CME JOURNAL OF MAGNETIC RESONANCE IMAGING 40:1437–1444 (2014)

Original Research

Whole-Body MRI-Based Fat Quantification: A Comparison to Air Displacement Plethysmography Ute A. Ludwig, PhD,1* Florian Klausmann, Dipl.Phys.,1 Sandra Baumann, PhD,1 €vener, PhD,1,2,3 Daniel Ko €nig, MD,4 Matthias Honal, PhD,1 Jan-Bernd Ho 4 1 €chert, PhD Peter Deibert, MD, and Martin Bu Purpose: To demonstrate the feasibility of an algorithm for MRI whole-body quantification of internal and subcutaneous fat and quantitative comparison of total adipose tissue to air displacement plethysmography (ADP). Materials and Methods: For comparison with ADP, whole-body MR data of 11 volunteers were obtained using a continuously moving table Dixon sequence. Resulting fat images were corrected for B1 related intensity inhomogeneities before fat segmentation. Results: The performed MR measurements of the whole body provided a direct comparison to ADP measurements. The segmentation of subcutaneous and internal fat in the abdomen worked reliably with an accuracy of 98%. Depending on the underlying model for fat quantification, the resultant MR fat masses represent an upper and a lower limit for the true fat masses. In comparison to ADP, the results were in good agreement with r 0.97, P < 0.0001. Conclusion: Whole-body fat quantities derived noninvasively by using a continuously moving table Dixon acquisition were directly compared with ADP. The accuracy of the method and the high reproducibility of results indicate its potential for clinical applications. Key Words: Dixon imaging; fat quantification; subcutaneous adipose tissue; visceral adipose tissue; air-displacement plethysmography; continuously moving table MRI J. Magn. Reson. Imaging 2014;40:1437–1444. C 2014 Wiley Periodicals, Inc. V

IN RECENT YEARS, obesity has become a serious health issue primarily in industrialized countries. The total amount of adipose tissue (AT) in the body is 1 Department of Radiology - Medical Physics, University Medical Center Freiburg, Freiburg, Germany. 2 German Consortium for Cancer Research (DKTK), Heidelberg, Germany. 3 German Cancer Research Center (DKFZ), Heidelberg, Germany. 4 Department for Rehabilitation, Prevention and Sports Medicine, University Medical Center Freiburg, Freiburg, Germany. *Address reprint requests to: U.A.L., University Medical Center Freiburg, Department of Radiology - Medical Physics, Breisacher Strabe 60a, 79106 Freiburg, Germany. E-mail: [email protected] Received March 12, 2013; Accepted October 14, 2013. DOI 10.1002/jmri.24509 View this article online at wileyonlinelibrary.com. C 2014 Wiley Periodicals, Inc. V

related to a multitude of diseases like diabetes mellitus, cancer, depression, as well as cardiovascular diseases and stroke (1). In addition to the total amount, the fat distribution in the body has also recently been recognized as an important factor in the pathogenesis of several diseases (2). It was shown that increased health risks are highly correlated to the amount of internal and visceral fat in the abdominal region of the body (3). For this reason, the quantification of internal and visceral fat is of clinical interest and its fraction of the whole-body fat can provide further information for the treatment of patients. MRI is capable of spatially resolved imaging of fat distribution in humans noninvasively and in vivo, thus allowing the quantification of different adipose tissue compartments in the body. For these tasks, several approaches have been introduced, based on T1-weighted MRI (4–8) or Dixon-related fat–water separation methods (9–11). The latter have the advantage that no further tissue analysis is required (12). The whole-body fat volume and distribution may be obtained using continuously moving table (CMT) MRI (9,13,14). In contrast to non–whole-body MRI, this approach enables comparison with other standard techniques for whole-body fat determination, like Bio Impedance Analysis (BIA) (15), dual energy X-ray absorptiometry (DEXA) (6) or air displacement plethysmography (ADP) (16). These methods have been discussed in several reviews (17–20); in particular, some studies have applied both MRI and ADP (21–23), and good correlation was found between the two modalities. However, to our knowledge, there is no direct comparison of whole-body fat mass measured with MRI and ADP. In this work, we introduce our approach to automatically quantify the whole-body, subcutaneous and internal fat volume and mass, using CMT, two-point Dixon MRI. The method presented includes a simple approach to compensate for B1-inhomogeneities and partial volume effects (24), and provides the subcutaneous and internal fat volume and mass using an automated active contour segmentation procedure. Our purpose was to demonstrate the feasibility of this algorithm for whole-body quantification of internal

