Imaging-based assessment of fatty acid composition in human bone marrow adipose tissue at 7 T: method comparison and in vivo feasibility

Abstract

Purpose

To demonstrate the feasibility and accuracy of chemical-shift-encoded imaging of the fatty acid composition (FAC) of human bone marrow adipose tissue (BMAT) at 7 T, and to determine suitable image acquisition parameters using simulations.

Methods

The noise performance of FAC estimation was investigated using simulations with a range of inter-echo times (ΔTE), and accuracy was assessed using a phantom experiment. Further, one knee of eight knee-healthy subjects (aged 35-54 years) was imaged, and the fractions of saturated fatty acids (SFA) and polyunsaturated fatty acids (PUFA) were mapped. Values were compared between reconstruction methods, and between anatomical regions.

Results

Based on simulations, ΔTE = 0.6 ms was chosen. The phantom experiment demonstrated high accuracy of especially SFA using a constrained reconstruction model (slope = 1.1, average bias = -0.2 %). The lowest accuracy was seen for PUFA using a free model (slope = 2.0, average bias = 9.0 %). For in vivo images, the constrained model resulted in lower inter-subject variation compared to the free model (e.g., in the femoral shaft, the SFA percent-point range was within 1.0 % vs. 3.0 %). Further, significant regional FAC differences were detected. E.g., using the constrained approach, the femoral SFA in the medial condyle was lower compared to the shaft (median [range]: 27.9 % [27.1 %, 28.4 %] vs. 32.5 % [31.8 %, 32.8 %]).

Conclusion

BMAT FAC quantification using chemical-shift encoding is feasible at 7 T. Both the noise performance and accuracy of the technique are superior using a constrained signal model.

Introduction

The adipose tissue of bone marrow (BMAT) is a distinct fat depot with unique properties and functions compared to e.g. subcutaneous and visceral adipose tissue (1). The fatty acid composition (FAC) of the stored fat, and a lower degree of fat unsaturation in particular, has previously been associated to e.g. bone frailty and osteoporosis (2–4), type 2 diabetes (5), and osteoarthritis (6–9). However, previous investigations of the BMAT FAC have mainly been conducted using methods based on magnetic resonance spectroscopy and semi-quantitative measures. To further investigate the role of the FAC in bone health, a high-resolution imaging-based and quantitative method of FAC assessment is needed.

To meet this need, a few studies have implemented the use of magnetic resonance imaging (MRI)-based estimation of the FAC in BMAT (2,10). Using chemical-shift-encoded imaging, spectral components are separated using iterative linear-least squares fitting to multi-echo gradient-echo imaging data while simultaneously correcting for T2* dephasing and off-resonance frequency (B0) (11–13). For estimation of the FAC, the signal amplitudes of individual fat peaks are described as functions of the number of double bonds (ndb), the number of methylene-interrupted double bonds (nmidb), and the chain length (cl), reducing the number of unknowns necessary to characterize the full fat spectrum (14–17). From these, the more generally used fractions of saturated (SFA), monounsaturated (MUFA), and polyunsaturated fatty acids (PUFA) can be easily calculated (17–19).

Chemical-shift-encoded MRI has been developed and evaluated mainly in white adipose tissue (WAT). For example, previous optimization of acquisition strategies has resulted in recommendations for the inter-echo time (ΔTE),which has an impact on quantification accuracy and precision (14,17,20). In reconstruction, alternative signal models where cl and/or nmidb are expressed as empirical functions of the ndb have been suggested to further reduce the number of unknowns and increase robustness (15). For both a free (estimation of ndb and nmidb) and constrained model (estimation of ndb, only), validation has shown strong correlation with the independent gold standard gas chromatography (21).

In comparison to WAT, however, the BMAT may present additional challenges. First, the presence of trabecular bone shortens T2*, and thus the time available for chemical-shift encoding. Second, even in healthy tissue the BMAT may contain a larger amount of water contributing to the signal. Both aspects may impact the optimal acquisition and reconstruction methods. Especially, the optimal ΔTE, total read-out time, and choice between a free and a constrained fat model may be different in BMAT compared to WAT (22).

