Vanderbilt University Institute of Imaging Sciences, Department of Radiology and Radiological Sciences, Vanderbilt University School of Medicine, Nashville, Tennessee, USA
Division of Gastroenterology, Hepatology, and Nutrition, Department of Medicine, Vanderbilt University School of Medicine, Nashville, Tennessee, USA; Vanderbilt University Institute of Imaging Sciences, Department of Radiology and Radiological Sciences, Vanderbilt University School of Medicine, Nashville, Tennessee, USA
Improved understanding of how depot-specific adipose tissue mass predisposes to obesity-related comorbidities could yield new insights into the pathogenesis and treatment of obesity as well as metabolic benefits of weight loss. We hypothesized that three-dimensional (3D) contiguous “fat-water” MR imaging (FWMRI) covering the majority of a whole-body field of view (FOV) acquired at 3 Tesla (3T) and coupled with automated segmentation and quantification of amount, type, and distribution of adipose and lean soft tissue would show great promise in body composition methodology.
Design and Methods:
Precision of adipose and lean soft tissue measurements in body and trunk regions were assessed for 3T FWMRI and compared to dual-energy X-ray absorptiometry (DXA). Anthropometric, FWMRI, and DXA measurements were obtained in 12 women with BMI 30-39.9 kg/m2.
Test–retest results found coefficients of variation (CV) for FWMRI that were all under 3%: gross body adipose tissue (GBAT) 0.80%, total trunk adipose tissue (TTAT) 2.08%, visceral adipose tissue (VAT) 2.62%, subcutaneous adipose tissue (SAT) 2.11%, gross body lean soft tissue (GBLST) 0.60%, and total trunk lean soft tissue (TTLST) 2.43%. Concordance correlation coefficients between FWMRI and DXA were 0.978, 0.802, 0.629, and 0.400 for GBAT, TTAT, GBLST, and TTLST, respectively.
While Bland–Altman plots demonstrated agreement between FWMRI and DXA for GBAT and TTAT, a negative bias existed for GBLST and TTLST measurements. Differences may be explained by the FWMRI FOV length and potential for DXA to overestimate lean soft tissue. While more development is necessary, the described 3T FWMRI method combined with fully-automated segmentation is fast (<30-min total scan and post-processing time), noninvasive, repeatable, and cost-effective.
Quantifying the amount, type, and distribution of adipose and lean soft tissue is critical to understanding not only the pathological roles of these tissues in the metabolic consequences of obesity but also the metabolic benefits of dietary, pharmacological, and surgical weight loss interventions (1, 2). While more sophisticated methods for body composition analysis would be applicable to numerous research questions, our research interests are focused on obesity in women as adipose tissue distribution and its metabolic consequences are quite different between women and men. Moreover, as women age, the percentage of body fat and amount of visceral adipose tissue (VAT) increases significantly (3). The increasing adiposity observed over the life cycle in women is an important factor in the lack of achievement of the same successes that have occurred in men with regard to reduced cardiovascular risk over the past few decades. It is conceivable that the gaps in the evidence to explain the biological underpinnings of cardiometabolic risk may be due to the relative strengths and weaknesses of information available from current body composition assessment techniques.
Imaging methods such as dual energy X-ray absorptiometry (DXA), magnetic resonance imaging (MRI) and computed tomography (CT) are considered the most accurate for quantifying body composition (4). Because DXA scanners are widely available, they are used most frequently to estimate body composition. However, since DXA acquires data as a two-dimensional coronal projection it does not distinguish type of fat, e.g., VAT vs. subcutaneous adipose tissue (SAT), and it is not able to discern ectopic fat. Further, it has been demonstrated that the error in DXA derived estimation of total and regional fat mass (typically overestimation) increases in proportion to the amount of body fat mass (adiposity) and abdominal sagittal thickness (5). In addition, the trunk region typically determined using DXA includes chest, abdomen, pelvis and some gluteal fat, which confounds measures in obese women (6). Even with manipulation of the region of interest, DXA tends to estimate intra-abdominal adipose tissue best in nonobese persons (7). With regard to lean soft tissue estimation, DXA has been shown to overestimate skeletal muscle mass, partly due to the effects of hydration status (8, 9).