1437

1438

and subcutaneous fat and its accuracy compared with reference standards. MATERIALS AND METHODS MRI Data Acquisition Whole-body MRI measurements were performed on 11 healthy volunteers (10 men, 1 woman, average age 30 6 10 years, average body mass index (BMI) 29.5 6 3.8 kg/m2) on a 1.5 Tesla (T) short bore, whole-body MRI system with an inner diameter of 70 cm (system A, Magnetom Espree, Siemens Healthcare, Erlangen, Germany). Written and informed consent in compliance with the institutional review board requirements was obtained before the measurements. Data were acquired with an axial two-dimensional (2D) gradient echo multislice sequence with two bipolar readout gradients. The patient table was moved continuously during the acquisition while adapting the frequency of the radiofrequency (RF) excitation pulses to compensate for the table motion (25). The signal was received using the manufacturer’s volume head and neck coil, three surface phased-array coils to cover the abdomen and a phased-array leg coil in addition to the spine coil array integrated in the patient table. During the continuous data acquisition of the extended field of view (FOV), the receiver coils were dynamically switched on and off according to the body region currently being measured. This way, the noise from unused coils was excluded and the overall signal to noise ratio (SNR) improved. Geometric distortions in this multislice experiment (12–18 slices per slice package, no slice gap) were reduced using the sliding multislice technique (26). Sequence parameters were as follows: echo time TE1 ¼ 2.38 ms (opposed-phase image), TE2 ¼ 4.76 ms  (in-phase image), TR ¼ 93 ms, flip angle 70 , bandwidth 450 Hz/Px, matrix size 320  259 and slice thickness 5 mm. The FOV in the anterior–posterior direction was adapted to the object size resulting in a FOV ¼ 450  (310–394) mm2. This leads to a voxel size of 1.4  (1.2–1.5)  5 mm3. For the acceleration of data acquisition, partial Fourier imaging (7/8) (27) as well as generalized auto calibrating partially parallel acquisition (GRAPPA, acceleration factor 2) (28) was applied. To cover the necessary longitudinal FOV, 324–386 slices (mean value 359 slices depending on the height of the volunteer) were acquired resulting in a total pure acquisition time of the order of 12 min. In the abdominal region the CMT acquisition was interrupted several times as data were acquired during breathholding (exhaled position). Each breathholding phase lasted 15 s. From the distal abdomen on, the CMT acquisition was continued without further breathing interruptions. For each slice position, one fat and one water image were calculated by a 2-pointDixon image reconstruction method automatically with the software provided by the manufacturer. Due to increasing field inhomogeneities outside the isocenter, one arm was acquired separately, after imaging the whole body. Hence, the volunteers were

Ludwig et al.

asked to move as far as possible to one side of the patient table to position one arm beside the body in an area with a homogeneous B0-field. For patient comfort, only one arm was acquired, and the second was assumed to be the same. Whole-body and arm MRI were joined at the lowest part of the axilla. To test the variability of the method, the inner fat volume (IFV) of the abdomen of one volunteer (age 25, BMI 30.0 kg/m2) was measured 3  5 times: 5 consecutive scans were acquired on 2 consecutive days each without moving the volunteer at the wide-bore system A (which was also used for the other volunteers), and an additional five examinations were conducted on system B with a bore diameter of 60 cm one week later (1.5T, Magnetom Avanto, Siemens Medical Solutions, Erlangen, Germany). Image Segmentation Segmentation of subcutaneous and internal adipose tissue, including visceral AT, muscular fat and bone marrow (following the classification by Shen et al) (29) was performed for the abdominal region ranging from the highest cranial extension of the liver (exhaled position) to the first cranial slice displaying the femoral heads (Fig. 1a). Similar to other published reports (4,7,12,30–33), subcutaneous and internal AT were not further subdivided. For image segmentation, active contour snakes were used. A detailed description of those parameterized dynamic curves was published by Kass et al (34). The automatic segmentation of the selected volume was performed slice by slice using two masks and three snakes (v1-v3). The masks were obtained from the axial fat images provided by the Dixon reconstruction. One binary mask of each fat image was produced by means of thresholding at 8% of the maximum signal intensity, which is referred to as “fat mask” in the following sections (Fig. 1d). A second mask was created in a similar manner but with an increased threshold of 10%, and further processed with a morphological opening operation using a disc shaped structuring element of two pixel radius (35). This procedure erases smaller objects and leads to a “segmentation mask” (Fig. 1b) with smoother object contours and better defined gaps between different fat compartments. Snake segmentation as described in Figure 1 leads to a subcutaneous fat region defined by the area between snakes v1 and v3. The internal region was defined as the area enclosed by v3 (Fig. 1e). To evaluate the robustness of the automatic approach, the MRI data were re-segmented manually, using the automatic segmentation as a starting point. Fat quantification was performed for the automatic as well as for the manual segmentation. Postprocessing Simply computing the number of voxels in segmented regions does not quantify the fat content of these compartments accurately (36). Partial volume effects arising from voxels that do contain only some fat,

