Comparison of fat quantification methods: A phantom study at 3.0T




To compare different imaging methods with single-voxel MR spectroscopy (MRS) for the quantification of fat content in phantoms at 3.0T.

Materials and Methods

Imaging and spectroscopy was performed on a GE Signa system. Eleven novel homogeneous fat-water phantoms were constructed with variation in fat content from 0% to 100%. These were imaged using three techniques and compared with single-voxel non-water-suppressed MRS. Pixel-by-pixel maps of fat fraction were produced and mean values compared to MRS-determined measurements. Preliminary in vivo examinations were subsequently performed in the breast and spine to compare the best imaging technique with MRS.


All imaging methods significantly correlated with MRS (P < 0.001): IDEAL (r2 = 0.985), IOP (r2 = 0.888), WS (r2 = 0.939), and FS (r2 = 0.938). In addition, IDEAL provided artifact-free maps of fat fraction with superior uniformity. In vivo results using IDEAL produced values that were between 9% to 16% of the corresponding MRS values.


This work demonstrates that imaging may be utilized as a high-resolution alternative to MRS for the quantification of fat content. In the future we intend to replace MRS with IDEAL in our clinical studies involving fat measurement. J. Magn. Reson. Imaging 2007. © 2007 Wiley-Liss, Inc.

FAT SUPPRESSION is a common MRI technique routinely used to suppress the signal from adipose tissue. It may be used to reduce chemical shift artifacts, increase conspicuity of contrast-enhanced tumors, or improve tissue characterization. Techniques for suppressing fat have more recently been utilized as a means of providing a measurement of fat content. Several groups have demonstrated the usefulness of fat quantitation in bone marrow (1), breast (2, 3), muscle (4), brain (5), and liver (6, 7). The majority of these studies have used the gold-standard of MR spectroscopy (MRS) to separate fat and water resonances and subsequently provide an accurate fat fraction measurement. However, the limited spatial resolution of this technique has prompted some groups to investigate imaging as an alternative method. Indeed, this study was motivated by our own desire to quantify fat fraction in our clinical work investigating bone marrow composition. To date, there has been little work comparing imaging and MRS for fat measurement.

Fat suppression, and therefore fat measurement, relies on discriminating between the signal from lipid and water protons by exploiting differences in either frequency (phase) or relaxation properties (8). Two techniques that rely on the former differences are chemical shift saturation and opposed-phase imaging (9, 10). Chemical shift saturation is a frequency-selective method that is recommended for suppression of signal from large amounts of fat and by adjusting the appropriate center frequency may alternatively be used to suppress water. This technique is sensitive to magnetic field nonuniformity and performs better with increased chemical shift dispersion at high magnetic field. Opposed-phase imaging utilizes the phase shifts between fat and water signals that accrue at specific field-dependent echo times. Both these techniques have recently been used qualitatively to demonstrate correlations between signal intensity and pathologic fat content in the liver (11). Opposed-phase imaging has additionally been used in the liver to correlate with MRS-determined values of fat fraction (6) or verify known concentrations in phantoms (7).

This work investigates three imaging methods: The first two, chemical shift saturation and opposed-phase imaging, are readily available on clinical systems. The third technique, IDEAL (iterative decomposition of water and fat with echo asymmetry and least-squares estimation), is the latest implementation of opposed-phase imaging based on a modified three-point Dixon technique (12). Both of these use three echoes to obtain separate water and fat images and additionally correct for field inhomogeneities. IDEAL further extends this method by iteratively calculating the field map and also uses asymmetric phase shifts to maximize signal-to-noise performance (13). It has been shown to provide qualitative improvements in fat-water separation compared to other techniques (14, 15). More recently, Reeder et al (16) obtained a fat fraction measurement using IDEAL in the liver.