While both CT and MRI provide quantification of tissue volume, MRI is preferred because, unlike CT, it does not expose individuals to ionizing radiation. In fact, the potential of MRI for assessment of amount, type, and distribution of adipose tissue has long been recognized (10), although low-sensitivity and long-scan times have limited widespread usage. In more recent investigations, multigradient-echo MRI acquisitions have been used with 1.5 Tesla (1.5T) scanners to acquire body composition data sets more quickly than standard MR anatomical imaging protocols (11). However, the use of multi-slice MRI methods with large inter-slice gaps requires interpolation and geometric assumptions to calculate tissue volumes (12). We suggest that such MRI methods could be optimized to sample contiguously almost the entire body with no gap between slices, thus, eliminating the necessity to interpolate data.
In fact, conventional T1-weighted MRI has been validated for quantification of adipose tissue in cadaver studies (13) and used successfully in vivo (10, 12, 14). However, T1-weighted spin-echo or T1-weighted spoiled gradient-echo scans attempt to discriminate adipose tissue vs. nonadipose tissues based on signal intensity because the relatively short T1 (longitudinal magnetization relaxation rate time constant) of adipose tissue causes it to be bright on T1-weighted scans. As other tissues such as blood vessels and bowel contents may also appear bright on T1-weighted images, semiautomatic segmentation may be necessary to eliminate segmentation errors. In contrast to T1-weighted imaging, MRI-based chemical shift imaging, termed “fat-water MRI,” is molecule-specific and takes advantage of the small but measurable differences in the resonance frequencies of water and lipid protons. With this method, high fat content is more specific to adipose tissue than obtained with short T1. Fat-water MRI can be used to assess body tissues with a three-dimensional (3D) (no gap between slices) multigradient-echo MRI pulse sequence. Notably, such a contiguous acquisition can be combined with fully automated segmentation of adipose and lean soft tissue as well as quantification of abdominal visceral and subcutaneous depots (15, 16). Moreover, the ability to use higher magnetic field strength 3 Tesla (3T) scanners offers the potential for improved signal-to-noise ratio and contrast-to-noise ratio, which can increase spatial resolution and/or reduce scan time (17).
Our overarching hypothesis is that utilization of fat-water MRI (FWMRI) to identify and quantify adipose and lean soft tissue will provide superior body composition information beyond what can be acquired from current imaging methods. As a critical first step, in this study we determined the precision of FWMRI in a manner similar to that recently reported for other novel magnetic resonance approaches (18). Thus, we rigorously compared repeated measurements of adipose and lean soft tissue in obese women acquired by FWMRI to comparable regions acquired by DXA, the most frequently used imaging tool for assessment of body composition.
Methods and Procedures
Twelve women with BMI 30-39.9 kg/m2 who were newly enrolled in a clinical trial for weight loss were recruited to have FWMRI and DXA scans prior to initiation of a calorically restricted diet. All subjects provided written informed consent and the study was approved by the Vanderbilt University Institutional Review Board. Subjects had no history of cancer, diabetes, cardiovascular, liver or kidney disease; no weight change of >2.3 kg within 3 months; no metal implants or other contraindications for MRI; and tested negative for pregnancy immediately before scanning. Acquisition of anthropometry, DXA, and FWMRI data from each subject was obtained at the Vanderbilt Clinical Research Center (CRC) and the Vanderbilt University Institute of Imaging Science (VUIIS) within a 4-h interval in the morning. Subjects had been instructed to avoid alcohol, excessive caffeine intake, and physical activity for 24 h before study visits, and to fast from 9:00 PM until arrival at 7:00 AM. Subjects remained in the fasted state and emptied their bladders before all scans.
