The British Journal of Radiology, 82 (2009), 123130

Whole-body adipose tissue analysis: comparison of MRI, CT and dual energy X-ray absorptiometry
1

J KULLBERG, MS, 2J BRANDBERG, MD, 3J-E ANGELHED, PhD, 1H FRIMMEL, PhD, 3E BERGELIN, 3 M, MD, PhD, 1L JOHANSSON, PhD and 2,3,4L LO NN, MD, PhD L STRID, RN, 1H AHLSTRO
1

RN,

Department of Radiology, Uppsala University Hospital, Uppsala, 2Department of Radiology, Sahlgrenska University Hospital, Go teborg, 3Department of Metabolism and Cardiovascular Research, The Sahlgrenska Academy at Go teborg University, Go teborg, Sweden and 4Faculty of Health Sciences, Radiology and Vascular Surgery, Copenhagen University, Denmark

ABSTRACT. The aim of this study was to validate a recently proposed MRI-based T1mapping method for analysis of whole-body adipose tissue (AT) using an established CT protocol as reference and to include results from dual energy X-ray absorptiometry (DEXA). 10 subjects, drawn from the Swedish Obese Subjects Sibling-pairs study, were examined using CT, MRI and DEXA. The CT analysis was based on 28 imaged slices. T1 maps were calculated using contiguous MRI data from two different gradient echo sequences acquired using different flip angles. CT and MRI comparison was performed slice-wise and for the whole-body region. Fat weights were compared between all three modalities. Strong correlations (r>0.977, p,0.0001) were found between MRI and CT whole-body and AT volumes. MRI visceral AT volume was underestimated by 0.790.75 l (p50.005), but total AT was not significantly different from that estimated by CT (MRI CT 5 0.611.17 l; p50.114). DEXA underestimated fat weights by 5.231.71 kg (p50.005) compared with CT. MRI underestimated whole-body volume by 2.031.61 l (p50.005) compared with CT. Weights estimated either by CT or by DEXA were not significantly different from weights measured using scales. In conclusion, strong correlations were found between whole-body AT results from CT, MRI-based T1 mapping and DEXA. If the differences between the results from T1mapping and CT-based analysis are accepted, the T1-mapping method allows fully automated post-processing of whole-body MRI data, allowing longitudinal wholebody studies that are also applicable for children and adolescents.

Received 26 November 2007 Revised 29 February 2008 Accepted 11 April 2008 DOI: 10.1259/bjr/80083156
2009 The British Institute of Radiology

The prevalence of overweight and obese individuals is rapidly increasing in many countries around the world. Both the amount and the distribution of adipose tissue (AT) are associated with premature death [1]. Therefore, there is an increasing need for accurate and automated tools in the field of body composition. CT is an imaging modality that allows rapid acquisition and accurate body composition analysis [2, 3]. However, in order to reduce the exposure to ionizing radiation, only a limited number of image slices can be used. A well-established and validated CT-based wholebody analysis method that uses 28 slices to divide the body into 12 main compartments of tissues, organs and gas has previously been described [2]. Body weights estimated using this method show only minor deviations from real body weights. Dual energy X-ray absorptiometry (DEXA) was originally constructed for bone density analysis. However, it can also be used in studies of body composition [4, 5]. DEXA body composition analysis gives the total weights of fat, lean tissue and bone mineral content as output. A total body DEXA investigation exposes the subject to
Address correspondence to: Joel Kullberg, MRT, Entrance 24, Uppsala University Hospital, SE-751 85 Uppsala, Sweden. E-mail: joel.kullberg@radiol.uu.se

minimal amounts of radiation (#15 mSv) [6]. As the scanner does not have a gantry, the patient compliance level is high, with respect to claustrophobia. MRI allows the analysis of body composition [3, 79] without any known long-term side effects, allowing large coverage, repeated acquisition and studies of children and adolescents [10]. Whole-body MR analysis is motivated by the need for accurate phenotype determinations. Age and ethnicity have an impact on body composition; hence, studies of these factors probably gain from extensive analysis. Larger coverage, more dense sampling and repeated acquisitions demand automation of the data processing. Automation is complicated because MRI intensity levels are given in arbitrary units (AUs) and images are often affected by intensity inhomogeneities. An MRI-based T1-mapping technique, hereafter denoted as the MRI method, has recently been shown to simplify automated analysis of AT from whole-body data [11]. However, the technique has not yet been validated in a contiguous whole-body region nor have the results been compared with those from previously validated modalities. The objective of this study was to validate the wholebody MRI method using the validated CT-based wholebody method as reference. Whole-body DEXA analysis was also performed. MRI, CT and DEXA results were
123

The British Journal of Radiology, February 2009

J Kullberg, J Brandberg, J-E Angelhed et al

compared on a whole-body level, as well as on a slicewise level (MRI and CT only).