We have developed 11 homogeneous fat-water phantoms covering a fat content ranging from 0% to 100%. Previous studies requiring such phantoms have often compromised by using separate fat and water compartments (17) or utilizing partial volume effects at the interface to introduce varying fat fractions (7). Bovine liver tissue and oil homogenates have also been used to obtain physiologically ‘real’ samples (18). The phantoms described here have enabled us to reliably compare three imaging techniques with measurements determined by MRS, taken to be the current gold standard for fat fraction measurement in vivo. We produced pixel-by-pixel maps of fat fraction for each imaging method with the ultimate aim of adopting one of these for our clinical studies at 3.0T.

In vivo examples using the best imaging technique have been included at two distinct anatomical sites together with MRS acquisitions to illustrate its usefulness in clinical situations.


Phantoms of varying fat-water content were constructed. In order to create homogenous oil-in-water emulsions that would remain stable for the duration of scanning, 15 mmol of the anionic surfactant (emulsifying agent) sodium dodecyl sulfate (Sigma-Aldrich, St. Louis, MO) was added to a liter of deionized and demineralized water and 5 g of carrageenan was dissolved in the solution using a magnetic stirrer hotplate heated to 50°C. Eleven (100 mL) phantoms were constructed with the following compositions; soya oil (fat only), original unadulterated solution (water only), and nine emulsions containing increasing amounts of soya oil from 10 mL to 90 mL. These were emulsified using a homogenizer (Ultra-Turrax T25) for 2 minutes with a 1-minute rest interval to reduce the presence of microbubbles. The gels remained stable; creaming did not occur during the time of imaging and coalescence did not occur at all. Imaging in the sagittal plane was used to verify that the fat and water components had not separated during the time of the examination.

MR imaging and spectroscopy was performed on a 3.0T whole-body GE Signa system (HDx platform; Milwaukee, WI) with all 11 phantoms positioned centrally and tightly packed in a commercially available single-channel RF headcoil. Initial localizing scans were followed by the spectroscopy acquisitions. These data were acquired using the PRESS sequence with TE/TR = 32/2000 msec. Nonwater-suppressed spectra were obtained by setting water suppression voltages and flip angles to zero. A single voxel was prescribed from a region at the center of each phantom, well away from the side walls, equal to 3.7 cm3. After local shimming and gradient adjustment, a total of 32 signals were collected with a spectral bandwidth of 5000 Hz and 4096 data points.

Data were analyzed offline using the SAGE package (GE Medical Systems) and processing included zero-filling, spectral apodization (using a Gaussian function with a 2.5 Hz linewidth), first-order automatic phasing, and DC baseline correction. Peak areas obtained by integration of the water resonance and the main lipid peak (at 1.3 ppm) were recorded and the percentage fat fraction (FF) was calculated from the fat-to-water peak ratios (FWR) according to 100 × FWR/(FWR +1).

Imaging was also performed in the coronal plane through the full extent of the phantoms using a 3-mm slice thickness and 1-mm gap, a 256 × 256 image matrix, and 25 cm field-of-view (FOV). In total, three imaging techniques were investigated. First, chemical shift saturation was examined for both water-selective suppression (WS) and fat suppression (FS). This was done by acquiring a T2-weighted FSE sequence (TE/TR = 32/2000 ms), first with no suppression and then repeated for fat and subsequently water suppression pulses turned on with an acquisition time (for both) of 1 minute 54 seconds. Second, standard in- and out-of-phase imaging (IOP) was acquired together with a methodology previously used in the liver (7). This consisted of two fast gradient echo sequences (TR = 150 msec) with TE either in (2.1 msec) or out (3.2 msec) of phase and flip angles of 20° and 70° to identify the dominant signal component (ie, to resolve the ambiguous signal given by either water minus fat and fat minus water inherent with magnitude reconstructed images). The total time for four acquisitions was 5 minutes 16 seconds. The third method used the T2-weighted IDEAL-FSE sequence (13) (TE/TR = 32/2000 ms), which produces separate water-only and fat-only images as part of the protocol, normally used in a qualitative manner and acquired in 3 minutes.