Whole-Body Fat Quantification: MRI Versus ADP

1439

Figure 1. a: Coronal MR fat image provided by a Dixon reconstruction based on data of a CMT acquisition. The volume for the segmentation of subcutaneous and internal fat in the abdomen is marked by the orange lines. b: A snake v1 was led to the outer border of the subcutaneous fat in a segmentation mask based on an axial fat image of the segmentation volume. c: Next, all mask pixels outside v1 were set to one and a second snake v2 was led to the inner border of the subcutaneous fat. d: A fine tuning of v2 was performed on a binary mask of the fat image resulting in v3. e: Original fat image with the final result of the segmentation.

especially in the inner abdomen, lead to an overestimation of the fat content in that case. To solve this issue, each voxel was weighted according to its signal intensity in the fat image (24). Before this step, signal intensities were corrected for inhomogeneities arising from the varying spatial sensitivities of the receiver coils and the RF transmitter field (10).

Voxel Weighting According to Their Fat Content To reduce partial volume effects during quantification, the signal intensities of the inhomogeneity-corrected data were weighted using a reference value Iref, resulting in data where a signal intensity of 1 represents maximum fat signal, and 0 no fat signal. Iref was derived slice-wise from the fat-only voxels in the

Inhomogeneity Correction The signal inhomogeneity correction was performed as illustrated in Figure 2 and described in the following: Voxels with high fat content were identified using the in-phase and the fat image (Fig. 2a). The background noise was eliminated in the fat image using a 5% threshold. This noise-corrected fat image was divided by the in-phase image to obtain an image with signal fractions of the fat voxels. The signal fractions do not represent volume fractions due to different proton densities and relaxation times of fat and water (37) but nearly pure fat voxels (referred to as fat-only voxels) can be found by searching for signal fractions exceeding a value of 0.95. The detected fat-only voxels do not exhibit the same signal intensity in the fat image, but rather reflect the intensity inhomogeneity profile. To correct for those inhomogeneities in the whole fat image, this profile has to be interpolated to all fat voxels (Fig. 2b). For that purpose, a discrete convolution with a 2D-Gaussian-kernel (38) with a size of 51  51 pixels and a standard deviation of 5 pixels was used. All undefined values in the original profile were set to zero. The occurring normalization errors were compensated for using a second convolution on a binary mask of the original profile. The final profile was derived by dividing the first interpolation result by the second one. In a last step, the noise corrected fat image was divided by the final profile to get the intensity corrected final fat image.

Figure 2. Flowchart with the image processing steps for the correction of B1 related intensity inhomogeneities in a fat image. a: Identification of nearly pure fat voxels (fat-only voxels) representing the inhomogeneity profile in the image. b: Interpolation of the profile to all fat voxels and correction of the fat image.

1440

Ludwig et al.

corresponding final fat image. These intensities should be the same for all fat-only voxels. However, there are still minor variations due to image noise and smoothing effects of the convolution. Iref was not chosen to be the maximum intensity across all voxels in a slice, but a lower value which is represented by the mean intensity of all voxels of 90–95% signal intensity. Fat Volume Quantification Vfat Data corrected for field inhomogeneities were slicewise divided by Iref and limited to be less than or equal to 1. The final fat volume Vfat was then computed as the sum of the volumes of all voxels, weighted by their intensity (Eq. [2]). Vfat ¼

XN n¼1

vn ¼

vn  Vvowel

[1]

Icf atðnÞ ; Iref

where Vvoxel is the volume of a voxel, N the total number of voxels, and Ic_fat(n) the signal intensity of voxel n in the inhomogeneity-corrected data. This choice of Iref, however, leads to a slight overestimation of the fat volume because small nonfatty contents are ignored. By summing up the whole-body, subcutaneous, and internal abdominal voxels, the corresponding volumes are obtained. Fat Mass Quantification For comparison to ADP, the fat mass Mfat is required. The mass is obtained by multiplying a fat volume by the density of AT. However, it is important to realize that whole-body AT density varies from 0.925 to 0.970 kg/l (39). These variations are based on individual differences in the composition of AT where most of the body’s fat is stored and are generally unknown. In addition to fat, AT consists of water, proteins, and minerals at variable fractions (39,40). We use two models to estimate the fat mass: Model 1. For this relatively simple model, the adipose tissue density was assumed to correspond to the known adipose lipid density in humans, with r fat ¼ 0.905 kg/l (39). Thus, the fat mass was calculated by multiplying the fat volume Vfat with the density r fat (Eq. [2]) Mfat ¼ Vfat  r fat :

[2]