Methods and materials

Subjects and anthropometric measurements
10 volunteer subjects (see Table 1) were scanned using the CT-based 28-slice protocol [2], the MRI protocol [11] and the DEXA protocol [12]. The subjects were members of two nuclear families and also participated in the Swedish Obese Subjects (SOS) Sibling-pairs study. The SOS Sibling-pairs study is an extension of the SOS study, which is a longitudinal study aiming to investigate the health effects from weight loss surgery [13, 14]. The focus of the Sibling-pairs study is to find relationships between genotypes and phenotypes. This study was approved by the local ethics and radiation committees and all subjects gave informed consent. We certify that all applicable institutional and governmental regulations concerning the ethical use of human volunteers were followed during this research. Height and weight were measured using standardized equipment. The coefficient of variation values were below 1% for these measurements. The acquisition of the CT, MRI and DEXA data from each subject was performed at the same site within a 3 h interval. Patients were asked to fast for 5 h prior to scanning. The positioning of the subjects during the CT and MRI acquisitions was as similar as possible to minimize intermodality positioning differences. The similarity of position was achieved by use of experienced personnel and identical cushioning. The shapes of the CT and MRI table-tops differed slightly.

repositioned, after which the slices from the L3 level to the wrists (positions 1229) were acquired. Obese subjects that did not fit into the field of view (FOV) were repositioned to exceed only the FOV on one side of the body. Thus, the other side of the body, including at least 50% of all tissue areas, was scanned and analysed. Scan parameters were 120 kVp, 40240 mAs depending on body part and body size, 480 mm FOV, 5 mm slice thickness and 256 6 256 matrix size. The charge was set according to a dose reduction scheme [15]. The average effective dose per examination was 0.63 mSv (maximum 1.61 mSv).

CT analysis
The acquired CT data was semi-automatically analysed using software developed at the Department of teborg University, Sweden. Measurements Medicine, Go of distance, circumference and area, and calculation of volumes, have previously been described in detail [2]. Calculation of the distance between slice 11 and slice 12 was made by manual determination of the position of a common skeletal feature in both half-body scouts. The analysis subdivides each scanned slice in subareas of tissues, organs and gas [2]. Subcutaneous adipose tissue (SAT) and visceral adipose tissue (VAT, i.e. the sum of intra- and retro-peritoneal AT) were separated using semi-automated delineation. VAT was measured, when present, between the symphysis upper border and lowest diaphragm dome, i.e. in slices 718. The total volume of each tissue was calculated using the measured tissue slice areas and the distances between the slices using Equation 1: V~ Xn{1
i~1

di

ai zaiz1 2

CT hardware and reference protocol

The CT scanner (HiSpeed Advantage, version RP2; General Electric, Milwaukee, WI) was used to acquire 28 axial images from subjects in the supine position (Figure 1). In addition to the 28 image slices, two positions were determined from the scouts in order to measure the total length of the subjects in the supine position. Positions were numbered 130 (including the 28 scans: positions 229) from toes to finger tips (Figure 1). The exact positions of the scans were obtained using the scout images and have previously been described [2]. Owing to limitations in the hardware, the slices were acquired using one repositioning. The slices from the ankle joint to the upper border of the crista (positions 211) were first acquired. The subject was then
Table 1. Subject characteristics
Females (n56) Mean Range Males (n54) Mean Range

Age (years) Height (m) Weight (kg) BMI (kg m2)

51 1.65 76.4 27.9

3870 1.611.71 59.3113 22.938.7

57 4571 1.80 1.771.83 105 92.4124 32.5 28.837.1

BMI, body mass index.