All FAC imaging has previously been conducted using clinical magnetic field strengths, except for an initial feasibility study in the WAT and liver of mice at 7 T (23). At 7 T, the increased spectral resolution is expected to be an advantage, whereas the shorter T2* may be a challenge.

The aim of this work is to demonstrate the feasibility of BMAT FAC imaging using chemical-shift-encoded MRI at 7 T. Accuracy is tested in phantoms and a pilot in vivo experiment is conducted in the knee of healthy volunteers. In addition, simulations are performed to determine optimal image acquisition parameters for quantitative BMAT FAC imaging at 7 T.

Methods

For all experiments, two reconstruction models (free and constrained) were compared, as well as each of the models’ performance in BMAT and WAT tissues. The reconstruction models were adapted from the works by Trinh et al (21) and Peterson et al (24), and all image, data, and statistical analysis was performed using Matlab (R2020a, MathWorks, Natick, USA).

Signal model

The signals S at the N echo times t from a voxel containing M spectral components with amplitude p_m and frequency f_m may be described according to:

$$\mathit{S}=\mathit{C}\mathit{D}\mathit{A}\mathit{\rho }$$ [1]

$${\mathit{S}}_{\mathit{N}\mathit{x}\mathbf{1}}=\left[\begin{array}{c}{S}_1\\ S_2\\ ⋮\\ S_N\end{array}\right],{\mathit{D}}_{\mathit{N}\mathit{x}\mathit{N}}=\left[\begin{array}{ccc}{e}^{i2\pi \widehat{\psi }{t}_1}& \cdots & 0\\ ⋮& \ddots & ⋮\\ 0& \cdots & e^{i2\pi \widehat{\psi }{t}_N}\end{array}\right]$$ [2]

$${\mathit{A}}_{\mathit{N}\mathit{x}\mathit{M}}=\left[\begin{array}{ccc}{e}^{i2\pi f_1{t}_1}& \cdots & e^{i2\pi f_M{t}_1}\\ ⋮& \ddots & ⋮\\ e^{i2\pi f_1{t}_N}& \cdots & e^{i2\pi f_M{t}_N}\end{array}\right],{\mathit{\rho }}_{\mathit{M}\mathit{x}\mathbf{1}}=\left[\begin{array}{c}{p}_1\\ ⋮\\ p_M\end{array}\right]$$ [3]

$${\mathit{C}}_{\mathit{N}\mathit{x}\mathit{N}}=\left[\begin{array}{ccc}{e}^{{\left(-1\right)}^{1}i{\theta }_{bip}}{e}^{I\left(1\right)i{\theta }_{int}}& \cdots & 0\\ ⋮& \ddots & ⋮\\ 0& \cdots & e^{{\left(-1\right)}^{N}i{\theta }_{bip}}{e}^{I\left(N\right)i{\theta }_{int}}\end{array}\right]$$ [4]

with $\widehat{\psi }$ representing the complex field map including both T2* dephasing and the off-resonance frequency. The matrix C describes the complex errors associated with the bipolar (θ_bip) and interleaved (θ_int) acquisition strategy, respectively (24). The function I is defined as: I_1xN = [1 1 − 1 − 1 1 1 − 1 − 1 …].

FAC quantification

The same 2-step reconstruction algorithm was used for all data. First, the fat and water signals were separated together with B0, a joint T2*, and complex errors θ_bip and θ_int using an iterative least-squares approach (24), and assuming an a priori 8-peak fat model (16). Note that in this work two complex errors were estimated compared to only one in the work by Peterson et al. (24), with otherwise identical methodology. The purpose was to correct both for phase discrepancies between echoes of opposed polarity, and the phase discrepancy between interleaves (25). Second, using the resulting B0, T2*, and complex error maps, a least-squares fitting was used to estimate water, fat, and FAC parameters (ndb, nmidb, and cl). From the estimated FAC parameters, SFA and PUFA were calculated (17). In both reconstruction steps, the first 10 echoes were used.