This measure is an upper limit for the fat mass. Model 2. For this model, only the voxels (in the weighted and inhomogeneity-corrected data) with an intensity of one (Eq. [3]) were used to obtain a volume Ufat, i.e., this volume contains only voxels fully filled with fat and thus is smaller than Vfat (Eq. [3]). XM Ufat ¼ v  Vvowel [3] m¼1 m where M is the number of voxels where vm ¼ 1. Furthermore, a more realistic estimation of the adipose lipid density was attempted. Following the approach by Martin et al (39), the individual adipose tissue density r fat depends on the tissue’s fat fraction, F. Both qfat and F were calculated iteratively as

follows: (i) Starting with an estimation of r fat ¼ 0.94 kg/l and the total body volume, the body’s fat mass was calculated and divided by the body’s weight to obtain the body’s fat fraction, A. (ii) As laid out by Martin, A allows obtaining F which leads to an updated estimation of r fat . The process was repeated three times, which led to stable values for qfat and F. The total mass of fat Mfat was calculated from the volume Ufat, the fat fraction F, and the density for pure human adipose lipids r fat (Eq. [4]). Mfat ¼ Ufat  F  rfat :

[4]

This measure is a lower limit for the fat mass, because only voxels with a corrected intensity of 1 are taken into account. Image segmentation and fat quantification were performed on a personal computer (AMD Athlon Dual Core 2.2 GHz processor and 1.93 GB RAM, Microsoft Windows XP Professional operating system). The algorithms were implemented in MATLAB (The MathWorks, Inc., Natick, MA). ADP Measurements Before the ADP measurements, total body weight and height of the volunteers were determined with a scale (Life Measurement Inc., Chicago, IL, accuracy < 0.01 kg) and a caliper attached to the examination room. ADP fat measurements were performed for 11 volunteers using a BodPod (Life Measurement Inc.). The ADP principles have been described in detail (16). Briefly, the body volume and weight are measured to calculate the whole-body density which leads to the whole-body fat mass according to the relation which is based on a two compartment model of the human body (41). To improve measurement accuracy, the body volume has to be corrected for the residual lung volume (20). While more complex and more accurate methods are available, ADP is a routine method used in many clinics, including ours, to determine whole-body fat mass. Thus, in awareness of its shortcomings, we chose ADP for this study. Statistical Analysis Automated and Manually Refined Segmentation The IFV was evaluated on 66 manually selected slices by (i) automated and (ii) manually refined segmentation by one observer (3 years of experience, F.K.). The mean values and standard deviations were calculated. Interobserver Variability of Segmentation The interobserver (N ¼ 4) variability of IFV was evaluated using five randomly selected datasets (5 male, average age 35 6 15 years, average BMI 26.6 6 2.0 kg/m2). All observers were blinded to the results obtained by the others. The results of the primary observer were compared with those of the other observers (3–12 years of experience). Graphical assessment against a line of perfect agreement was used to visualize the interobserver variability (Fig. 5).

Whole-Body Fat Quantification: MRI Versus ADP

1441

The mean difference and maximum deviation as well as the coefficients of variation (cv ¼ standard deviation [SD] / mean) of IFV were calculated. Reproducibility of MRI The reproducibility of the MR results was assessed by calculating the mean intraday, interday, and intersystem cv of IFV for the automatic as well as for the manually refined segmentation. Comparison ADP and MRI Both ADP and MRI measurements were performed within 48 h except for one volunteer, where the ADP examination took place 6 days after the MRI examination. For a quantitative comparison of ADP and both MRI models, a correlation analysis was performed using Pearson’s test. A P value of < 0.0001 was considered statistically significant.

Figure 3. Automatic whole-body MRI fat segmentation and quantification for one female (no. 2) and 10 male volunteers, sorted with increasing BMI. The abdominal subcutaneous and the internal fat volume, as well as the residual nonabdominal body fat volume are marked separately. Note that the composition of the body fat may vary among volunteers with similar total fat volume.

RESULTS Comparison ADP and MRI Fat volumes were quantified using the proposed MR imaging procedure and subsequent postprocessing of the obtained fat and in-phase images. The total examination time spent for one volunteer study took approximately 45 min with the workflow presented in the Methods section. More than half that time was spent for the positioning of the volunteer and the receiver coils as well as for the additional arm measurements. Postprocessing time for segmentation and quantification in MATLAB was 3.8 s per slice. With the presented imaging strategy, breathing motion artifacts in the abdominal region were suppressed, while the total imaging time was only marginally prolonged. Subsequent image segmentation

and fat quantification provided fat volumes of the internal and subcutaneous compartments in the abdomen as well as the total fat volume in the human body. These quantities are presented in Figure 3. Volunteers with almost equal total fat volume may have a different fat distribution in their body, i.e., different fractions of subcutaneous and internal fat, e.g., volunteers 2 and 4, or 4 and 7. On average, the automatic segmentation led to a slight overestimation of (0.6 6 1.3)% for subcutaneous fat and to an underestimation of (0.2 6 1.1)% for the internal fat compared with the manually corrected segmentation results. The whole-body fat masses were derived using the two different models described in the methods and compared with ADP. As illustrated in Figure 4a, the

Figure 4. Comparison of the absolute fat masses of 11 volunteers obtained with MRI and ADP. The MRI fat masses were determined with two different models representing an upper and lower limit for the actual fat masses. Data of the only female subject is marked with a w. a: Absolute fat masses (dashed line: identity). b: Deviation of the MRI derived fat masses in percentage compared with the ADP derived fat masses.

