Proton and phosphorus magnetic resonance spectroscopy of the healthy human breast at 7 T

In vivo water‐ and fat‐suppressed 1H magnetic resonance spectroscopy (MRS) and 31P magnetic resonance adiabatic multi‐echo spectroscopic imaging were performed at 7 T in duplicate in healthy fibroglandular breast tissue of a group of eight volunteers. The transverse relaxation times of 31P metabolites were determined, and the reproducibility of 1H and 31P MRS was investigated. The transverse relaxation times for phosphoethanolamine (PE) and phosphocholine (PC) were fitted bi‐exponentially, with an added short T 2 component of 20 ms for adenosine monophosphate, resulting in values of 199 ± 8 and 239 ± 14 ms, respectively. The transverse relaxation time for glycerophosphocholine (GPC) was also fitted bi‐exponentially, with an added short T 2 component of 20 ms for glycerophosphatidylethanolamine, which resonates at a similar frequency, resulting in a value of 177 ± 6 ms. Transverse relaxation times for inorganic phosphate, γ‐ATP and glycerophosphatidylcholine mobile phospholipid were fitted mono‐exponentially, resulting in values of 180 ± 4, 19 ± 3 and 20 ± 4 ms, respectively. Coefficients of variation for the duplicate determinations of 1H total choline (tChol) and the 31P metabolites were calculated for the group of volunteers. The reproducibility of inorganic phosphate, the sum of phosphomonoesters and the sum of phosphodiesters with 31P MRS imaging was superior to the reproducibility of 1H MRS for tChol. 1H and 31P data were combined to calculate estimates of the absolute concentrations of PC, GPC and PE in healthy fibroglandular tissue, resulting in upper limits of 0.1, 0.1 and 0.2 mmol/kg of tissue, respectively.