For the second step, two previously described reconstruction models based on gas chromatography data in WAT (21) were compared: 1) a free model estimating ndb and nmidb as free parameters, using the following matrices and E_m(t) = e^i2πf_mt):

$${\mathit{A}}_{\mathit{N}\mathit{x}\mathbf{4}}=\left[\begin{array}{cccc}1& a\left(1\right)& b\left(1\right)& c\left(1\right)\\ ⋮& ⋮& ⋮& ⋮\\ 1& a\left(N\right)& b\left(N\right)& c\left(N\right)\end{array}\right],{\mathit{\rho }}_{\mathbf{4}\mathit{x}\mathbf{1}}=\left[\begin{array}{c}W\\ kF\\ kFndb\\ kFnmidb\end{array}\right]$$ [5]

$$\begin{array}{lll}a\left(t\right)\hfill & =\hfill & E_1\left(t\right)+4E_2\left(t\right)+6E_4\left(t\right)+6E_6\left(t\right)+73.92E_7\left(t\right)+9E_8\left(t\right)\hfill \end{array}$$ [6]

and 2) a constrained model estimating the ndb only, assuming nmidb = 0.45ndb − 0.71(21), resulting in:

$${\mathit{A}}_{\mathit{N}\mathit{x}\mathbf{4}}=\left[\begin{array}{ccc}1& a\left(1\right)& b\left(1\right)\\ ⋮& ⋮& ⋮\\ 1& a\left(N\right)& b\left(N\right)\end{array}\right],{\mathit{\rho }}_{\mathbf{3}\mathit{x}\mathbf{1}}=\left[\begin{array}{c}W\\ kF\\ kFndb\end{array}\right]$$ [7]

$$\begin{array}{lll}a\left(t\right)\hfill & =\hfill & E_1\left(t\right)+4E_2\left(t\right)-1.42E_3\left(t\right)+6E_4\left(t\right)+2.84E_5\left(t\right)+6E_6\left(t\right)\hfill \\ \hfill & \hfill & +72.50E_7\left(t\right)+9E_8\left(t\right)\hfill \end{array}$$ [8]

In both models, cl was set to cl = 16.32 + 0.38 ndb (21) for in vivo data and to cl = 14.25 + 1.12ndb for phantom data to correct for the lower cl of butter compared to human tissue (15). The altered phantom cl model was based on reference FAC values of butter and margarine provided by the Swedish National Food Administration.

Phantom construction

For assessment of quantification accuracy, oil (long T2*) and butter (short T2*) was used to mimic WAT and BMAT. Fourteen phantom vials were prepared with 1-3) vegetable oils (100 % fat) 4-7) clarified butter (100 % fat), 8-11) butter (80 % fat), 12-13) low fat butter (40 % and 60 % fat), and 14) water. For each phantom set, a range of FACs was chosen. For oils SFA = 12 % – 18 % and PUFA = 7 % – 69 %, for butters SFA = 23 % – 70 % and PUFA = 3 % – 40 %. Although both butters and margarines were used, they will all be referred to as butter for simplicity.

Noise performance

The noise performance was investigated as the maximum effective number of signals averaged (NSA), a parameter which describes how efficiently the signal from several echoes are used in fitting with a maximum value equal to the number of echoes (26). The NSA was calculated analytically from the Cramer-Rao bound, estimated using the Slepian-Bangs formula assuming Gaussian noise (27,28). The NSA was estimated for a range of ΔTEs (0 ms - 3 ms), comparing BMAT with T2* = 5 ms, and WAT with T2* = 35 ms, and with proton density fat fraction (PDFF) = 95 %, ndb = 2.7 (with nmidb and cl calculated according to (21)), off-resonance frequency = 40 Hz, θ_int = θ_bip = 0.01 + 0.01i, 10 echoes, and a first echo time of 1.2 ms.