1442

Ludwig et al.

Table 1 Variability of IFV of One Volunteer Measured on 3 Days and Two Systems*

Automatic Refined by one observer

Mean intraday cv

Interday cv

Mean interday, intersystem cv

0.73 % 0.77 %

1.17 % 1.23 %

0.85 / 0.78% 1.03 / 0.81%

*Intraday coefficients of variation (cv) were calculated from five IFV values acquired on 3 days each and averaged; interday cv was calculated from 10 IFV values obtained on two consecutive days (five each); intersystem, interday cv was calculated by using either of the intra-day datasets (each N ¼ 5 on system A) and five additional IFV values from system B.

values obtained by model 1 represent an upper limit and the values obtained by model 2 represent a lower limit for the fat masses derived with ADP. There is a significant correlation between ADP and both MRI models (model 1: r ¼ 0.98, P < 0.0001; model 2: r ¼ 0.97, P < 0.0001). Compared with ADP, the absolute fat masses obtained with MRI were on average higher by 17.9% when using model 1 and lower by 11.1% when using model 2 (Fig. 4b). It is known that both MRI models under- and overestimate the fat mass, respectively. The mean of both models, for BMI >28, is only 2% less than the fat mass determined by ADP. Reproducibility of MRI The repeated MR measurements of the abdominal region of one volunteer showed high reproducibility regarding the automatic quantification of the total, subcutaneous and internal fat volume. The relative standard deviation for five measurements on system A on 1 day were less than 1.3% for all fat compartments. The standard deviations for all 15 measurements were less than 1.6% for all compartments. No relevant differences were observed when system B was used. Automated and Manually Refined Segmentation The variability of internal fat volume for the automatic and manually refined segmentation were cv < 1.5% (Table 1). The refinement of the segmentation did not improve the reproducibility. (In fact, the manually refined cv was found to be worse in one case.) Interobserver Variability The interobserver variability for manual segmentation resulted in a mean difference over the observers of less than 1% total fat volume for each of the selected datasets. Percentage differences for internal and subcutaneous fat were less than 5% and 4%, respectively, in each dataset. The results are shown in Figure 5. The refinement by different observers had a mean coefficient of variability of 0.23%. DISCUSSION The first aim of our study was to evaluate the performance of an automatic segmentation of

Figure 5. Scatter plots show the interobserver variability of internal fat volume measurements of five datasets analyzed by four operators. For subcutaneous and total fat tissue, the reproducibility is slightly better (not shown). Correlation coefficients of the analysis results between two operators were 0.99 or better for all cases. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]

subcutaneous and internal fat in the abdomen based on continuously moving table (CMT) whole-body MRI measurements. This included an explicit signal weighting correcting partial volume effects. The successful acquisition of complete whole-body datasets using CMT MRI allowed a direct comparison with ADP which was the second aim of our study. Several studies have already been published based on either a stationary (7,8) or a CMT approach (9,11) for whole-body fat assessment. Compared with the stationary approach, the intrinsically fast CMT technique enables seamless coverage from head to toe in a reasonable acquisition time. Despite the generally fast CMT acquisition time, the intended high spatial resolution, the use of local array coils and separate arm measurements lengthened the examination time. While in some studies patient positioning with arms extended above the head (7,8,11) were used which hinders the use of local array coils, a conventional supine position was preferred for this study due to patient comfort issues. The limited 500 mm FOV in xy direction of both MR systems required additional arm measurements to obtain complete, high precision MR whole-body data, to allow for comparison with ADP measurements. These separate measurements of the left arm performed in the same arm position as for the whole-body measurements might be a source of error, due to the transition region and side to side differences. However, we assume these errors to be negligible for the calculation of the total amount of body fat. The internal fat is not affected at all by this. Because for most clinical applications of fat water quantification a restriction to the abdominal region is sufficient an extra acquisition of the arms could be omitted. Additional measures to reduce acquisition times are to increase slice thickness and slice distance (7,8). Hence, in practice, measurements might be performed