Equation 1 determines the total tissue volume V for each tissue type from n slices (n.1) by using the distance di between slice i and i+1 and the measured tissue areas ai and ai+1 from slice i and i+1, respectively. When the ipsilateral arm of a subject exceeded the FOV (in slices 2428), the measurements were performed in the contralateral arm, and the results were doubled. When slices 623 were affected, the measurements of total and SAT areas from the unaffected side were doubled. The centre line used was manually determined. Total subject weight and fat weight were estimated by the use of density values reported in the literature [2] (AT, 0.923 kg dm3). Precision errors (standard errors) for SAT and VAT from repeated interpretation have previously been determined as 0.5% and 1.2%, respectively [2]. The reproducibility of a CT protocol that acquired and analysed two of the slices (thigh and L4: slices 5 and 11) from the 28-slice protocol used in this study has recently been investigated in 50 obese subjects at the site performing the CT analysis used in this study (Brandberg et al, unpublished data). The standard errors reported from repeated interpretation were 0.5% and 1.1%, and from repeated acquisition were 2.2% and 6.0%, for SAT and VAT, respectively. The time required for the post-processing of the CT data from one subject, using the semiautomatic software, was approximately 3 h.
The British Journal of Radiology, February 2009

124

Whole-body adipose tissue comparison of MRI, CT and DEXA

Figure 1. Illustration of the slice

positions.

MRI hardware and protocol

The contiguous whole-body MRI acquisition was performed on a 1.5 T clinical MRI scanner (Gyroscan NT; Philips Medical Systems, Best, the Netherlands) using a spoiled T1 weighted gradient echo sequence [11]. The main scan parameters were: repetition time, 177 ms; echo time, 2.3 ms; FOV, 530 mm; and slice thickness, 8 mm. Three whole-body volumes were acquired. The first was acquired using a flip angle of 80 and was denoted as Flip80 [11]. The Flip80 data were used for automatic segmentation of the bodies and lungs. It was also used in the visual selection of the MR slices that best corresponded to the acquired CT slices. Scan parameters were turned off to ensure a constant MR signal scaling. Two more wholebody volumes were acquired using flip angles of 80 and 30 . These volumes were denoted as Flip80off and Flip30off, respectively, and were used in the calculation of the T1 relaxation map [11]. Owing to limitations in the hardware, the whole-body volumes needed to be acquired half-body-wise using one subject repositioning. The acquisition has previously been described in detail [11].

Visual selection of the 28 slices within positions 130 from the continuous whole-body data was performed by an experienced radiologist. The slice selections were performed based on the congruence of anatomical structures, using the acquired CT slices as a reference. The MR slice selections were performed using the Flip80 data and the resulting slice positions were used in the slice-wise evaluation. The simple thresholding of the T1-mapped data includes bone marrow (BM) in the AT volumes measured. Accurate exclusion of BM is difficult to achieve in many MR slices acquired using this protocol. However, to assess the effect on the SAT areas measured in the comparison with CT, the BM was manually segmented by an experienced operator in eight slices (2 9) from the binary AT volumes.

DEXA analysis
The DEXA scanner used was a LUNAR DPX-L (Lunar Co., Madison WI) with software version 1.35 and an extended analysis program for total body analysis (LUNAR Radiation, Madison, WI). Body fat, lean tissue mass, bone mineral content and body weight were assessed. Quality assurance tests were conducted on a daily basis. The DEXA hardware used in this study had been validated in a previous study performed by Lantz et al [12]. However, a different version of the software (v1.31) was used. Repeated examinations of 10 females (body mass index, 22.11.6 kg m2), using the old versions of the software, showed body fat and lean tissue mass differences of 1.7% and 0.7%, respectively.

MRI analysis
Registration of the half bodies was performed as previously described [11], with the exception that the registration in the feethead direction was manually performed by an experienced radiologist. As T1-mapped values outside the bodies and in the lungs are of no interest, an automated image processing algorithm was used to segment bodies and lungs from the Flip80 data (see Appendix A). The total subject volumes were estimated from MRI using the segmented whole-body volumes. AT was segmented from the T1-mapped data using thresholds automatically derived from the whole-body T1 histograms using a commonly used method. The histograms were analysed by fitting of two Gaussian functions: one was fitted to the histogram AT peak and the other to the lean tissue peak. The optimization was performed using the summed least squares criterion and the downhill simplex method of Nelder and Mead [16]. The intersection between the two Gaussian functions was chosen as the T1 threshold value. SAT and VAT were separated by an experienced radiologist using manual classification. VAT was classified using binary AT results from the thresholded T1 maps by use of ImageJ software (National Institute of Health, Bethesda, MD) [17]. The total SAT and VAT volumes were estimated using the results from the manually classified binary AT data.
The British Journal of Radiology, February 2009