Human subjects

For the 7 T in vivo feasibility investigation, eight healthy volunteers (aged 35-54 years, 3 male, 5 female) were recruited after written informed consent and approval of the regional ethics committee. One knee of each subject was randomly selected for imaging.

MR imaging

Imaging was conducted using a transmit and receive QED Knee Coil 1TX / 28RX and a 7 T MR scanner (Philips Achieva AS, Best, the Netherlands). Two sagittal multi-echo gradient echo sequences covering the knee joint (or phantom vials) were acquired with interleaved echo times and bipolar read-out gradients to reduce the ΔTE. The following imaging parameters were used: TE₁ = 1.2 ms, effective ΔTE = 0.6 ms (ΔTE = 1.2 ms per sequence), total number of echoes = 16 (of which the first 10 were used in reconstruction), TR = 30 ms, pixel bandwidth = 1378 Hz, and acquired voxel size 0.8x0.8x3 mm³. Using SENSE with an acceleration factor of 1.2, the total acquisition time of each of the two sequences was 2 minutes and 8 seconds. A flip angle of 8° was chosen as a compromise between SNR and T1 weighting.

The inline mDixon Quant application (Philips Achieva AS, Best, the Netherlands) was used to generate magnitude and phase images, as well as a B0 map, which was used as an initial guess in the iterative algorithm for FAC quantification to avoid fat/water swaps. For each data set, quantitative parameter maps were reconstructed of PDFF, T2*, SFA, and PUFA using the method described above.

Image analysis

A region-of-interest (ROI) was manually defined in each phantom vial. For in vivo images, three sagittal slices were chosen for each subject: a central slice and two slices centered over the medial and lateral femoral condyles, respectively. In these, a total number of nine ROIs were manually delineated: Femoral and tibial shaft, patella, infra-patellar fat pad and the posterior subcutaneous adipose tissue (central slice), femoral and tibial lateral condyle (lateral slice), and finally the femoral and lateral medial condyle (medial slice). To avoid any extreme values, voxels with T2* < 1 ms and PDFF < 40 % were excluded prior to estimating the average of each of the T2*, SFA, and PUFA parameters within each ROI. For estimation of average PDFF, only an exclusion criterion of T2* < 1 ms was used. The median and interquartile range for each parameter were presented for the subject group. Due to a very thin layer of subcutaneous adipose tissue, this ROI was excluded for three subjects and the median and subject ranges were presented instead.

Statistical analysis

Wilcoxon signed rank tests were performed to evaluate if there was a statistically significant difference in the FAC parameters between the free and constrained models, and if FAC parameters were different in the various ROIs compared to the femoral shaft. p < 0.05 was considered a statistically significant difference. No statistical comparisons were conducted in the subcutaneous adipose tissue due to fewer data points.

The accuracy of the butter phantoms with PDFF = 100 % and PDFF = 80% was assessed using the regression slope and the average bias to true values. As the FAC range of oils were different compared to butter, no comparison of regression slope and average bias between oils and butter was made.

Results

Noise performance

NSA was lower in BMAT compared to WAT, especially using longer ΔTEs, indicating a higher impact of noise on estimated FAC parameters in this tissue (Figure 1). Comparing the free and constrained fat models in reconstruction, the constrained model yielded a higher NSA compared to the free model. Based on these results, a ΔTE of 0.6 ms was chosen for further experiments.

Figure 1.

Simulated number of signal averages (NSA) as function of inter-echo time (ΔTE), where a higher NSA represent a better noise performance of the method. The constrained approach demonstrated a better noise performance compared to the free approach. Based on the results, ΔTE = 0.6 ms was chosen for further experiments.

