (1 Rx)	RAPID Gmbh	GE	Aarhus, DK
11	Flex	23Na/13C (1 Rx)	RAPID Gmbh	GE	Aarhus, DK
12	Breast Array	13C (2 Tx/Rx, 6 Rx)	RAPID Gmbh	GE	Cambridge, UK
13	Head Birdcage	1H/13C (1 Tx/Rx)	RAPID Gmbh	GE	Cambridge, UK

Note: A description of type, tuning and location of the specific system is stated. Siemens mMR system in Copenhagen, GE Discovery MR750 system in Aarhus, GE Premier XT scanner in Cambridge.

For multi‐element coils, data were reconstructed using SoS and SENSitivity (SENSE) encoding, implemented with the Berkeley Advanced Reconstruction Toolbox (BART) [40]. GRE SNR was calculated for both reconstruction methods. Wilcoxon analysis with H₀: SENSE ≠ SOS was performed to statistically compare the two methods.

2.3. Comparison of CSI and GRE

To compare the two SNR methods, a Spearman correlation was performed as well as a Passing‐Bablok regression to validate that the two methods are interchangeable. Using the equation from the Passing‐Bablok fit, the CSI data was normalized to GRE values in order to have SNR data in the same absolute range. A Bland–Altman analysis was performed to further assess agreement between the methods. The analysis was performed using Python.

3. Results

3.1. GRE Signal Scaling

SNR of the coils as a function of averages at RH and AUH can be seen in Figure S1. Across all 14 measurements, SNR as a function of average increase was well described by the fitted model (mean R² = 0.80), with an overall SNR overestimation of 4.5%. The GRE SNR, across the two sites, ranges from 10.6 dB at one average (#1 Abdomen) to 36.2 dB at 32 averages (#3 Head). At AUH, the GRE SNR ranges from 11.4 dB at 1 average (#12 ²³Na/¹³C Flex) to 35.3 dB at 32 averages (#5 Flex), all listed in Table 1.

As shown in Figure S1, the average increase in SNR across all coils is 13.6 dB when increasing the number of averages from 1 to 16, whereas the increase from 16 to 32 averages results in a mean gain of 2.4 dB. In Figure S2, an efficiency plot is shown. Here it is visible that the efficiency in SNR gain at 32 averages (mean efficiency = 0.99) is like 16 averages (mean efficiency = 1.82). Based on this, the GRE sequence with 16 averages was chosen for further QC implementation. Results from CAM showed comparable findings, both compared to AUH (same scanner vendor) as well as when compared to data from RH (similar coils).

3.2. Comparison of GRE and CSI

SNR, as acquired with GRE (using 16 averages) and CSI are presented in Figure 2. The dual‐tuned coils, including #1, #2, and #4 (for RH) showed lower SNR at RH than the single‐tuned #3. #4 at RH showed a higher SNR than the other dual‐tuned coils. The maximum inter‐site difference in SNR utilizing the CSI sequence was 24.5 dB (a 16.8‐fold difference in non‐dB values) between RH and AUH. The difference observed with the GRE sequence was not as pronounced as that observed with the CSI sequence, with a maximum difference of 12.9 dB (a 4.4‐fold difference in non‐dB values), demonstrating that both the absolute and relative differences between RH and AUH are reduced employing the GRE sequence. Coil #1, #2, and #3 all scanned at RH1 and RH2 showed a 1.1‐fold difference for the GRE method and an 8.6‐fold difference for the CSI method. The standard error between the two scan planes for the GRE, Figure 2A, and CSI, Figure 2B, revealed less deviation between the two scan planes when using the CSI sequence than the GRE, which was apparent across all sites.

FIGURE 2.

Comparison of signal‐to‐noise ratio (SNR) of the GRE sequence, using 16 averages, and a previously presented CSI sequence. The error bars represent the standard error of the axial and sagittal scan planes. Conducted using Prism 10 (GraphPad, Boston, MA, USA).

Looking only at the GRE sequences in Figure 2A, the coil with the overall highest SNR is the single‐tuned breast coil (#13 in Table 1, CAM) with an SNR of 33.4 dB. The dual‐tuned coil with the highest SNR was also from CAM, being the head coil with an SNR of 32.2 dB (#14 in Table 1, CAM).