Whole-body and slice-wise comparisons

Total body, AT, SAT and VAT volumes from the MRI method were evaluated using the volumes from the CTbased analysis as a reference. Body weights estimated by CT and DEXA were evaluated using the body weights from a scale as a reference. Fat weights estimated by MRI and DEXA were compared using the values derived by CT as a reference. Slice-wise total body, SAT and VAT areas from MRI were evaluated using CT as a reference. To visualize the differences in areas estimated by CT and the MRI-based analysis, the contiguous slices were standardized for body size. The 28 selected MRI slice positions from one subject were used as a template and the contiguous slice results from all subjects were linearly interpolated (nearest neighbour) between these 28 slice positions. The slice-wise comparisons did not include DEXA.
125

J Kullberg, J Brandberg, J-E Angelhed et al

did not differ significantly from the fat weights estimated by CT. DEXA was found to underestimate the total fat weights compared with both CT and MRI. Bivariate plots of whole-body volume differences between MRI and CT are given in Figures 24 for total body, SAT and VAT, respectively. The total fat weights measured by CT, MRI and DEXA for the 10 subjects are given in Table 3.

Slice-wise comparisons
The results from the slice-wise comparisons of total area, SAT and VAT between CT and MRI are displayed in Figure 5.
Figure 2. Whole-body total volume differences (MRI CT)
plotted as a function of CT volume measured (the reference). The linear dependence was found to be: 0.045CT + 1.920; r250.39; p50.054.

Statistical analysis
Linear correlations were studied and are reported as correlation coefficients (r-values). Differences were investigated using the Wilcoxon signed-rank test. Bivariate plots and linear regression analysis were used to study dependencies. p-Values of ,0.05 were considered significant.

Total body areas Total slice areas were often underestimated by the MRIbased method compared with CT (see Figure 5a,c). The absolute slice area was underestimated in 11 slice positions and overestimated in 2 slice positions. Linear regression on all absolute area differences showed a significant dependence on slice area (MRI CT 5 0.053CT + 19.1; r50.584, p,0.0001). The slice positioned at the top of the skull (position 27) was found to overestimate the area from the MRI-based method compared with CT by more than 30 cm2 in eight of the subjects. Subcutaneous adipose tissue areas SAT areas were often overestimated by MRI compared with CT (see Figure 5a,d). Significant differences were found in 19 slice positions. Slices from MRI positioned at the ankle joint, knee, pelvis, thorax and wrists were found to overestimate the SAT areas, whereas slices at the L3 level and at the level of the lower orbital border (slices 12 and 25) underestimated the SAT areas when compared with CT. Linear regression on all differences showed a dependence on the SAT area measured by CT (MRI CT 5 0.041CT + 16.7; r50.161, p,0.0001). The slice-wise differences in SAT after manual exclusion of BM are shown in Figure 5c (dotted line).

Results
Whole-body comparisons
Correlations, differences and linear dependencies between CT, MRI and DEXA measurements are given in Table 2. Strong correlations (r>0.977, p,0.0001) were seen for all estimated volumes. The MRI analysis was seen to overestimate SAT volumes, whereas VAT and total body volumes were underestimated. Total subject weights estimated from the CT and the DEXA analyses were not seen to differ from the weights assessed by scales. The fat weights estimated from the MRI analysis

Table 2. Whole-body correlations, differences and linear dependencies between CT, MRI and DEXA measurements (n510)
Volume comparisons (MRI CT) Volume Correlation Total AT SAT VAT 0.998 0.995 0.977 0.987 Difference (l) 22.031.61, 20.611.17, 2.772.41, 20.790.75, p50.005 p50.114 p50.007 p50.005 Linear dependence (l) 20.045CT+1.920, p50.054 20.004CT0.754, p50.916 20.007CT+2.563, p50.933 20.209CT+0.310, p50.002

Subject weight comparisons Modalities Correlation CT scales DEXA scales 1.000 1.000

Difference (kg) 0.400.60, p50.114 20.220.42, p50.114 Difference (kg) 20.561.08, p50.114 25.231.71, p50.005 24.672.38, p50.005

Fat weight comparisons Modalities Correlation MRI CT DEXA CT DEXA MRI 0.995 0.990 0.979

Correlations are given as correlation coefficients (r-values). All correlations were significant (p,0.0001). Differences are given as mean standard deviation. AT, adipose tissue (SAT + VAT); SAT, subcutaneous adipose tissue; VAT, visceral adipose tissue; DEXA, dual energy X-ray absorptiometry.