Subjects' height was obtained without shoes to the nearest 0.1 cm on a wall-mounted stadiometer, and weight was measured while wearing medical scrubs on a digital platform scale to the nearest 0.1 kg. Waist and hip circumferences (±0.1 cm) were measured by positioning a flexible measuring tape above the right iliac crest and at the full extension of the buttocks, respectively. Two derivative indices of adiposity, BMI from dividing weight by height squared and body adiposity index (BAI) from dividing hip circumference by height squared (19) were calculated from the anthropometric data.
Dual energy X-ray absorptiometry
DXA measurements were acquired by a certified densitometrist using a Lunar iDXA whole-body scanner (GE Healthcare, Madison WI). Analysis of DXA images was performed using enCORE 2007 software version 11.40.004. Before image acquisitions, total body phantom calibration was performed according to manufacturer procedures. Subjects removed all extraneous personal items that would interfere with the accuracy of DXA results. No subjects had body width that exceeded the dimensions of the scanner window. Scan time ranged from 6.5-12.5 min depending on DXA-calculated abdominal sagittal thickness (20).
Subjects were scanned on a Philips Achieva 3T scanner (Philips Healthcare, Best, the Netherlands) equipped with a dual gradient set capable of 40 mT/m peak strength and 200 mT/m/ms peak slew rate in the serial operation mode. A table-top extender was used to extend the stroke of the patient couch in the head-to-foot direction to enable an increased FOV length to 1.632 m (∼5 feet 4 inches) in the superior–inferior direction. Subjects entered the MRI scanner feet first in the supine position with arms bent above the head using pillow and table supports. To monitor respiratory activity and breath-hold compliance, an air-filled bellows was placed on the subject's abdomen with a Velcro strap. Breath-holds were performed for 15 s each at table positions from the shoulders to the waist.
The integrated quadrature body coil was used for transmit and receive of MRI sequences. The multi-station protocol included 17 table positions. Each of the 17 stacks comprised a multi-slice, multigradient-echo (known as multiple fast field echo) acquisition with 12 8-mm slice with slice gap set to zero. Other acquisition details included repetition time = 75 ms; echo times (TE) of TE1/TE2/TE3 (ms) = 1.34/2.87/4.40; flip angle (FA) = 20º; water-fat shift = 0.325 pixels; readout sampling bandwidth = 1,335.5 Hz/pixel; axial in-plane field of view (FOV) = 500 mm × 390 mm; acquired in-plane matrix size = 252 × 195; and acquired voxel size = 2 mm × 2 mm× 8 mm. Automated linear shimming was performed for each slice stack. Flyback gradients were employed between echoes so that the chemical shift direction for all echo readouts was the same. Data acquisition time averaged 4 min and 16 s with ∼5 additional minutes for table movement, preparation phases at each table position, and breath-holding pauses. Subjects' total time in the scanner including subject positioning, scout imaging and image planning was 20-25 min.
Regions of interest
From DXA and FWMRI acquisitions, we report four tissue compartments: gross body adipose tissue (GBAT), total trunk adipose tissue (TTAT), gross body lean soft tissue (TLST), and total trunk lean soft tissue (TTLT). With FWMRI, GBAT and GBLST included all tissue visible in the scanner FOV. We also report SAT and VAT from FWMRI. We manually customized the DXA trunk region of interest to match the upper and lower anatomical boundaries of the visceral region delineated by the automated MR image segmentation algorithm, which defined the abdominal subvolume as extending from the base of the lungs to the femoral heads. Figure 1 displays the coronal DXA image (on the left) and 1st echo maximum intensity projection from the FWMRI acquisition (on the right) for subject #1. The DXA region of interest in the head-to-foot direction was manually set to match the abdominal subvolume identified by the FWMRI segmentation algorithm by placing a rostral horizontal line at the lower base of the lung and a caudal horizontal line at the top of the femoral head. The left-to-right boundaries of the region of interest were adjusted to exclude as much of the arms as possible.