All images were processed offline using in-house developed software (Matlab; MathWorks, Natick, MA). For the WS and FS methods this involved calculating the ratio of the suppressed image to the unsuppressed image (taking note of the fact that FS would give water fraction and so FF = 100-WF). For the IOP method, FF was calculated using both sets of flip angle images and the final value determined from a comparison of the two results, recording either FF or 100-FF (7). In the case of IDEAL, FF was calculated from the ratio of the separate water-only and fat-only images (FWR) using the equation 100 × FWR/(FWR +1).

Pixel-by-pixel color-scale maps of fat fraction were produced for each imaging technique and measurements of the mean and standard deviation (SD) pixel values were then taken from regions of interest (ROIs) drawn within each phantom. For accuracy measurements an ROI was used that matched the size and location of the MRS voxels. The imaging and MRS values were correlated with a Pearson correlation coefficient using SPSS 13.0 (Chicago, IL). Bland–Altman plots were also constructed by plotting the mean of the MRS and imaging values against their differences for each technique to demonstrate the levels of agreement.

Uniformity was assessed qualitatively and a comparison between the techniques was made by calculating 100 × SD/mean in the 30% and 70% fat phantoms, using a larger ROI (≈80% of the cross-sectional area avoiding flask edges), and reporting the mean of the two values.

MRS and IDEAL was also performed in the lumbar spine of a normal subject, a male age 34 years, and in a primary breast tumor in a patient age 64 years using the same protocols as outlined above. In the case of the spine examination, MRS was acquired within three vertebral bodies (L3, L4, L5). An ROI was drawn in the fat fraction maps corresponding to the prescribed spectroscopic voxel position and an average taken over the appropriate number of slices consistent with the voxel thickness.


Figure 1 shows example spectra acquired in the phantoms. Figure 1a is the spectrum corresponding to the 50:50 (fat:water) phantom; Fig. 1b is the spectrum from the 90:10 phantom. These two spectra have been scaled with respect to one another and overlaid in Fig. 1c to illustrate the assignment of the water (4.7 ppm) and the main lipid peak (1.3 ppm) corresponding to protons from the methylene chain. In addition to the resonance at 1.3 ppm, several other resonances from lipid protons are present that are commonly seen in vivo (3, 19, 20). These are labeled in Fig. 1b for terminal methyl protons 0.9 ppm, carboxyl group methylene protons 2.06 ppm, and olefinic protons 5.35 ppm. There are a further three lipid resonances (indicated by arrows in Fig. 1b) of low amplitude seen in this oil that are not readily observed in vivo. Figure 2 shows a plot of fat fraction (by volume) for each phantom against the peak area obtained from the 1.3 ppm resonance (r2 = 0.993).

Figure 1.

Example spectra acquired in (a) 50:50 (fat:water) phantom and (b) 90:10 phantom. Arrows indicate three additional lipid peaks that are not usually seen in vivo. c: The two spectra are shown scaled and overlaid.

Figure 2.

A plot of percentage fat fraction (by volume) for each phantom against the peak area of the lipid peak at 1.3 ppm as measured by MRS (r2 = 0.993).

Pixel-by-pixel maps of FF, with an in-plane resolution equal to 1 × 1 mm, are shown in Fig. 3a–d for each of the following imaging techniques investigated: (a) chemical shift selective fat saturation (FS), (b) water saturation (WS), (c) in- and out-of-phase imaging (IOP), and (d) IDEAL. The maps are taken from the central slice of the set of phantoms, with the containers arranged in order of increasing fat content (bottom to top, left to right). Each image reproduces the increase in fat content as can be seen by the color scale. However, Fig. 3a–c demonstrate more pixel intensity variations than Fig. 3d. Air bubbles present in the original images for the 80:20 phantom have affected the final FF map in the case of Fig. 3c.

Figure 3.

Pixel-by-pixel parameter maps of fat fraction (FF) determined from (a) chemical shift selective fat saturation (FS), (b) water saturation (WS), (c) in- and out-of-phase imaging (IOP), and (d) IDEAL.