126

The British Journal of Radiology, February 2009

Whole-body adipose tissue comparison of MRI, CT and DEXA

Figure 3. Whole-body subcutaneous adipose tissue (SAT)

volume differences (MRI CT) plotted as a function of CT volume measured (the reference). Differences (mean standard deviation) are illustrated using lines. No linear dependence was found.

Figure 4. Whole-body visceral adipose tissue (VAT) volume differences (MRI CT) plotted as a function of CT volume measured (the reference). The linear dependence was found to be: 0.209CT + 0.310; r250.72; p50.002.

Significant differences were seen in four slice positions (compared with five before exclusion of BM). SAT was overestimated in slices at the symphysis lower border and acetabulum upper border (slices 6 and 8) after exclusion of BM, whereas the slices at the ankle joint and calf (slices 2 and 3) were underestimated after the exclusion of BM. The absolute differences in SAT areas measured between MRI and CT were significantly reduced in slices 2, 4 and 69 by the exclusion of BM.

Visceral adipose tissue areas VAT areas were underestimated by MRI compared with CT (Figure 5a,e). Significant differences were seen in eight slice positions. Slices between the upper border of the acetabulum and the L1 level (positions 813) and at the positions of the lowest diaphragm dome and 7 mm above this position (slices 16 and 17) were underestimated by MRI compared with CT. Linear regression on all differences showed a dependence on the VAT area measured by CT (MRI CT 5 20.155CT 0.459; r50.575, p,0.0001).

Affected slices
Two MRI slices used in the slice comparison were affected by artefacts or visible geometrical alternations. The body of one subject mildly exceeded the FOV in position 22. The affected area was estimated to be 0.5% of the total in-slice body area. Position 25 in another subject was affected by metal, resulting in an artefact that caused a reduction in the estimated skull area of approximately 18%. In the CT analysis, 42 slices were affected. Subjects exceeded the FOV in positions 626 and 1627. The arms (positions 2427) were, in part, out of the FOV in six subjects. The DEXA examinations did not manage to fully include the arms into the FOV in the three heaviest subjects.

Discussion
We have shown that results from a previously presented MRI method [11] and from DEXA analysis
The British Journal of Radiology, February 2009

give measurements of total and AT volumes and weights that correlate highly with the results from the wholebody CT protocol used as reference. We have also shown that the MRI method gives measurements of total AT volumes with few differences from the whole-body CT protocol. To our knowledge, no whole-body comparisons between CT and MRI, and CT and DEXA, have previously been presented. Total slice areas were seen to be underestimated by the MRI-based analysis compared with CT. Larger areas showed larger absolute and relative underestimations, indicating that geometrical distortion in MRI might be a contributing factor. The overestimations of the total slice areas in the calf and at the top of the skull are probably caused by systematic differences in the slice selection. The automated segmentation algorithm might also be causing the differences in the slice from the top of the skull. The MRI method was frequently seen to overestimate the SAT areas measured when compared with CT. A probable cause is that BM is included in the MRI analysis. Slice area differences for SAT measured by the two methods were smaller after manual exclusion of BM from the analysis (see Figure 5d). Figure 5a,d also shows how the SAT is overestimated in the knees, pelvis and thorax, i.e. in slices containing relatively large amounts of BM. The independence between the absolute difference and total SAT volume measured by CT strengthens the hypothesis that the inclusion of BM, which probably shows a limited variation with differing amounts of SAT, causes the overestimation. Ideally, BM should be excluded from the MRI whole-body volumes. However, accurate exclusion of BM from the whole-body region was found to be difficult at the current MRI resolution. VAT was seen to be underestimated by MRI compared with CT. Absolute differences in VAT slice area and whole-body volume were seen to increase with increasing VAT areas and volumes measured by CT, respectively. However, the relative differences showed no dependence on VAT areas or volumes measured by CT. A previous study has shown that the intestinal content might contribute significantly to the VAT area measured
127