Fat-water MRI processing
Complex (real and imaginary) images for each of the acquired three echo times were reconstructed and exported in Digital Imaging and Communications in Medicine standard format for off-line processing. Separation of water and fat signal requires knowledge of the main magnetic field map, and essentially all water/fat separation algorithms rely on an assumption of spatial smoothness of the main magnetic field. Water and fat images were reconstructed using a recently published algorithm by Berglund et al. (21) in which fat and water signal components were found by least squares fitting (22) after resolving the magnetic field map using a 3D multiple seed-point region growing algorithm with a dynamic path sensitive to field in homogeneity, solving regions of homogeneous magnetic field before inhomogeneous or noisy regions. Berglund's 3-point Dixon algorithm, which directly finds two analytical solutions for the background static field phasor term, is valid as long as the echo spacing between the 1st and 2nd echoes matches the spacing between the 2nd and 3rd echo. When applied to whole-body 3D data sets acquired at 1.5T, the method was found by two blinded radiologists to have the best image quality (in terms of fat water “swaps”) compared to two reference methods: an advanced version of the “iterative decomposition of water and fat with echo asymmetric and least-squares estimation” IDEAL approach (23) as well as a multi-resolution golden section search method (24). The employed fat-water separation algorithm used a multi-peak model with nine peaks set with fixed spectral locations and relative amplitudes. Values of the chemical shifts and of the relative amplitudes were taken from a liver spectroscopy study of 121 human subjects (25). The reconstructed fat and water image volumes were input into a fully automated segmentation algorithm (16) that first localized the abdominal subvolume and then separated VAT from SAT within the identified abdominal subvolume. Lean soft tissue was quantified by measuring the total volume of the voxels with fat signal fractions <50%. Adipose and lean soft tissue volumes were multiplied by published density values of 0.923 and 1.05 kg/l, respectively, to convert volume to mass (kg) (13, 26). Typical post-processing times for the fat-water separation and automated segmentation algorithms were two and 3 min, respectively for the size of image files in this study. Figure 2 presents a single coronal slice (top left), single sagittal slice (top middle), a montage of 20 axial slices (top right), and a histogram of fat signal percentage for all voxels in the 3D MRI fat volume of subject #5. As seen in these images, VAT is displayed in red. Across the axial slices, the identified VAT voxels are spatially consistent and are not contaminated by SAT or other voxels. Importantly, the algorithm is designed to automatically exclude bone marrow and intramuscular adipose tissue in the area of the spine, which would confound adipose tissue measures. On the scale of fat signal percentage in the displayed histogram, two distinct populations of voxels are discernable. The populations are widely separated with long tails. Applying the well-known Otsu method (27) to determine an optimal threshold that simultaneously minimizes intraclass variance and maximizes interclass variance yields an optimal threshold of 49.8% for this example data to distinguish the fatty and lean voxels which is very close to the employed threshold of 50%. Additional evidence of the robustness of the fat-water imaging approach and automatic VAT segmentation method is provided by six movies showing fat and water MR image volumes of one subject that are included as Supplementary Videos S1-S6 online at the journal website. Fat images appear on the left in the coronal (see Supplementary Videos S1 and S2 online) and sagittal (see Supplementary Videos S3 and S4 online) orientations or on the top for the axial orientation (see Supplementary Videos S5 and S6 online). Three animations (see Supplementary Videos S2, S4, and S6 online) display the segmented VAT as red voxels superimposed on both the fat and water images.
Statistical analyses were conducted using SPSS version 19.0 (IBM, Armonk, NY) and MATLAB version 184.108.40.2069 R2010a (MathWorks, Natick, MA) with the level of significance set at α = 0.05. Data are presented as mean ± SD. As values were not normally distributed, Wilcoxon signed-rank tests were performed to assess the significance of the observed differences between DXA and FWMRI for TAT, TTAT, TLST, and TTLST. Percent differences were calculated as (FWMRI – DXA)/DXA × 100%.