Whole-Body Fat Quantification: MRI Versus ADP

much faster. Limiting CMT MR imaging to the abdominal region would reduce total examination time to approximately 10–15 min including preparation time. A further improvement in patient comfort as well as an acceleration of the data acquisition can be achieved with additional motion correction methods (42), allowing free breathing during the measurement. With regard to treatment of obesity, the fat quantification for patients with a high BMI would be especially desirable, but the mentioned hardware limitations disallow imaging of subjects with BMIs higher than approximately 45 kg/m2. Qualitatively, the automatic segmentation of subcutaneous and internal fat in the abdomen worked reliably. The comparison of the manually corrected and the automatic segmentation with fixed parameters, which were determined a priori (35), showed good robustness of the overall quantification method with only minor inaccuracies in segmentation. The reason for this are voxels in the transition area between subcutaneous and internal fat, which typically have only small or zero intensities in the fat images, and therefore do not contribute significantly to the quantification results. The idea of using snakes for the segmentation of abdominal fat compartments has already been reported (4,8). Compared with these approaches, the used implementation of the proposed method initially needs user input to define the boundaries of the abdominal region. A combination with methods as proposed to obtain a fully automatic procedure is possible (8), but was not the main focus of this study. Nevertheless, the signal weighted fat quantification method is not restricted to the segmentation with snakes, but can be combined with other segmentation techniques as long as partially filled fat voxels are not excluded from the analysis. To set up an accurate algorithm for fat quantification in vivo, a fast method for B1-inhomogeneity correction similar to the one published (10) was implemented. Compared with other approaches (43,44) the proposed method is simple and easy to implement, and worked reliably with the given data. However, the fat quantification method suggested in this study could be combined with other B1 inhomogeneity correction methods as proposed (43). There were also some limitations in ADP quantification: imperfections in the ADP quantification procedure may lead to differences in the determined fat quantities. Individual variations from the underlying two compartment model of the human body restrict the accuracy of ADP-derived absolute fat quantities to approximately 6 10% (6 3.8% of total body mass) (41). Possible larger differences between ADP and MRI fat quantification results in subjects with small BMI, as for the two male subjects with the smallest BMI in this study, seem to be systematic and be caused by simplifications done in both MR models. They should be increasingly negligible for people with higher fat contents due to their increasing fat fractions in adipose tissue of up to 94% (40). Furthermore, they will cancel out completely if only ratios between subcutaneous and internal fat parts are to be considered.

1443

The expectations that the MR fat masses obtained with the two models represent an upper and a lower limit for the true fat content are confirmed by our data. For further simplification one could take the mean values of the results from both models. However, a generalization regardless of gender of the underlying relations from Martin et al. in the second model might be problematic because they only rely on measurements in six male cadavers (39). This fact could also explain the different result obtained with the female volunteer. Despite these limitations, the presented MRI segmentation and quantification results proved to be highly reproducible even across different MR systems and over multiple days. Differences in the volunteer’s gastric and intestinal fluid volume have been shown to be negligible. These results suggest that the proposed method for fat quantification is able to resolve temporal changes in total fat amount or fractions of subcutaneous and internal fat. ADP only measures the body volume and assumes that all nonfatty tissue is homogeneous among subjects. We could not find any published study comparing absolute fat quantities obtained with MRI and ADP as done in this study. In addition to the comparison with ADP, a direct comparison to a technique like the “four-compartment-model” in quantifying total body fat (45,46) would, therefore, have been desirable. However, the determination of body fat with the four-compartment model requires additional modalities to be applied like DEXA, hydrodensitometry and isotopic dilution (45), which were not available for this study. There is room for improvement regarding further separation of internal fat into compartments. Meanwhile there are various methods published using advanced postprocessing (9). For an abdominal subvolume of a 3D whole-body dataset the intermuscular fatty tissue and bone marrow of the spine and the pelvis have been separated using a probability model based on a visceral AT reference (9). Our method could be combined with this approach. Anyhow not all applications may need a further separation of internal fat. In conclusion, a direct quantitative comparison of whole-body MRI with ADP data was successfully performed. An automatic algorithm for whole-body fat quantification and segmentation of subcutaneous and internal fat in the abdomen was presented. The proposed method includes an explicit signal weighting correcting partial volume effects. Quantification results show high reproducibility and good correlation to the fat quantities derived with ADP. Unlike ADP, which provides only whole-body fat quantities, the MRI based approach enables spatially resolved separation and quantification of internal and subcutaneous fat. ACKNOWLEDGMENT JBH wishes to thank the Innovationsfonds BadenW€ urttemberg and the Academy of Excellence of the DFG. The authors declare no conflict of interest.

1444

Ludwig et al.