J Kullberg, J Brandberg, J-E Angelhed et al

Figure 5. (a) Slice-wise mean total body, subcutaneous adipose tissue (SAT) and visceral adipose tissue (VAT) areas from the CT and the MRI method are displayed. The 30slice positions used are illustrated in (b). Slice-wise total area, SAT and VAT differences (MRI CT) are given as mean confidence interval in (c), (d) and (e), respectively. In (d), the mean SAT values, after exclusion of bone marrow (BM), are given for the eight slices (dotted line).

using single-slice CT analysis [18]. Intestinal content could possibly affect the results from the separate modalities differently. The presence of air is, for
Table 3. Total fat weights in the 10 subjects, measured by CT, MRI and DEXA
Whole-body fat weight (kg) Subject CT MRI DEXA

1 2 3 4 5 6 7 8 9 10

32.78 48.43 33.95 23.29 39.94 42.05 25.24 30.58 25.35 58.33

30.90 50.06 33.84 23.24 39.88 40.27 24.35 29.18 25.49 57.10

26.07 39.09 28.42 19.77 35.96 37.27 21.19 26.33 20.52 53.03

DEXA, dual energy X-ray absorptiometry.

example, well known to cause T2* relaxation, resulting in a reduced MR signal. Differences observed between MRI and CT were probably caused by errors resulting from soft-tissue motion, motion artefacts, the MRI protocol used, partial volume effects, differences in slice positions, image resolution differences, image artefacts, FOV exceedings and non-identical positioning of the subjects. Partial volume effects probably affect the CT and MRI methods differently, especially as the out-of-phase MRI protocol was used. The fact that half bodies were acquired and registered in the post-processing in both the CT and MRI protocols limits the accuracy of both methods. The presence of dental metal in one subject gave imaging artefacts on MRI and the fact that six subjects exceeded the FOV during CT acquisition also limits the accuracy of the results. The previous validation of the CT protocol was based on a comparison with total body weight measured by scales, and did not validate the AT measurements specifically. Minor differences in AT
The British Journal of Radiology, February 2009

128

Whole-body adipose tissue comparison of MRI, CT and DEXA

measurements between MRI and CT could therefore be caused by inaccuracies in either of these two methods. Multiple sources of the differences/errors mentioned above affect both MRI and CT and introduce bias to both methods. This should be kept in mind when the CT results are used as a reference. The Flip80 data was occasionally seen to have voxels in the skinbackground interface with values lower than Tbody (see Appendix A), e.g. by the mouth or by the knees, where the SAT layers are absent or thin. The region-growing of the background was seen to enter the body at these locations. Therefore, the region growing was limited by use of the binary mask Mmin. The binary mask needed to fill a volume slightly smaller than the body. The mask created using slice-wise dilation and erosion gave an accurate border, and the mask created using dilation and erosion in three dimensions was used to fill the body, as this was not always the case using the slice-wise approach in, for example, slices at the level of the mouth or nose. The segmented lungs were occasionally, by visual inspection, seen to underestimate the lung volume. This was caused by motion artefacts from the heart and by the segmentation method used. Fat weight was underestimated by DEXA analysis. This is partly explained by the three heaviest subjects exceeding the DEXA FOV. However, the underestimation was not eliminated by excluding these subjects, suggesting that there were additional causes of the differences. Large obese subjects (.120 kg) have previously been reported to give deviating results [19, 20]. Inter- and intra-manufacturer variations in DEXA measurements have also previously been reported [12, 21, 22]; indeed, inter-manufacturer variations of up to 10% of fat mass have been found [22]. Measurements of abdominal fat have previously been shown to underestimate the fat mass by 10% compared with CT [23]. When comparing AT volume results from imaging techniques with the fat weights estimated by DEXA, an assumption on AT fat density is needed. This density assumption will affect directly the comparison. MRI and CT, as opposed to DEXA, allow direct separation of VAT and SAT from the acquired images. These modalities are therefore often chosen in studies of different adipose tissue depots. One limitation of this study is the small number of subjects included. The body weight and body mass index range were therefore not optimal. A shortcoming of MRI, and sometimes also of CT and DEXA, is the hardware, which limits the size and weight of the subjects to be examined. The acquisition and postprocessing required for the MRI method might also limit the availability of this method. In the comparison of total subject weights (Table 2), MRI was not included, as the estimation of the non-AT class density is outside the scope of this study. In conclusion, strong correlations between body composition analysis results in whole-body regions from CT, MRI and DEXA have been shown. We believe that the differences between the results from MRI vs CT are acceptable. Additionally, the MRI method allows a fully automated post-processing of whole-body MR data, allowing longitudinal whole-body studies that are also applicable to children and adolescents.
The British Journal of Radiology, February 2009