Coefficient of variation
To determine the precision of DXA and FWMRI methods, we performed test–retest measurements and then calculated coefficients of variation (CV) for the tissue masses estimated by each modality. CV was determined from the square root of the mean within-group variance of a one-way ANOVA (with each subject as a group) divided by the grand mean (28), and multiplied by 100 to express CV as a percentage. Thereafter, all analyses were performed using the averaged value of the test–retest data. As complete retest FWMRI scans were not available for subjects #5 and #7, FWMRI values for these subjects represent one measurement only and they were excluded in CV analysis.
The concordance correlation coefficient (ρc), often used in comparing quantitative methods, measures accuracy by determining how much the linear relationship between two variables deviates from the line of concordance and precision by how far each observation deviates from the fitted line (29). The more commonly known Pearson correlation coefficient (ρr) is related to ρc via the bias correction factor (Cb), which is a ratio of the two coefficients (ρc/ρr) (30). The value of Cb is within the range of 0 to 1, and thus, measures how far the best-fit line deviates from the 45° line (accuracy). Only using the Pearson correlation coefficient would measure how far each observation deviates from the best-fit line (precision) but not accuracy. Here we report ρc, bias correction factors and asymptotic 95% confidence intervals for ρc. Spearman's Rho correlation coefficients (ρs) were also calculated to identify any monotonic relationships between adipose and lean soft tissue discrepancies (difference between DXA and FWMRI) with anthropometric measures.
The Bland–Altman method (31) was used to assess the limits of agreement (mean difference ± 1.96 SD of the difference) between DXA and FWMRI measures. As such the difference between the methods (y-axis) was plotted against the mean of the methods (x-axis). Linear dependence trends in the Bland–Altman plots were calculated using a least squares fit to a first order polynomial. The coefficient of determination (r2) and significance level of each linear fit were also calculated.
Descriptive characteristics of the 12 female subjects are presented in Table 1. Five of the women self-identified as Caucasian and seven as African American, but no significant differences were detected by ethnicity. Subjects had an average BMI of 34.3 ± 2.8 kg/m2 and an average BAI 39.5 ± 4.8% indicating class I obesity.
Table 1. Demographic characteristics of female subjects (N = 12)
Average adipose and lean soft tissue masses measured by DXA and FWMRI are presented in Table 2 for each subject. For adipose tissue, mean GBAT was 39.9 ± 7.5 kg by FWMRI and 39.8 ± 6.9 kg by DXA, and mean TTAT was 14.0 ± 3.4 kg by FWMRI and 15.7 ± 3.5 kg by DXA. The mean FWMRI measurements for SAT and VAT were 11.6 ± 2.6 kg and 2.1 ± 1.0 kg, respectively. For lean soft tissue, mean GBLST was 39.8 ± 5.4 kg by FWMRI and 45.8 ± 7.3 kg by DXA, and mean TTLST was 11.9 ± 1.7 kg by FWMRI and 15.1 ± 2.7 kg by DXA.
Table 2. Average mass (kg) for adipose and lean soft tissue depots
Coefficient of variation
CVs for GBAT, TTAT, GBLST, and TTLST are presented and show that reproducibility measurements were <3% for DXA and FWMRI (Table 2). For DXA, CVs were 0.58, 1.35, 0.42, and 1.02%, for GBAT, TTAT, GBLST and TTLST, respectively. For FWMRI, CVs were 0.80, 2.08, 0.60, and 2.43%, respectively. The CVs for FWMRI quantification of SAT and VAT were 2.11 and 2.62%, respectively.