REFERENCES 1. Haslam DW, James WPT. Obesity. Lancet 2005;366:1197–1209. doi: 10.1016/S0140–6736(05)67483-1. 2. Powell K. Obesity: the two faces of fat. Nature 2007;447:525–527. doi: 10.1038/447525a. 3. Fox CS, Massaro JM, Hoffmann U, et al. Abdominal visceral and subcutaneous adipose tissue compartments: association with metabolic risk factors in the Framingham Heart Study. Circulation 2007;116:39–48. doi: 10.1161/CIRCULATIONAHA.106.675355. 4. Positano V, Gastaldelli A, Sironi AM, Santarelli MF, Lombardi M, Landini L. An accurate and robust method for unsupervised assessment of abdominal fat by MRI. J Magn Reson Imaging 2004;20:684–689. doi: 10.1002/jmri.20167. 5. Ross R, Shaw KD, Martel Y, de Guise J, Avruch L. Adipose tissue distribution measured by magnetic resonance imaging in obese women. Am J Clin Nutr 1993;57:470–475. 6. Vogt FM, Ruehm S, Hunold P, et al. [Rapid total body fat measurement by magnetic resonance imaging: quantification and topography]. Rofo 2007;179:480–486. doi: 10.1055/s-2007–962833. 7. Machann J, Thamer C, Schnoedt B, et al. Standardized assessment of whole body adipose tissue topography by MRI. J Magn Reson Imaging 2005;21:455–462. doi: 10.1002/jmri.20292. 8. W€ urslin C, Machann J, Rempp H, Claussen C, Yang B, Schick F. Topography mapping of whole body adipose tissue using A fully automated and standardized procedure. J Magn Reson Imaging 2010;31:430–439. doi: 10.1002/jmri.22036. 9. Kullberg J, Johansson L, Ahlstr€ om H, et al. Automated assessment of whole-body adipose tissue depots from continuously moving bed MRI: a feasibility study. J Magn Reson Imaging 2009;30: 185–193. doi: 10.1002/jmri.21820. 10. Leinhard OD, Johansson A, Rydell J, et al. Quantitative abdominal fat estimation using MRI. In: Proceedings of the 19th International Conference on Pattern Recognition, ICPR. 2008:1–4. doi: 10.1109/ICPR.2008.4761764. 11. Berglund J, Johansson L, Ahlstr€ om H, Kullberg J. Three-point Dixon method enables whole-body water and fat imaging of obese subjects. Magn Reson Med 2010;63:1659–1668. doi: 10.1002/ mrm.22385. 12. Donnelly LF, O’Brien KJ, Dardzinski BJ, et al. Using a phantom to compare MR techniques for determining the ratio of intraabdominal to subcutaneous adipose tissue. AJR Am J Roentgenol 2003;180:993–998. 13. B€ ornert P, Keupp J, Eggers H, Aldefeld B. Whole-body 3D water/ fat resolved continuously moving table imaging. J Magn Reson Imaging 2007;25:660–665. doi: 10.1002/jmri.20861. 14. Sommer G, Fautz H-P, Ludwig U, Hennig J. Multicontrast sequences with continuous table motion: a novel acquisition technique for extended field of view imaging. Magn Reson Med 2006;55:918–922. doi: 10.1002/mrm.20838. 15. Lukaski HC, Johnson PE, Bolonchuk WW, Lykken GI. Assessment of fat-free mass using bioelectrical impedance measurements of the human body. Am J Clin Nutr 1985;41:810–817. 16. Fields DA, Goran MI, McCrory MA. Body-composition assessment via air-displacement plethysmography in adults and children: a review. Am J Clin Nutr 2002;75:453–467. 17. Lee SY, Gallagher D. Assessment methods in human body composition. Curr Opin Clin Nutr Metab Care 2008;11:566–572. doi: 10.1097/MCO.0b013e32830b5f23. 18. Di Sebastiano KM, Mourtzakis M. A critical evaluation of body composition modalities used to assess adipose and skeletal muscle tissue in cancer. Appl Physiol Nutr Metab 2012;37:811–821. doi: 10.1139/h2012-079. 19. McCarthy EA, Strauss BJG, Walker SP, Permezel M. Determination of maternal body composition in pregnancy and its relevance to perinatal outcomes. Obstet Gynecol Surv 2004;59:731–742; quiz 745–746. 20. Ellis KJ. Human body composition: in vivo methods. Physiol Rev 2000;80:649–680. 21. Peters A, Bosy-Westphal A, Kubera B, et al. Why doesn’t the brain lose weight, when obese people diet? Obes Facts 2011;4:2–2. doi: 10.1159/000327676. 22. Winsley RJ, Fulford J, MacLeod KM, Ramos-Ibanez N, Williams CA, Armstrong N. Prediction of visceral adipose tissue using air displacement plethysmography in children. Obes Res 2005;13: 2048–2051. doi: 10.1038/oby.2005.252. 23. Schautz B, Later W, Heller M, M€ uller MJ, Bosy-Westphal A. Associations between breast adipose tissue, body fat distribution and

24.

25.

26.

27.

28.

29.

30.

31.

32.

33.

34. 35.

36.

37.

38.

39.

40. 41. 42.

43.

44.

45.