In vivo water-and fat-suppressed 1 H magnetic resonance spectroscopy (MRS) and 31  The application of 1 H spectroscopy in breast cancer diagnostics and treatment response has been the subject of several reviews, [6][7][8][9][10][11][12] in which extensive entries to the literature can be found. In vivo tChol 1 H MRS in the human breast is challenging because of the high water and (potentially) high fat signals, which are several orders of magnitude larger than the tChol signal. Appropriate placement of the MRS voxel, excluding as much fatty tissue as possible, is therefore important. In addition, robust water and fat suppression is required, 12 for which we have chosen a VAPOR (variable power and optimized relaxation delays) 13 scheme (water suppression) prior to the sequence and two chemical shift-selective radiofrequency (RF) pulses surrounded by crusher gradients applied within the TE of the MRS sequence (MEGA 14 scheme fat suppression).
At high magnetic field, the most accurate localization technique available so far is a single-voxel semi-LASER sequence, 15 as it uses adiabatic highbandwidth RF pulses that have small chemical shift displacement artifacts and slice profiles with sharp transitions. In the case of MR coils and systems with limited B 1 , the use of a semi-LASER with FOCI (frequency offset corrected inversion) or GOIA (gradient-modulated offset independent adiabaticity) pulses could provide a better alternative to minimize the chemical shift artifact, although the use of these pulses leads to longer TEs and higher specific absorption rate (SAR). In order to effectively incorporate single-shot frequency alignments to mitigate field fluctuations during the scan time, metabolite cycling can be applied, 16,17 in which the large signal of water can be removed by subtraction, but used for accurate frequency and phase alignments. When residual water or lipid signals remain visible, Bolan et al 18 10 However, the intrinsic sensitivity of 1 H MRS is much higher than that of 31 P MRS.
Recent advances in 31 P MRS, as reviewed by Khlebnikov et al, 23 have resulted in an increased sensitivity. Moreover, the general formula of the sensitivity ratio between 1 H and 31 P, as known from the NMR field (γ 1 H/γ 31 P) 3 , 24 is not what is practically encountered when applied in the human breast in vivo. First, in the regime of tissue load dominance of the receive coil, i.e. if noise from the sample dominates thermal noise from the coil, which is the case in human in vivo applications, the sensitivity ratio is reduced by approximately γ 1 H/γ 31 P to (γ 1 H/γ 31 P) 2 , as the thermal noise of a conducting sample is proportional to the resonance frequency. 24 Second, when in the regime in which T 2 * substantially dominates over T 2 (i.e. as a result of susceptibility differences between lipids and glandular tissue in the breast: T 2 * < < T 2 ), the sensitivity ratio between 1 H and 31 P MRS is reduced by approximately the square root of the ratio γ 1 H/γ 31 P to (γ 1 H/γ 31 P) 3/2 . This includes two counteracting effects. The sensitivity of 31 P is increased by γ 1 H/γ 31 P, because of a sharper line width (peak height) for 31 P, as a result of the lower resonance frequency of the 31 P nucleus relative to the 1 H nucleus, but decreased by the square root of γ 1 H/γ 31 P, because of the longer sampling time of this sharper peak.
(If T 2 * < < T 2 , then T 2 *( 1 H)/T 2 *( 25 P) ≈ (γ 1 H/γ 31 P) −1 . The effect of the longer sampling time for 31 P is analogous to decreasing the number of sampling averages, and varies with the reciprocal square root (γ 1 H/γ 31 P) -1/2 . 24 It should be noted that all of these sensitivity comparisons only hold under the assumption of RF tissue load dominance and susceptibility dominance.) Moreover, as 31 P signals cannot be detected from adipose lipids, voxel selection in 31 P MRS becomes substantially less critical. Third, when using multi-echo sequences, such as AMESING (adiabatic multi-echo spectroscopic imaging), the sensitivity of 31 P MRS can be increased. 26 Finally, polarization transfer (not used here) can potentially increase the sensitivity by a factor γ 1 H/γ 31 P. [27][28][29][30] In this study, we compare state-of-the-art 1   During the first scan session, a screen print was made of the voxel placement (in three directions) and, during the second scan session, this screen print was used as a map to visually place the voxel at approximately the same position for the second scan session. As water was used as an internal reference for quantification, any variation in fibroglandular tissue content in the voxel would be accounted for and would not lead to different tChol concentrations as long as the tChol concentration is homogeneous over the fibroglandular tissue. Water suppression was performed by a VAPOR scheme 13 and fat suppression by a MEGA scheme 14 with narrow-band adiabatic full passage (AFP) pulses. Metabolite cycling for tChol was performed by measuring a dynamic series of 2 × 14 scans for each TE (118, 119, 120 and 121 ms) and selectively inverting (with a narrow-band AFP pulse with a frequency sweep of 300 Hz) the signal at 3.2 ppm in the even scans, and subsequently subtracting even from odd scans after phasing the residual water signal, which was a few per cent of the unsuppressed water signal. 1 H spectroscopy was cardiac triggered with a TR of three heart beats for the choline scan (at four different TEs: 118, 119, 120 and 121 ms) and nine heart beats for the water reference scan with TE = 26 ms. The total scan time for the 1 H spectroscopy protocol at an average heart rate of 60 beats/min was 6 min and 12 s.  1 H spectra at the four different TEs were phased and frequency aligned on the residual water signal, and averaged subsequently. The averaged spectra from the even scans were subtracted from the averaged odd scans in order to obtain the tChol signal.