The GRE sequence shows a lower variance in SNR across sites, representing the variability expected to be found between MR scanners and multiple MR coils. The CSI sequence was only conducted on the sites where the GRE sequence was evaluated before implementation at CAM.

The SoS vs. SENSE comparison resulted in a p‐value of 0.14, thereby rejecting H₀ and no statistical difference between the two reconstruction methods was found.

The Passing‐Bablok regression is presented in Figure 3, with a slope of 1.22 and a Y‐intercept at 76.54 along with 95% confidence intervals. To further validate the correlation of the SNR measurements, a Spearman correlation was performed showing a strong correlation of 0.78 [41]. The Bland–Altman analysis of the normalized CSI data and GRE data, shown in Figure S3, resulted in all measurements being within 95% limits of agreement.

FIGURE 3.

Passing‐Bablok regression, with a slope of 1.22 and a Y‐intercept at 76.54, of the two SNR methods. A Spearman correlation of 0.78 further validates a strong correlation between the two methods.

In Figure S4, all the obtained GRE images are shown. Please note that only a single well is visible in the breast coil from RH, due to the other being broken. A comparison of CSI vs. GRE images of a few coils is shown in Figure S5.

4. Discussion

4.1. GRE Implementation and Practical Advantages

We present a streamlined workflow for SNR estimation of clinical ¹³C coils using a GRE sequence, validated across three sites (RH, AUH, and CAM), four 3 T MRI systems, and two vendors (Siemens and GE). The GRE sequence was acquired in 83 s with 16 averages, compared with 576 s for the CSI sequence, representing close to a 7‐fold reduction in acquisition time. This reduction is primarily attributable to the absence of a spectral dimension in the GRE sequence, which eliminates the additional phase‐encoding step required by CSI [42].

A key feature of the GRE‐based approach is its accessibility. GRE sequences are universally implemented across all MRI platforms and are routinely utilized in clinical imaging protocols. In contrast, while the CSI sequence is a standard component of X‐nuclei MR systems, the associated QC workflow is more complex, requiring spectral analysis, noise FID acquisition at 0° flip angle, and for multi‐element coils, knowledge of individual element sensitivities, as that method is based on SENSE coil combination for multi‐element coils, opposing the SoS reconstruction utilized throughout this work. The GRE workflow operates entirely in the image domain on coil‐combined DICOM data, lowering the implementation barrier for sites aiming to perform ¹³C coil QC in a clinical setting. On Siemens systems with the multi‐nuclear option (MNO), the GRE can be activated for ¹³C using the IDEACMDTOOL and associated password, ensuring compatibility with MNO‐specific workflows.

The efficiency analysis further supported the choice of 16 averages as the recommended acquisition protocol. Increasing from 1 to 16 averages yielded a mean SNR gain of 13.6 dB, whereas doubling again to 32 averages provided only an additional 2.4 dB, at the cost of doubling the acquisition time. The mean efficiency at 16 averages (1.82) substantially exceeded that at 32 averages (0.99), confirming that 16 averages offer the optimal trade‐off between SNR and scan time.

4.2. Statistical Comparison of GRE and CSI

SoS reconstruction was selected over SENSE due to its compatibility with standard DICOM magnitude data, which lacks the complex information required for SENSE. SoS also ensures uniform reconstruction across coils regardless of coil elements. Although no statistically significant difference between the two methods was found, SENSE reconstruction yielded a lower SNR (mean 1.9 dB) than SoS, while preserving the SNR trends across coils, sites and systems.

The Passing‐Bablok regression yielded a slope of 1.22 and a Y‐intercept of 72.54, and the Bland–Altman analysis confirmed that all measurements fell within the 95% limits of agreement, supporting the agreement between the two methods. A Spearman correlation of 0.77 further demonstrated a strong association between GRE‐ and CSI‐derived SNR values [41].

The GRE sequence showed smaller inter‐site differences in both absolute and relative SNR compared with the CSI sequence. The maximum inter‐site difference using CSI was 24.5 dB (16.8‐fold in linear scale), whereas the GRE sequence yielded a maximum difference of 12.9 dB (4.45‐fold in linear scale). Comparing Coil #1, #2 and #3 that were scanned at two equivalent systems at RH, the same pattern is observed with a 1.1‐fold difference of the GRE compared to an 8.6‐fold difference for the CSI method. This enhanced consistency suggests that the GRE‐based approach provides more robust SNR estimation across different hardware configurations and vendors, which is a desirable property for a QC method intended for multi‐center deployment.