46.

cardiometabolic risk in women: cross-sectional data and weightloss intervention. Eur J Clin Nutr 2011;65:784–790. doi: 10.1038/ejcn.2011.35. Hu HH, Nayak KS. Quantification of absolute fat mass using an adipose tissue reference signal model. J Magn Reson Imaging 2008;28:1483–1491. doi: 10.1002/jmri.21603. Ludwig U, Sommer G, Zaitsev M, Ghanem N, Hennig J, Fautz H-P. 2D axial moving table acquisitions with dynamic slice adaptation. Magn Reson Med 2006;55:423–430. doi: 10.1002/mrm.20779. Fautz H-P, Kannengiesser SAR. Sliding multislice (SMS): a new technique for minimum FOV usage in axial continuously movingtable acquisitions. Magn Reson Med 2006;55:363–370. doi: 10.1002/mrm.20762. McGibney G, Smith MR, Nichols ST, Crawley A. Quantitative evaluation of several partial Fourier reconstruction algorithms used in MRI. Magn Reson Med 1993;30:51–59. Griswold MA, Jakob PM, Heidemann RM, et al. Generalized autocalibrating partially parallel acquisitions (GRAPPA). Magn Reson Med 2002;47:1202–1210. doi: 10.1002/mrm.10171. Shen W, Wang Z, Punyanita M, et al. Adipose tissue quantification by imaging methods: a proposed classification. Obes Res 2003;11:5–16. doi: 10.1038/oby.2003.3. Liou T-H, Chan WP, Pan L-C, Lin P-W, Chou P, Chen C-H. Fully automated large-scale assessment of visceral and subcutaneous abdominal adipose tissue by magnetic resonance imaging. Int J Obes 2006;30:844–852. doi: 10.1038/sj.ijo.0803216. Poll LW, Wittsack H-J, Koch J-A, et al. Quantification of total abdominal fat volumes using magnetic resonance imaging. Eur J Med Res 2002;7:347–352. Poll LW, Wittsack H-J, Koch J-A, et al. A rapid and reliable semiautomated method for measurement of total abdominal fat volumes using magnetic resonance imaging. Magn Reson Imaging 2003;21:631–636. doi: 10.1016/S0730-725X(03)00093-6. Brennan DD, Whelan PF, Robinson K, et al. Rapid automated measurement of body fat distribution from whole-body MRI. AJR Am J Roentgenol 2005;185:418–423. Kass M, Witkin A, Terzopoulos D. Snakes: active contour models. Int J Comput Vis 1988;1:321–331. doi: 10.1007/BF00133570. Klausmann F, Ludwig U, Honal M, K€ onig D, Deibert P, Huff S. Accuracy of wholebody fat quantification with MRI: a comparison to air-displacement plethysmography. In: Proceeding of the 18th Annual Meeting of ISMRM, Stockholm; 2010. (abstract 57). Peng Q, McColl RW, Ding Y, Wang J, Chia JM, Weatherall PT. Automated method for accurate abdominal fat quantification on water-saturated magnetic resonance images. J Magn Reson Imaging 2007;26:738–746. doi: 10.1002/jmri.21040. Liu C-Y, McKenzie CA, Yu H, Brittain JH, Reeder SB. Fat quantification with IDEAL gradient echo imaging: correction of bias from T(1) and noise. Magn Reson Med 2007;58:354–364. doi: 10.1002/mrm.21301. Haddad RA, Akansu AN. A class of fast Gaussian binomial filters for speech and image processing. IEEE Trans Signal Process 1991;39:723–727. doi: 10.1109/78.80892. Martin AD, Daniel MZ, Drinkwater DT, Clarys JP. Adipose tissue density, estimated adipose lipid fraction and whole body adiposity in male cadavers. Int J Obes Relat Metab Disord 1994;18:79–83. Thomas LW. The chemical composition of adipose tissue of man and mice. Q J Exp Physiol Cogn Med Sci 1962;47:179–188. Siri WE. The gross composition of the body. Adv Biol Med Phys 1956;4:239–280. Honal M, Leupold J, Huff S, Baumann T, Ludwig U. Compensation of breathing motion artifacts for MRI with continuously moving table. Magn Reson Med 2010;63:701–712. doi: 10.1002/mrm.22162. Vovk U, Pernus F, Likar B. A review of methods for correction of intensity inhomogeneity in MRI. IEEE Trans Med Imaging 2007; 26:405–421. doi: 10.1109/TMI.2006.891486. Positano V, Cusi K, Santarelli MF, et al. Automatic correction of intensity inhomogeneities improves unsupervised assessment of abdominal fat by MRI. J Magn Reson Imaging 2008;28:403–410. doi: 10.1002/jmri.21448. Withers RT, LaForgia J, Pillans RK, et al. Comparisons of two-, three-, and four-compartment models of body composition analysis in men and women. J Appl Physiol 1998;85:238–245. Friedl KE, DeLuca JP, Marchitelli LJ, Vogel JA. Reliability of body-fat estimations from a four-compartment model by using density, body water, and bone mineral measurements. Am J Clin Nutr 1992;55:764–770.