Phantom results

Butter was a relevant reference for BMAT tissue with high PDFF and short T2*. The mean (standard deviation) T2* was 45 ms (21 ms) in oil phantoms, 13.2 ms (5.7 ms) in 100 % butter, 6.7 ms (1.1 ms) in 80 % butter, and 8.5 ms (1.7 ms) in low fat butter. PDFF was estimated to 99.8 % (1.0 %) in oil, 97.8 % (0.7 %) in 100 % butter, 81.6 % (1.5 %) in 80 % butter, and 43.9 % (1.3 %) and 66.1 % (0.4 %) in low fat butter.

In 100 % butter phantoms, high agreement of the estimated SFA to true values was seen, with slopes close to 1 and small average biases (Figure 2), whereas accuracy of PUFA was lower. Overall, the accuracy using a constrained approach was higher compared to a free model in butter phantoms.

Figure 2.

Results of the oil and butter phantom experiment using a constrained (left column, a, c, e) and free reconstruction model (right column, b, d, f). Each row of vials is described as type of phantom (oil/butter) and PDFF. Mainly homogenous estimation is seen in the SFA and PUFA image examples in a) and b). The average value within each ROI is compared to true values in scatter plots (c, d, e, and f), and regression slopes and average bias for 100 % butter and 80 % butter phantoms is displayed in each plot. The black line represents a perfect agreement to reference values. Especially SFA in high PDFF butter phantoms is estimated with high accuracy. In e), PUFA in the PDFF = 60 % butter phantom was out of range, estimated to -67 %.

Accuracy of SFA was affected by water content, with lower accuracy in 80 % compared to 100 % butter. In PUFA, little difference was seen between 80 % and 100 % butter. Both SFA and PUFA results was more affected by the choice of reconstruction model in the two low PDFF butter vials, with PDFF-dependent results especially using a free model and for PUFA.

As the reconstruction model was not optimal for vegetable oil, the lower accuracy in oil compared to butter was expected. In oil, a closer agreement to true values was obtained using the free model.

In vivo results

Example in vivo images are presented in Figure 3 and show promising image quality, especially using a constrained signal model. Image quality was superior in the SFA maps compared to PUFA, which appeared less precise. A mainly homogenous estimation was seen in the femoral and tibial BMAT, whereas a larger, likely artefactual, spatial variation was apparent in the subcutaneous adipose tissue. Example MUFA maps are displayed in Figure S1 and complex field and error maps in Figure S2. MUFA results showed a marked difference between the two reconstruction models.

Figure 3.

Quantitative in vivo image examples. From calculation step 1 (a), PDFF and T2* were mapped. From calculation step 2, SFA and PUFA were mapped using either a free (b) or constrained (c) reconstruction model. The image quality of especially SFA and using the constrained approach is promising with stable estimation within BMAT areas.

In Figure S3, the corresponding quantitative images are shown for a subject with a lower fat content in the BMAT of the femoral shaft. The assessment of SFA and PUFA appears robust across the margin between red and yellow bone marrow. The SFA / PUFA values in the femoral shaft for this subject were: 32.3 % / -3.4 % for the free model and 37.0 % / 8.5 % for the constrained model.

Numerical results within ROIs are shown in Figure 4 and Table 1. The PDFF and T2* were estimated in a joint first step using a basic fat/water separation and were therefore the same for the free and constrained models. As expected, both the tibial and femoral condyles showed significantly faster T2* dephasing, and the subcutaneous and infrapatellar adipose tissue regions showed significantly slower dephasing compared to the femoral shaft.

Figure 4.

Subject group median and inter-quartile range of in vivo SFA (top row) and PUFA (bottom row) results in femoral ROIs (left column), tibial ROIs (middle column), and other anatomical regions (right column), comparing the free and constrained reconstruction models. (For subcutaneous adipose tissue (SCAT), the full range is shown due to a small number of data points N = 5.) The constrained model appears to yield more robust results with more narrow inter-subject variation.

Table 1. In vivo subject group median (interquartile range) of the number of voxels within each ROI, as well as the PDFF and T2* values from reconstruction step 1, and SFA and PUFA values from reconstruction step 2 using either the free or constrained reconstruction models.