Comparison of tissue masses
Concordance correlation coefficients, coefficients of bias, and mean differences between FWMRI and DXA are presented in Table 3. None of the 95% confidence intervals include zero, indicating statistically significant correlations between FWMRI and DXA for GBAT, TTAT, GBLST, and TTLST. The strength of the concordance correlation was highest for GBAT with ρc = 0.978 (0.923, 0.994) and Cb = 0.996 (ρr = 0.982, P < 0.0001). A Wilcoxon signed-rank test also showed no significant difference (0.1 ± 1.5 kg, P = 0.75) between FWMRI and DXA measures for GBAT. However, concordance correlations for TTAT, GBLST and TTLST showed negative biases toward FWMRI. Correlation between FWMRI and DXA for TTAT was ρc = 0.802 (0.468, 0.935) and Cb = 0.884 (ρr = 0.907, P < 0.0001), but a Wilcoxon signed-rank test revealed a significant bias of –1.7 ± 1.5 kg (–10.4 ± 9.1%, P = 0.008).
Table 3. Correlations and differences between FWMRI and DXA adipose and lean soft tissue masses
Concordance correlation of GBLST was ρc = 0.629 (0.324, 0.816) and Cb = 0.646 (ρr = 0.974, P < 0.0001), and a Wilcoxon signed-rank test showed a significant bias of –6.0 ± 2.4 kg (–12.8 ± 3.8%, P = 0.003). Concordance correlation of TTLST was the lowest of the compared tissue masses with ρc = 0.400 (0.120, 0.620) and Cb = 0.423 (ρr = 0.944, P < 0.0001). A Wilcoxon signed-rank test confirmed a significant bias of –3.2 ± 1.2 kg (–20.9 ± 5.2%, P = 0.003).
Bland–Altman plots for GBAT and TTAT ( Figure 3) show good agreement between FWMRI and DXA with the limits of agreement for GBAT from –2.8 kg to +3.1 kg and for TTAT from –4.6 kg to +1.2 kg. These ranges include zero and linear regression of the method difference against method average showed no statistically significant trend with r2 = 0.17 (P = 0.18) for GBAT and r2 = 0.00 (P = 0.92) for TTAT.
Bland–Altman plots for GBLST and TTLST ( Figure 4) indicate there is less agreement between FWMRI and DXA for lean soft tissue than for adipose tissue. Neither the limits of agreement for GBLST (–10.8 kg to –1.3 kg) or TTLST (–5.6 kg to –0.9 kg) include zero. The lean soft tissue plots reveal an increasing negative bias in the FWMRI minus DXA difference with increasing lean soft tissue mass. Linear regression of the method difference against method average confirmed that this trend was statistically significant for GBLST (r2 = 0.66, P = 0.0013) and TTLST (r2 = 0.67, P = 0.0011).
FWMRI vs. DXA differences
No anthropometric measures or derived values (BMI and BAI) were significantly correlated with the difference between FWMRI and DXA for TTAT. Waist circumference, hip circumference, and BMI were also not significantly correlated with the differences observed in GBLST or TTLST. However, as presented in Figure 5, the differences between FWMRI and DXA for GBLST and TTLST were significantly correlated with subject height (ρs = –0.811 and ρs = –0.790), weight (ρs = –0.895 and ρs = –0.832) and abdominal sagittal thickness (ρs = –0.804 and ρs = –0.601). Additionally, BAI was correlated with TTLST difference (ρs = 0.608, P = 0.004).
The key finding from this study is demonstration that the precision of the implemented contiguous 3D fat-water MR imaging technique at 3T makes it a compelling alternative to DXA for quantifying body composition in obesity primarily because this method offers the important advantage of quantifying VAT and SAT. As recent MR imaging body composition studies have reported findings from adipose tissue mass only (16, 32), the present study further extends the evidence by presenting data on lean soft tissue mass which is an important determinant of insulin sensitivity and has a critical role in the pathophysiology of many disease processes including obesity-associated sarcopenia and cachectic syndromes.