The GRE method additionally offers higher spatial resolution than CSI, enabling easier detection of spatial inhomogeneities in coil sensitivity profiles. The generally lower per‐acquisition SNR of the GRE sequence compared with CSI is effectively compensated by signal averaging, as confirmed by the scaling validation showing a mean R² of 0.80 for the expected √N relationship between SNR and number of averages.

The standard error of the mean between the two scan planes was larger for the GRE sequence than for the CSI, which may reflect the inherently lower SNR per excitation of gradient echo acquisitions. However, the use of 16 signal averages mitigates this by reducing the standard error, and the dual‐plane acquisition itself provides a measure of reproducibility for each coil configuration.

4.3. Coil Performance Considerations

The SNR calculation methodology employed in this study defines signal from the highest‐intensity pixels within the phantom rather than a mean ROI of the entire phantom, thereby characterizing maximum SNR rather than the mean across the entire imaging volume. This approach is particularly informative for surface coils, where proximity to the scanned object increases sensitivity, and for evaluating coils intended for applications where maximum local SNR is the primary objective [29, 30].

The distinction between single‐tuned and dual‐tuned coils represents an important consideration for ¹³C imaging [43]. Most coils in this study were dual‐tuned, which, as expected, generally yielded lower SNR than the single‐tuned alternatives (#3 and #13 in Table 1). However, dual‐tuned coils provide the critical capability of anatomical ¹H reference imaging without patient repositioning, making them preferable for clinical applications where co‐registration of metabolic and anatomical data is essential. The dual‐tuned flex coil at RH notably approached the SNR of the single‐tuned head coil, which may be attributed to its ability to be positioned in closer proximity to the phantom, illustrating that coil placement geometry can partially offset the SNR penalty of dual tuning.

Inter‐site variability was observed across all sites and is consistent with previous multi‐center coil characterization studies [32, 44]. The lower SNR observed at RH1 compared with RH2 is attributed to the older scanner system, with differences in hardware configuration and Faraday cage characteristics. AUH showed greater internal variability across coil configurations despite using the same transmitter, while RH2 and CAM results were more consistent with each other. Such variability underscores the inherent complexity of comparing SNR across different MR systems and reinforces the value of systematic QC to identify and characterize these differences.

A halo artifact was observed in GRE images, arising from the ethylene glycol triplet spectrum of the phantom. This artifact is systematic and does not affect the SNR measurements, as the signal was defined from pixels within the phantom interior. Though the BW chosen in this work benefits the SNR, it is likely a contributor to this artifact, as the triplet's ˜142 Hz splitting is not fully collapsed at this BW [32]. A higher BW could be implemented if the artifact disturbs the readout, at a cost of SNR. TE of 14 ms was selected on the basis of empirical optimization to maximize SNR, and ethylene glycol was chosen for its availability and low cost [15, 45].

The relatively high FA and short TR used for SNR maximization introduce greater spatial variation in the reported SNR for the surface coils. The transmit field, and thereby the FA, is spatially varying and decreases as the distance to the surface coil increases. As GRE signal depends nonlinearly on the FA, this can cause the reported SNR to be a combination of sensitivity and local longitudinal magnetization saturation [46]. Low FA acquisitions, which are less susceptible to these artifacts, are generally employed for sensitivity mapping in parallel imaging, but were not used in this work, as maximum SNR was prioritized [47].

4.4. Implications for Multi‐Center ¹³C MRS Trials

The enhanced cross‐site and cross‐vendor consistency of the GRE sequence, compared with CSI, demonstrates the suitability of this method for standardizing QC in future multi‐center trials. Such standardization is increasingly important as clinical ¹³C MRS expands, with approximately 1200 subjects scanned across 15 sites as of 2025. Recent consensus recommendations have emphasized the need for harmonized protocols to ensure data comparability across institutions [16, 27].