				SFA (%)			PUFA (%)
	Number of voxels	PDFF (%)	T2* (ms)	Free	Constrained	p-value	Free	Constrained	p-value
Femur
Shaft	3168 (2961, 3469)	97.1 (96.3, 97.2)	6.6 (5.8, 7.5)	33.2 (32.5, 35.2)	32.5 (31.8, 32.8)		13.2 (9.1, 16.0)	12.2 (11.9, 12.7)
Medial condyle	3598 (3185, 3990)	93.5 (92.2, 94.3)	2.8 (2.7, 3.1) **	30.9 (28.7, 31.1) **	27.9 (27.1, 28.4) **	††	18.0 (16.5, 19.1) **	16.0 (15.6, 16.6) **
Lateral condyle	3551 (3154, 4016)	95.1 (94.3, 95.8)	3.0 (2.4, 3.1) **	29.8 (28.4, 30.9) **	27.6 (27.0, 28.3) **	††	17.3 (16.3, 18.2) *	16.2 (15.7, 16.7) **
Tibia
Shaft	3965 (3672, 4293)	97.8 (97.3, 98.2) **	6.7 (5.9, 8.1)	30.5 (29.7, 32.2) *	31.8 (30.3, 32.4)	†	8.9 (8.6, 10.4)	12.8 (12.3, 14.0)	††
Medial condyle	1323 (1103, 1515)	93.2 (90.7, 94.5)	2.8 (2.5, 3.2) **	31.2 (29.8, 33.4) *	30.4 (29.4, 31.5)		11.1 (9.7, 13.9)	13.9 (13.0, 14.8)
Lateral condyle	1544 (1246, 1897)	94.8 (93.3, 95.9)	3.3 (2.6, 3.5) **	30.9 (29.2, 33.0)	30.4 (29.6, 30.9) *		10.8 (9.2, 13.6)	13.9 (13.5, 14.6) *
Adipose tissue
Subcutaneousa	907 (531, 2060)	97.0 (60.3, 98.0)	24.6 (9.6, 32.0)	31.2 (27.1, 33.9)	34.4 (32.9, 40.2)	NA	0.9 (-8.1, 3.8)	10.6 (5.9, 13.0)	NA
Infrapatellar fat pad	1213 (1043, 1439)	90.8 (89.3, 91.3)	14.5 (13.0, 15.9) **	31.2 (28.7, 34.0)	37.3 (34.8, 38.3) **	††	-4.7 (-7.4, -2.1) **	8.3 (7.4, 10.3) **	††
Patella	1524 (1418, 1713)	91.1 (87.8, 92.9)	2.4 (2.1, 2.6) **	31.4 (30.2, 33.3) *	28.1 (27.0, 29.1)		20.7 (19.6, 21.8) **	15.8 (15.0, 16.7)	†

Note: For subcutaneous adipose tissue the range is presented instead of quartiles due to fewer data points. No statistical testing was performed for this ROI.

^*is used to denote a significant difference from the results in the femoral shaft (gray).

^†is used to denote a significant difference between the results using the free and constrained models.

*/† p <0.05 and **/†† p < 0.01

For FAC values, both the SFA and PUFA were significantly different comparing the free and constrained models in several anatomical regions. It was also apparent that the estimation using the constrained model was more robust with lower inter-subject variability.

Both models demonstrated lower SFA, and higher PUFA in both femoral condyles, as well as lower PUFA in the infrapatellar fat pad compared to the femoral shaft. However, the PUFA in the two adipose tissue depots were negative in some cases using the free reconstruction model, indicating a systematic bias of these results.

Discussion

In this study, the feasibility and accuracy of MRI-based quantification of the BMAT FAC has been demonstrated at 7 T, and suitable imaging parameters have been investigated using simulations. For BMAT FAC assessment, a short ΔTE and a constrained reconstruction model yielded the most accurate and noise efficient results. To our knowledge, this is the first presentation of human in vivo FAC maps using this technique at 7 T.