The coefficient of variation for repeated measurements using FWMRI in obese women was <1% for GBAT and GBLST. Confirming the excellent precision of the FWMRI method, the CVs for GBAT and GBLST were comparable to those of DXA (<0.6%). Additionally, FWMRI CVs for TTAT and TTLST were <2.5%, and CVs for SAT and VAT were <2.7%. Our findings demonstrate greater precision for FWMRI at 3T than those previously reported at 1.5T where GBAT CV was 2.32% and TTAT CV was 2.25% (16). We observed slightly less precision in the CVs for FWMRI trunk tissues compared to DXA where TTAT and TTLST CVs were <1.4%. We found no published reports comparing the reproducibility of FWMRI for GBLST or TTLST.
Consistent with the overall precision, we found that FWMRI measures of GBAT in these obese women correlated strongly with DXA (ρc = 0.978). However, it appeared that FWMRI underestimated TTAT. Kullberg et al. (32) have previously reported that FWMRI at 1.5T underestimated VAT when compared to CT and suggest that abdominal soft tissue motion and partial volume effects may contribute to VAT signal loss. Similar effects may explain the lower TTAT measured by FWMRI in this study. Partial volume effects are a potential confounding factor in general for the threshold adipose tissue segmentation approach applied in this study while being less of a concern for fat signal fraction integration approaches. However, a number of known confounding factors (radio frequency (RF) transmit and receive inhomogeneity, T1 relaxation, T2* relaxation, eddy currents, noise bias, multiple fat resonances) must be addressed for the signal fat fraction to accurately reflect the true proton density fat fraction. Another confounding factor arises from intramuscular adipose tissue included in the VAT volume. The automated segmentation algorithm (16) excludes bone marrow and intramuscular adipose tissue in the abdominal subvolume in the region of the spine and back. However, the exclusion of intramuscular tissue throughout the abdomen is not perfect, as is visible in some axial sections displayed in Figure 2, and is a focus of ongoing development.
Though correlations between FWMRI and DXA estimates of lean soft tissue were strong for GBLST and TTLST regions, our findings indicate a significant difference between the two methods that increased in magnitude with taller, heavier, and thicker subjects. As the women in the present study ranged in height from 149.9 to 177.8 cm, some extremity tissue was excluded from the FWMRI FOV, which was limited to a maximum of 163.2 cm. Thus, the discrepancy in GBLST is accounted for by missing the most distal portions of the head, feet, hands, and forearms (which were bent directly overhead). We recognize that tissue not captured within the FWMRI FOV is a primary source of error in the FWMRI quantification. However, future studies can address this shortcoming by matching overhead arm position for DXA and FWMRI prior to comparison and by software modifications to increase scanning coverage in the head-to-foot direction. Another limitation of the current quantification method is the use of a binary, all-or-none, classification of FWMRI voxels as either water (lean soft tissue) or fat (adipose tissue). Once confounding factors such as transmit and receive RF field inhomogeneity and T1 bias are better-compensated, more accurate estimates of fat volume (and mass) could be obtained using the fractional fat signal percentage in each FWMRI voxel.
However, incomplete coverage of the extremities does not explain the difference in TTLST observed between FWMRI and DXA because the entire trunk region was well sampled in the FWMRI FOV. The bias between FWMRI and DXA in TTLST, may be explained by overestimation of lean soft tissue by DXA. Such overestimation has been observed in several prior studies and attributed to assumptions in DXA algorithms for estimating appendicular skeletal muscle (8, 9). It must also be recognized that direct comparison of soft tissue measured by FWMRI and DXA methods is limited by the fact that FWMRI measures tissue volume, which must be converted to mass by using assumed densities for soft tissue (33). Unfortunately, comparison between FWMRI and CT was contraindicated in the present study because of the need for performing repeated measures to determine repeatability and the substantial concern with CT-associated radiation exposure.