Figure 4 plots FF values determined from the ROIs drawn in the pixel-by-pixel maps for each imaging method compared with the corresponding MRS values. In each case there was a significant correlation (P < 0.001) with MRS values. Results obtained with the IDEAL sequence correlated best. The regression lines produced the following results: IDEAL: 0.79 + 7.24 (r2 = 0.985), IOP: 0.72 + 7.77 (r2 = 0.888), WS: 0.68 + 9.89 (r2 = 0.939), and FS: 0.78 + 12.9 (r2 = 0.938). Bland–Altman plots are shown in Fig. 5. These revealed mean differences equal to −3.7% (IDEAL), −7.4% (IOP), −7.1% (WS), and +1.2% (FS). Maximum differences were 16.1% (IDEAL), 26.3% (IOP), 30.7% (WS), and 19.3% (FS). Of note, there was a general underestimate for all imaging sequences at high fat fraction values (from 80%). Excluding the values from the 90% and 100% fat phantoms the mean differences were equal to −0.9% (IDEAL), −4.0% (IOP), −2.2% (WS), and +5.2% (FS).

Figure 4.

Plots of fat fraction (FF) determined by each imaging method versus MRS determined values. A line of identity is shown; however, regression lines have been omitted for clarity.

Figure 5.

Bland–Altman plots for each method showing the level of agreement between the MRS values and the various imaging methods. Symbols are the same as in Fig. 4.

Qualitatively, the IDEAL images displayed a more uniform distribution of pixel values in each phantom compared to the other techniques. The mean value of uniformity for the 30% and 70% phantoms was as follows: IOP = 93.0%, FS = 95.2%, WS = 96.0%, and IDEAL = 98.6%.

Figure 6 illustrates fat fraction maps acquired using the IDEAL technique in vivo in breast (left) and spine (right) examinations. In both cases fat content was also estimated using MRS. The spectroscopic voxel location for the breast tumor is highlighted in white. The fat fraction in this region was found to be 11.3% and 9.5% for the MRS and IDEAL measurements, respectively. In the spine example MRS was acquired centrally in the vertebral bodies of L3, L4, and L5. This produced FF values of 57.2%, 58.6%, and 59.3% with corresponding imaging values of 62.5%, 67.3%, and 67.1%, respectively.

Figure 6.

Fat fraction maps produced from the IDEAL images acquired in the breast (left) and lumbar spine (right). In the breast image the location of the spectroscopic voxel location has been outlined in white. The same color scale (0% to 80% FF) has been used in both cases.


In this work we compared three imaging techniques to the gold standard of MRS for the quantification of fat in 11 phantoms of varying fat proportion. This formulation permits the construction of a homogenous and stable fat-water mix that is more representative of in vivo composition. The various fat resonances (both in terms of relative amplitudes and chemical shifts) are equivalent to in vivo spectra as previously described in MRS studies of breast (3, 19) and bone marrow (20), and assignments were made accordingly.

A more complete analysis would involve measuring the T1 and T2 relaxation times of the phantoms so that the contrast weightings could be normalized. While the sequence timings were the same for the spectroscopy and three of the imaging techniques, the IOP method introduced T1-weighting. Furthermore, differences in T2* of the water and soya oil may introduce further errors in the IOP method due to differences in decay rate between these components. However, the aim of this study was to create fat phantoms of widely varying proportion so that MRS and imaging methodologies could be tested as they would be in vivo. More specifically, we wanted a range of phantoms that had a consistent linear relationship between 1.3 ppm lipid resonance (Fig. 2). In this respect we were less interested in either reproducing absolute concentrations from MRS or correcting for differences in image weightings.

A further potential for deviation is B0 inhomogeneities. We attempted to minimize this by imaging within the smallest FOV possible and using small containers that where still large enough to adequately contain single voxels of typical size used in vivo. Susceptibility artifacts will be present around the edges of the flasks and can be expected to be worse for gradient echoes. The magnitude of this effect could be examined by imaging the flasks separately.