Acknowledgments
This study was supported by the Swedish Research Council, Grant no. K2006-71X-06676-24-3, and the teborg Medical Society. Go

References
1. Van Gaal LF, Mertens IL, De Block CE. Mechanisms linking obesity with cardiovascular disease. Nature 2006;444:87580. stro m L, Alpsten M, Kostanty J, Kvist H, 2. Chowdhury B, Sjo fgren R. A multicompartment body composition techniLo que based on computerized tomography. Int J Obes Relat Metab Disord 1994;18:21934. 3. Ross R. Advances in the application of imaging methods in applied and clinical physiology. Acta Diabetol 2003;40:4550. 4. Pietrobelli A, Formica C, Wang Z, Heymsfield SB. Dualenergy X-ray absorptiometry body composition model: review of physical concepts. Am J Physiol 1996;271:94151. 5. Bachrach LK. Dual energy X-ray absorptiometry (DEXA) measurements of bone density and body composition: promise and pitfalls. J Pediatr Endocrinol Metab 2000;13:9838. 6. Mattsson S, Thomas BJ. Development of methods for body composition studies. Phys Med Biol 2006;51:20328. 7. Ross R, Goodpaster B, Kelley D, Boada F. Magnetic resonance imaging in human body composition research. From quantitative to qualitative tissue measurement. Ann N Y Acad Sci 2000;904:127. 8. Ellis KJ. Human body composition: in vivo methods. Physiol Rev 2000;80:64980. 9. Kullberg J, Ahlstrom H, Johansson L, Frimmel H. Automated and reproducible segmentation of visceral and subcutaneous adipose tissue from abdominal MRI. Int J Obes (Lond) 2007;31:180617. 10. Formica D, Silvestri S. Biological effects of exposure to magnetic resonance imaging: an overview. Biomed Eng Online 2004;3:11. nn L, Brandberg J, Ahlstro m H, 11. Kullberg J, Angelhed JE, Lo Frimmel H, et al. Whole-body T1 mapping improves the definition of adipose tissue: consequences for automated image analysis. J Magn Reson Imaging 2006;24:394401. stro m L. 12. Lantz H, Samuelson G, Bratteby LE, Mallmin H, Sjo Differences in whole body measurements by DXA-scanning using two Lunar DPX-L machines. Int J Obes Relat Metab Disord 1999;23:76470. strom L, Larsson B, Backman L, Bengtsson C, Bouchard 13. Sjo C, Dahlgren S, et al. Swedish obese subjects (SOS). Recruitment for an intervention study and a selected description of the obese state. Int J Obes Relat Metab Disord 1992;16:46579. strom L, Lindroos AK, Peltonen M, Torgerson J, 14. Sjo Bouchard C, Carlsson B, et al. Lifestyle, diabetes, and cardiovascular risk factors 10 years after bariatric surgery. N Engl J Med 2004;351:268393. nn L, Cederblad A, Forssell-Aronsson E, 15. Starck G, Lo stro m L, Alpsten M. A method to obtain the same levels Sjo of CT image noise for patients of various sizes, to minimize radiation dose. Br J Radiol 2002;75:14050. 16. Nelder JA, Mead R. A simplex method for function minimization. Comput J 1965;7:30813. 17. Rasband WS. ImageJ (version vl.37s). 1.37s ed. Bethesda, MD: National Institute of Health; 2006. [Available from: http://rsb.info.nih.gov/ij/]. 18. Potretzke AM, Schmitz KH, Jensen MD. Preventing overestimation of pixels in computed tomography assessment of visceral fat. Obes Res 2004;12:1698701. 19. Svendsen OL. Body composition and fat distribution by dual energy X-ray absorptiometry in overweight postme-