Although MRI scanning time has been perceived to be an impediment to widespread use of MRI in the past, the FWMRI scan time in this study was comparable to that of DXA acquisition. Other faster variations of multi-echo DIXON methods different from the three-echo method employed in this study are available. For example, a two-echo method has the advantage of shorter acquisition time, but acquiring three echoes improves identification of the water and fat components, which simplifies signal separation. Acquiring more than three echoes enables additional unknowns such as T2* to be numerically fit, but the acquisition of more echoes increases scan time or requires a reduction in spatial resolution. The FWMRI scan time can be further accelerated using parallel imaging methods such as sensitivity encoding (SENSE) which requires an array of multiple receive-only RF surface coils (34). Scan time could also be reduced with a continuously moving table strategy (16) or by using a data sampling acceleration strategy combined with an advanced image reconstruction approach such as compressive sampling (35).
In the present study, the absence of surface RF coils maximized subject comfort, compatibility with larger subjects, and simplified the multiple table-station scanning approach. Importantly, the current speed of FWMRI acquisition makes it as affordable as DXA in research and clinical settings. For example, at our institution the charge for the 30-min time slot required to complete a FWMRI scan on the research 3T MR scanner is equivalent to the charge for a 30-min DXA time slot. It is also noteworthy that FWMRI in the present study was performed on a clinical scanner, which clearly increases the potential for widespread utilization of this method. Yet, we acknowledge that MRI scanning at 3T presents increased challenges related to static magnetic field (B0) and RF transmit field (B1) inhomogeneity (17) which can cause geometric distortion, signal loss and signal intensity inhomogeneity that can confound accurate fat and water separation as well as increase error in volume measurements. In cases where static magnetic field (B0) homogeneity becomes problematic, alternative shimming strategies such as slice-wise dynamic shimming can be pursued (36). Challenges related to RF transmit field (B1) inhomogeneity that may arise could be addressed by a parallel RF transmit 3T scanner (37).
In conclusion, we have demonstrated that gross body FWMRI imaging performed at 3T is rapid, precise, repeatable and potentially cost-effective. In the obese women studied, FWMRI had excellent concordance with DXA for the measurement of GBAT. Moreover, because FWMRI data is 3D and reliably divides the MR signal into separate fat and water components, this method is able to go advance the study of body composition and distinguish adipose tissue topography to quantify visceral and subcutaneous abdominal tissue depots. While recent reports on quantifying body composition using magnetic resonance spectroscopy (quantitative magnetic resonance) also show great promise with regard to sensitivity and specificity for certain applications (18, 38), such spectroscopy-based approaches provide little or no information about the spatial distribution of adipose and lean soft tissue. In addition, the chemical-shift-based separation of fat and water signals employed in the FWMRI approach described is more advanced than MRI fat quantification approaches that are based on standard relaxation contrast mechanisms, which are not amenable to automated segmentation and quantification algorithms (39). Importantly, future applications of the FWMRI method described will also allow quantification of ectopic fat (i.e., intermuscular adipose tissue and lipid burden within organs such as the liver, pancreas, and heart) (40). Finally, when compared to DXA and other non-MR methods (i.e., CT), FWMRI is prudent because it requires no exposure to ionizing radiation. Thus, it can be used safely in children and in longitudinal studies that require serial measurements to capture changes in body composition. Hence, the implementation of a 3T multi-station multigradient-echo based fat-water imaging method described in this work combined with a fully automated post-processing algorithm is currently being applied to investigate physiological and metabolic mechanisms that lead to the development of obesity and its metabolic consequences.
The authors thank Taren B. Spence, RD for study data collection, Charles DeMarcus Keil for data entry, and VUIIS MR Technologists, Robin Avison and Donna Butler, for MR scanning support. We also thank the study participants for their time and commitment. This work was supported by The Dr. Robert C. and Veronica Atkins Foundation (H.J.S.), the Vanderbilt CTSA grant 1 UL1 RR024975 from the National Center for Research Resources, National Institutes of Health (H.J.S.), the Tennessee Valley Healthcare System, NIH grants DK064857 and DK069927, and the Vanderbilt Diabetes Research and Training Center grant DK020593 (K.D.N.).