Hyperpolarized ¹³C experiments are resource‐intensive, both in terms of cost and preparation time, making rigorous QC essential to ensure that acquired data meets quality thresholds. In multi‐center and multi‐vendor studies, harmonization of coil performance assessment across sites reduces inter‐scanner variability, enhances data comparability, and ultimately increases the statistical power of pooled analyses. By implementing GRE‐based SNR analysis prior to hyperpolarized studies, coil performance can be quantitatively benchmarked, enabling sites to identify and address potential issues before patient scanning.

As ¹³C MRS is not body‐noise dominated, unlike ¹H imaging, downstream electronics of the coils can contribute to the SNR. Coil vendors typically specify noise figure in the documentation, and this should be considered when comparing across sites [48].

When multi‐center trails are conducted, it should be noted that even though coils are showing similar characteristics across sites, it does not guarantee similar results when doing hyperpolarized ¹³C MRS. This would further require similar polarization and MR acquisition protocols among others.

The analysis pipeline and results have been made publicly available [39] to facilitate adoption and to enable cross‐institutional coil comparisons. We encourage additional sites to contribute to this growing database of QC assessments, thereby strengthening multi‐center standardization efforts and advancing the clinical translation of ¹³C MRS methodologies.

5. Conclusion

We have developed and validated a novel streamlined, GRE‐based ¹³C coil characterization method, successfully implemented across three different sites, utilizing multiple RF coil configurations, scanners, and vendors. The proposed method has demonstrated reliable SNR assessment, establishing it as an accessible, efficient, and robust methodology freely available to the community. The GRE‐based QC method was compared with a previously presented CSI method. A Spearman regression with 0.78 and a Passing‐Bablok regression confirmed the interchangeability of these two methods. The relative difference within each method was the smallest of the GRE, with a difference of 4.4‐fold compared with 16.8‐fold.

We hope that additional sites will contribute to the growing database of QC assessments, thereby strengthening multi‐center standardization efforts.

Funding

This project received funding from the European Union's Horizon 2020 research and innovation programme under grant agreements no. 670261 (ERC Advanced Grant) and 668532 (Click‐It), the Lundbeck Foundation, the Novo Nordisk Foundation, the Innovation Fund Denmark, the Neuroendocrine Tumor Research Foundation, the Danish Cancer Society, Arvid Nilsson Foundation, the Neye Foundation, the Sygeforsikringen Denmark, the Research Foundation of Rigshospitalet, the Danish National Research Foundation (grant 126)—PERSIMUNE, the Research Council of the Capital Region of Denmark, the Danish Health Authority, the John and Birthe Meyer Foundation and Research Council for Independent Research. Andreas Kjaer is a Lundbeck Foundation Professor.

Conflicts of Interest

Andreas Kjaer is a Lundbeck Foundation Professor. The remaining authors declare no conflicts of interest.

Supporting information

Figure S1: Curves of the signal‐to‐noise ratio (SNR) measured in the ethylene glycol phantom using different averages using the gradient‐recalled echo (GRE) sequence. The data is from the Siemens site, RH‐1 (A) and RH‐2 (B), and the GE site, AUH (C). A scaling curve has been fitted (mean R² = 0.80) to the data, indicating that the SNR is expected to increase every time the averages are doubled. Conducted Python 3.9.

Figure S2: Efficiency curve of signal‐to‐noise ratio (SNR) vs. averages. These results show that the gain in SNR when increasing from 16 averages to 32 averages results in similar SNR results. Hence, the GRE method was widely implemented using 16 averages.

Figure S3: Bland–Altman plot of gradient‐recalled echo (GRE) and chemical shift imaging (CSI) data, normalized to GRE using the Passing‐Bablok regression. All coils are within the 95% limits of agreement. The non‐zero mean difference show that the CSI tends to show higher SNR result than the GRE method.

Figure S4: Overview of the signal‐to‐noise ratio images produced with the gradient‐recall echo method, measured in dB. The numbers of each image refer to the coils described in Table 1. Only one well on the RH Breast coil is visible due to a malfunction.

Figure S5: Chemical shift imaging signal‐to‐noise ratio (SNR) images compared to the gradient‐recalled echo SNR images, both measured in dB. The numbers refer to the specific coil presented in Table 1.

Data Availability Statement

A script implementing the GRE analysis is available at GitHub (29). Data are available upon reasonable request to the corresponding author.