| Spectral fitting of 31 P data
The acquired data were analyzed on a group level and on an individual level. Spectral fitting was performed with Lorentzian line shapes in JMRUI 31 using the AMARES 25 algorithm, employing a priori constraints for the chemical shift difference between PE and PC of 0.5 ppm 32 and 0.56 ppm for the chemical shift difference between (diacyl-)glycerophosphatidylethanolamine (GPtE) + GPC and (diacyl-) glycerophosphatidylcholine (GPtC). 33 In addition, a 10-Hz additional line width in the FID spectra for the PDEs, relative to Pi and PMEs, was set to compensate for the much shorter T 2 times of the mobile phospholipid PDE signals in the breast. 22,34 Fibroglandular tissue also contains small amounts of adenosine monophosphate (AMP), 35 with a chemical shift in between PE and PC at physiological pH. 36 The signal of GPE in the breast overlaps partially with the GPtE-MPL (membrane phospholipid) resonance, and concentrations of GPE in the breast are too low to give sufficient signal intensity for quantification. Line widths for PE, PC and Pi were kept free but identical, and line widths for γ-nucleoside triphosphate (γ-NTP) and α-NTP were also kept free but identical. As the spectral resolution in the breast at 7 T is insufficient to distinguish multiplets from J coupling with protons, multiplets were not considered and fitted as single resonances. Spectral fitting with these a priori constraints was used to obtain the values of the line widths and chemical shifts for all metabolites per volunteer.

| T 2 analyses of 31 P metabolites
Metabolite T 2 values were determined based on the weighted averaged group spectra (one FID and five echoes) of the eight volunteers. Each volunteer was measured twice, and so the group-averaged FID is the Pi-weighted average of 16 FID spectra. Weighted averaging was chosen to maximize SNR of the group spectra. Likewise, the group-averaged echo spectra were also weighted with the Pi intensity of the FID. Weighted averaged spectra were subsequently spectrally fitted for PE, PC, Pi, GPC + GPtE, GPtC, PCr, γ-NTP, α-NTP and NAD in JMRUI 31 with the AMARES 25 algorithm. Spectral peak areas of all metabolites in the FID and echoes were fitted mono-exponentially as a function of TE. The spectral amplitudes of PE, PC and GPC + GPtE were also fitted bi-exponentially. For the PMEs, it was assumed that, at short TE, the signal shows contributions from nucleoside monophosphate, for which a T 2 similar to that of γ-NTP was chosen.
The GPC + GPtE resonance was fitted assuming a T 2 of GPtE similar to GPtC.

| Reproducibility
All volunteers were measured twice on the same day (to minimize physiological influences) and the spectral fitting of these two datasets for each volunteer was used to calculate a standard deviation and a coefficient of variation for each metabolite in all volunteers. The standard deviation of the fitted relative peak area (with respect to the total 31 P signal) A m,n of a metabolite m in volunteer n can be written as: where A m;n is the average of the fitted peak area of metabolite m over the two measurements of volunteer n. Subsequently, the coefficient of variation CoV in the fitted peak area of a metabolite m in volunteer n can be written as: Analogously, we can define the average standard deviation and the average coefficient of variation for the peak area of a metabolite m over the whole group of eight volunteers as: and respectively.