MRS may be a time-intensive method both in terms of acquisition and significant input and postprocessing. Although the single-voxel data here were collected in 64 seconds, in vivo studies may require either more than one voxel or, more likely, many more signal averages to be collected. This would make the total acquisition time similar to the imaging described here without the benefits of imaging resolution. In either case, it is important to be mindful of the potential for patient motion when resolution gain results in increases in scan time. It should be noted, however, that MRS has the advantage of demonstrating the presence of all lipid resonances that may be important in certain clinical investigations (21).

All three imaging methods described here correlated significantly with MRS (IDEAL > WS > FS > IOP). There was a general trend toward underestimation at higher fat fractions. This is thought to be caused by the presence of the fat peak closest to the water resonance (5.35 ppm) which may erroneously be assigned as ‘water’ thereby causing an underestimation. At the opposite end, an examination of the intercepts of the regression lines is a particularly useful comparison, as it is very important clinically to distinguish between the presence and absence of fat. All imaging methods overestimated this from 7.2% to 12.9%, with IDEAL demonstrating the smallest error.

An attempt to compare uniformity has been made by calculating the percentage variation of pixel values within the 30% and 70% phantoms. These phantoms were chosen so that in each technique there was significant signal from either component, irrespective of how the final maps were calculated, so that the measure was less dependent on signal-to-noise.

The WS and FS methods are easy to implement and were the fastest of the three techniques in terms of total acquisition time for the two required sequences. However, the fat fraction maps demonstrated regions of variation within individual phantoms. This is due to artifacts in the suppressed images caused by B0 inhomogeneities, presumably as a result of susceptibility artifacts. Chemical selective saturation is readily available on all clinical systems, usually as fat suppression (FS). The water-equivalent suppression (WS) is a simple extension of this and was also available on our high-field system.

The IOP sequence is also readily available on most systems, although the exact methodology described here requires the acquisition of four series of images, making it the longest total acquisition time, with a slight increase in the degree of postprocessing. This technique did not demonstrate the artifacts seen with the frequency selective method, although the methodology used here, whereby the dominant fat or water signal is determined from the T1-weighting is, intuitively, prone to error where the fat and water content is similar. This can be seen with the large discrepancy around 50% fat fraction and indicates that the 70° flip angle sequence is not determining the dominant component correctly and is possibly exacerbated by calculating on a pixel-by-pixel basis.

Maps produced using the IDEAL sequence show this method to be superior in terms of providing uniform images that are robust with respect to artifacts in a phantom set-up which created problems (although not intentionally) for the other sequences. This may be pertinent in problematic areas of anatomy. It also has the advantage of requiring only one acquisition of around 3 minutes, which is marginally longer than the FS method. A more recent study has proposed further improvements to the accuracy of this sequence by reducing noise and T1-effects (22).

Our preliminary in vivo examinations have shown that FF maps produced with the IDEAL sequence are artifact-free and can be used to quantify fat content at high spatial resolution. Reeder et al (16) recently used the sequence in the liver and demonstrated values of 20% fat fraction, although there was no verification with MRS. We have gone on to compare this sequence against the gold standard of MRS in two quite dissimilar areas of anatomy, breast and spine. This has facilitated the comparison over a wider range of fat content. In both cases the IDEAL measurement was within 16% of the MRS determined value, although the values were closest in the breast compared to the spine. The larger discrepancy in spine may be due to T2* effects, although preliminary in vivo work suggests it is still the preferred technique in this region of anatomy (20).

In conclusion, this work provides evidence that high-resolution imaging techniques offer an alternative to MRS for fat quantification. This may prove crucial in a wide range of applications where the distribution of fat content is sought.


The authors thank Mr. Timothy Dunstan (Research Technician, Surfactant and Colloid Group, Department of Chemistry, University of Hull) for advice and help in the preparation of the phantoms.