Literature FAC values for the distal femur and tibia are scarce, but a few comparisons to our results are possible. Compared to previous results in the BMAT of the femoral head using a similar MRI-based technique (SFA = 50.7 %, PUFA = 14.1 %), our estimated SFA in the femoral shaft is markedly lower (10). However, both our SFA and PUFA results in the femoral shaft agree with literature gas chromatography results in femoral head BMAT (SFA = 33.7 %, PUFA = 11.2 %) (30).

In WAT (SFA = 29.5 %, PUFA = 16.5 %), PUFA is underestimated using the free estimation model, whereas SFA is similar to gas chromatography data. The results are slightly overestimated using the constrained model (31).

The used ΔTE (0.6 ms) and total read-out time (6.6 ms) in this study were in accordance with those previously presented for WAT at 3 T, which at 7 T would correspond to a maximum ΔTE of 0.8 ms (14) or 0.5 ms (20) and a total read-out time of 6 ms – 7.8 ms (14,20). Although 16 echoes were acquired, only the first 10 echoes were used, to reduce bias from any model inaccuracies (20).

In this work, constraints on cl and nmidb were based on empirical linear relationships to ndb obtained from gas chromatography data in WAT (21). Alternatively, a full quadratic nmidb model suitable for a larger range of FACs may be beneficial but requires modifications of the reconstruction algorithm (15). Although a model based on data in BMAT should be investigated in future work, a model based on WAT was considered sufficient in this feasibility study. Given the comparatively limited difference in FAC between various human tissues, the impact on FAC accuracy in vivo is likely small.

In phantoms with a large range of FAC, however, inaccuracies of the linear nmidb and cl models likely contribute to systematic bias. Although butter and oil were relevant mimics of BMAT and WAT in terms of T2* and PDFF, butter has a markedly lower cl, and vegetable oils higher ndb and nmidb in comparison to human tissue (15). Although the cl model was adjusted to better reflect the properties of butter, the used models were not optimal for quantification in vegetable oils. For this reason, it is difficult to compare the accuracies in short T2* (butter) and long T2* (oil) phantoms, based on the current results.

Previous work using the constrained approach has demonstrated unreliable results of the estimated MUFA map with close to constant values (21). However, the PUFA and especially the SFA results were correlated to GC results. Given this known limitation of the technique, MUFA results were not analyzed further in this work.

The low PDFF case, explained by the presence of red bone marrow in the femoral shaft due to either red marrow persistence or reconversion, offered an opportunity to observe the performance of the technique in the presence of a higher water signal. However, it was not possible to draw conclusions on accuracy and precision from this single subject. Based on the phantom experiment, less reliable results were obtained in low PDFF compared to 100 % PDFF butter phantoms. Further assessment of the method’s performance in lower PDFF is important in future work.

This study has several limitations. First, no reference values for comparison were available for human BMAT tissue in the distal femur, which limits the ability to draw conclusions on in vivo estimation accuracy. For this reason, a phantom experiment was conducted to assess accuracy in short T2* butter, demonstrating high accuracy of especially the SFA parameter and using a constrained reconstruction model. However, due to difficulties in constructing phantoms with representative properties to human BMAT tissue, in vivo validation against independent gas chromatography measurements is important future work. Secondly, only healthy volunteers, and in a limited age range, were recruited and the ability of the technique to detect pathology could thus not be investigated. Finally, a comparison between 3 T and 7 T was beyond the scope of the present work, and no conclusion can be drawn on which is preferred. A comparison between field strengths is an important topic for future work.

Conclusions

In conclusion, BMAT FAC quantification using chemical-shift encoding is feasible at 7 T, with a similar acquisition protocol used for quantification in WAT. Especially, SFA is accurately estimated from short T2* and high PDFF signals, and both the noise performance and accuracy of the technique are superior using a constrained signal model.