129

J Kullberg, J Brandberg, J-E Angelhed et al

nopausal women. Effect of energy-restriction and exercise. Dan Med Bull 1996;43:24962. Laskey MA. Dual-energy X-ray absorptiometry and body composition. Nutrition 1996;12:4551. Genton L, Hans D, Kyle UG, Pichard C. Dual-energy X-ray absorptiometry and body composition: differences between devices and comparison with reference methods. Nutrition 2002;18:6670. Kistorp CN, Svendsen OL. Body composition analysis by dual energy X-ray absorptiometry in female diabetics differ between manufacturers. Eur J Clin Nutr 1997;51:44954. Snijder MB, Visser M, Dekker JM, Seidell JC, Fuerst T, Tylavsky F, et al. The prediction of visceral fat by dualenergy X-ray absorptiometry in the elderly: a comparison with computed tomography and anthropometry. Int J Obes Relat Metab Disord 2002;26:98493.

20. 21.

22.

23.

Appendix A
Segmentation of bodies and lungs from the MRI Flip80 data
The automated segmentation of the bodies used a region-growing approach combined with two binary whole-body masks, denoted Mcrude and Mmin. Mcrude was used to remove motion artefacts outside the body and Mmin was used to prevent the region-growing of the background from penetrating the body. The segmentation of the lungs was performed by the use of thresholding, morphological operations and filtering based on object size. Mcrude was created by applying a region-growing in three dimensions of the body voxels of intensity value greater than a threshold value (Tcrude 5 130). The result from the region-growing is a binary body object containing holes from low-intensity regions inside the body. One iteration of morphological dilation (18 neighbourhood (NB)) was applied to close the surface of the body. A region-growing of the background was applied to create a binary body object without holes in the low-intensity body regions. One iteration of erosion (18 NB) was applied to correct for the previously performed dilation. Mmin was created from the union of two binary masks. The first binary mask was created by applying three iterations of erosion (18 NB) in three dimensions to the Mcrude mask. The three iterations were determined empirically to result in a binary body mask enclosed within the subject bodies. The second binary mask was created by region-growing of the body voxels of intensity value greater than Tcrude. The majority of the holes in the

body surface were closed by two iterations of dilation (4 NB, axial slices) in two dimensions. A region-growing of the background was applied to separate the background from the body. Three iterations of erosion (4 NB, axial slices) in two dimensions of the binary body were applied to ensure that the mask was slightly smaller than the body. Because image intensity levels were seen to vary between the acquired stacks, the bodies were segmented using a stack-wise approach. A threshold level, denoted Tbody, was automatically determined for each stack. An intensity-based region-growing of the background until Tbody was reached was used to determine the extension of the body in the skinbackground interface. The region-growing was limited to the region outside Mmin, and motion artefacts outside the body caused by pulsatile flow and motion were removed by masking using Mcrude. Tbody was determined as the average of the two intensity levels IBG and IAT derived from the intensity values of the background and AT histogram peaks, respectively. Least square fitting of Gaussian functions to the peaks was performed. The Gaussian function centres were used as the intensity levels. The underlying assumption was that voxels containing equal amounts of background and AT have an intensity value equal to Tbody. The segmentation of the lungs first used a thresholding of voxels in the intensity interval 080 AU inside the segmented body, resulting in a binary volume mainly containing voxels originating from the lungs, intestinal gas and cortical bone. One iteration of erosion (18 NB) was applied to remove most of the unwanted binary objects. The binary objects were then filtered based on their three-dimensional (18 NB) connected volumes. Objects with volumes smaller than 1000 voxels (<34 cm3) were removed. The remaining objects were at this point crude estimates of the lungs, and hence were denoted the crude lung mask, MCL. Voxels of intensity value ,Tlungs connected in three dimensions (18 NB) to MCL in a region, Rlungs, were considered to delineate the lungs. Rlungs was determined by the use of the binary volume resulting from application of three iterations of dilation to MCL. The threshold value Tlungs was determined by least square fitting of a Gaussian function to the lung peak in the histogram from Rlungs. The value corresponding to inclusion of 99.9 % of the Gaussian function area was selected as Tlungs.

130

The British Journal of Radiology, February 2009