| RESULTS AND DISCUSSION
3.1 | T 2 analysis of 31 P metabolites in fibroglandular breast tissue In addition, it should be noted that the peak labeled GPC + GPtE loses substantial intensity from FID to the first echo (GPtE has a short T 2 ), after which the decrease in intensity is much slower, leaving only the slowly decaying GPC signal. The short T 2 PDE signals (GPtC and GPtE) are thought to originate from highly mobile membrane phospholipids. 34,38,39 In the FID spectrum, there appears to be some signal between the peaks labeled PE and PC that seems to disappear in the echoes; this signal could be nucleoside monophosphate (NMP). 35 The β-NTP signal could not be observed above noise level and therefore the spectra are only shown between 10 and −10 ppm. The reason for the apparent absence of the β-NTP signal in the FID spectrum may be two-fold. First, the 95% excitation bandwidth of the AHP pulse is only 1400 Hz and the transmitter offset is approximately 2500 Hz away from the β-NTP resonance. Second, the β-NTP resonance is dependent on the Mg 2+ concentration, which may be low in healthy breast fibroglandular tissue. The group-averaged spectra depicted in Figure 1 allow for a pH estimation of fibroglandular tissue based on the chemical shift of Pi.
As the small PCr signal most probably originates from voxel bleeding from the pectoralis muscle (which may have a B 0 offset because shimming was aimed at the glandular tissue), it is safer to use the signal from GPC and/or α-NTP as chemical shift reference value and take, for these resonances, chemical shift values that have been measured in vivo, in the presence of PCr, which is by definition 0.0. Here, we adopt for GPC and α-NTP, values of 2.97 ppm and −7.57 ppm, respectively, as measured in mammalian brain. 40 For the FID spectrum, we use the α-NTP resonance as it is most clear, whereas the GPC signal in FID is affected by mobile phospholipid; for the echo spectra, we take GPC as chemical shift reference. In this way, we have six measurements (one FID and five echoes) for the chemical shift of Pi which is, on average, 5.31 ± 0.05 ppm, corresponding to a pH value of 7.51 ± 0.07. Figure 2 shows the T 2 fits for the different 31 P metabolites. It should be noted that the FID signal amplitudes of PE + NMP, PC + NMP and GPC + GPtC all appear well above the fitted mono-exponential curve. The additional bi-exponential fit for these resonances shows a better agreement with the data. As the T 2 values of the NMP and GPtC components are not known, we assume a similar T 2 for NMP as for the apparent T 2 of γ-NTP that we measured and, likewise, a similar T 2 of GPtC as for the GPtE component that we measured. An overview of the fitted T2 data is shown in Table 1.
The relative metabolite abundances expressed as a percentage of the total 31 P signal, for this group of volunteers measured twice, as obtained from the data in Figures 1 and 2, are: PE = 8%; AMP = 5%;

| Reproducibility of 31 P measurements
The average relative errors (Equation 4) in the amplitudes of the different metabolite signals, derived from spectral fitting of the FID spectra and the T 2 -weighted (T2w) average spectra, are shown in Figure 3a,b.
Errors on the individual monoester signals are largest because these signals have, on average (see Figure 1), the lowest SNR, overlap partially and are also influenced by the presence of a disturbing signal from NMP. The sum of monoesters, however, shows only small variability. Note the reduction in the relative error and standard deviation of the relative error when making use of T2w spectra, which combine FID and echo spectra, relative to FID spectra, which results in increased SNR and more reliable fitting.  Table 2. In the case of volunteers 3 and 5, one of the two determinations gave insufficient signal to determine a tChol concentration. Spectral fitting was performed with an unconstrained line width for tChol and also with a fixed line width for tChol, equal to the line width of the water signal, whilst adding the (frequency aligned) odd and even scans (i.e. tChol cancels out and residual water and fat peaks do not).

| 1 H spectroscopy
The group-averaged tChol concentration of the data analyzed with a free line width for tChol, as well as with a fixed line width, from ), as shown in Figure 3c, is 34% for the spectral fitting with free line width and 26% for the spectral fitting with fixed line width for tChol.
The interpretation of the relative errors depicted in Figure 3 should also take into consideration the difference in voxel sizes used for 31 P and 1 H spectroscopy. Here, we used 4 × 2 × 4 cm 3 (nominal  3 . In a worst case, this would mean an increase in the coefficient of variation for 31 P spectroscopy of a factor of four on using the 2 × 2 × 2 cm 3 nominal voxel size. As the larger voxel size does not imply that these voxels are completely filled with glandular tissue, they may contain fatty tissue (without 31 P signal) as well, such that the factor of four in the coefficient of variation is the limiting case.
Although the intrinsic sensitivity of 31 P MRS remains less than that  Although the noise level is expected to be the same in all 31 P acquisitions, the noise pattern of the 31 P spectrum in the top right of Figure 5 seems to include a higher spectral noise density than in the remaining spectra. This may be caused by truncation artifacts as a result of the relatively short acquisition window of 16 ms of FID acquisition, which affects the spectral region most strongly around sharp and high-intensity peaks.  Estimates of the average PME and PDE concentrations in healthy fibroglandular tissue based on a combination of 31 P and 1 H MRS, using the 31 P data shown in Figures 1 and 2, and the average tChol value of Table 1