International Journal of Obesity (2006) 30, 844–852 & 2006 Nature Publishing Group All rights reserved 0307-0565/06 $30.00 www.nature.com/ijo
ORIGINAL ARTICLE Fully automated large-scale assessment of visceral and subcutaneous abdominal adipose tissue by magnetic resonance imaging T-H Liou1,2, WP Chan3, L-C Pan4, P-W Lin2, P Chou1, and C-H Chen2,5 1
Community Medicine Research Center and Institute of Public Health National Yang-Ming University, Taipei, Taiwan; Department of Physical Medicine and Rehabilitation Taipei Medical University-Wan Fang Hospital, Taipei, Taiwan; 3 Department of Radiology Taipei Medical University-Wan Fang Hospital, Taipei, Taiwan; 4Department of General Education Taipei Medical University, Taipei, Taiwan; and 5Graduate Institute of Medical Informatics Taipei Medical University, Taipei, Taiwan 2
Objective: To describe and evaluate a fully automated method for characterizing abdominal adipose tissue from magnetic resonance (MR) transverse body scans. Methods: Four MR pulse sequences were applied: SE, FLAIR, STIR, and FRFSE. On 39 subjects, each abdomen was traversed by 15 contiguous transaxial images. The total abdominal adipose tissue (TAAT) was calculated from thresholds obtained by slice histogram analysis. The same thresholds were also used in the manual volume calculation of TAAT, subcutaneous abdominal adipose tissue (SAAT) and visceral abdominal adipose tissue (VAAT). Image segmentation methods, including edge detection, mathematical morphology, and knowledge-based curve fitting, were used to automatically separate SAAT from VAAT in various ‘nonstandard’ cases such as those with heterogeneous magnetic fields and movement artefacts. Results: The percentage root mean squared errors of the method for SAAT and VAAT ranged from 1.0 to 2.7% for the four sequences. It took approximately 7 and 15 min to complete the 15-slice volume estimation of the three adipose tissue classes using automated and manual methods, respectively. Conclusion: The results demonstrate that the proposed method is robust and accurate. Although the separation of SAAT and VAAT is not always perfect, this method could be especially helpful in dealing with large amounts of data such as in epidemiological studies. International Journal of Obesity (2006) 30, 844–852. doi:10.1038/sj.ijo.0803216; published online 17 January 2006 Keywords: image segmentation; mathematical morphology; knowledge-based curve fitting; magnetic resonance imaging (MRI); abdominal fat volume
Introduction Substantial evidence suggests that the volume and distribution of visceral abdominal adipose tissue (VAAT) plays a critical role in the pathogenesis of the metabolic syndrome.1–4 A high level of VAAT has been linked to cardiovascular and metabolic diseases, such as type 2 diabetes mellitus.5 Subjects are considered to be at risk, even if they are not obese, if they have increased adipose tissue deposits in the abdominal area. Correspondence: Dr C-H Chen, Department of Physical Medicine and Rehabilitation Wan Fang Hospital, 111 Hsing-Long Rd Sec. 3, Taipei 11696, Taiwan. E-mail:
[email protected] Received 23 March 2005; revised 19 October 2005; accepted 13 November 2005; published online 17 January 2006
The accurate measurement of the volumes of VAAT and total abdominal adipose tissue (TAAT) is often critical in metabolism and obesity-related research and clinical applications. Methods of adipose tissue assessment have progressed significantly from simple anthropometric measurements, such as body mass index (BMI) and waist-to-hip ratio, to expensive high-technology measurements, such as computed tomography (CT) and magnetic resonance imaging (MRI), using single- and multiple-slice techniques.6 The waist-to-hip ratio has significant correlation with the ratio of the area of intraabdominal fat to that of subcutaneous fat as determined by CT. However, it remains an indirect means for calculating intraabdominal adipose deposits. Sonography is a safe, inexpensive, and widely available imaging technique. However, its relatively inconsistent inter- and intraoperator reproducibilities are major shortcomings.7,8 Dual energy
Automated large-scale assessment of abdominal fat T-H Liou et al
845 X-ray absorptiometry (DEXA) provides a relatively rapid and accurate assessment of total body fat with relatively low radiation exposure but does not directly quantify VAAT.9–11 Whereas CT provides more accurate estimation of VAAT, subjects are exposed to ionizing radiation, which makes it unsuitable for large-scale research.12,13 With the advantages of noninvasiveness and excellent soft-tissue contrast by using different imaging parameters, MRI provides a method to measure adipose tissue safely and accurately.8,10,14–25,27 However, measuring the fat quantity and distribution in the abdomen by MRI is generally laborious. Two steps are required to identify the volume of VAAT. The first step is to separate fat from nonfat tissue. Thresholding on a histogram8,14–17,19,25 is frequently used for this task and could be accomplished automatically with good agreement between automatically and manually set thresholds.17,19 The second step is to locate VAAT on MR image slices. It is often timeconsuming to delineate the region of interest (ROI) of VAAT because of the complex structure of the viscera.26 Almost all the previous studies of this step have reported requiring more or less human interventions, which are either laborious or subject to technical problems8,10,14–17,19,24–25 (Table 1). For example, Elbers et al.15 measured visceral fat areas using an algorithmic seed-growing procedure. The neighboring pixels of this seed and their adjacent pixels that lay between a selected upper and lower signal intensity value were automatically included in the delineation of the fat area. However, both the seed points and the thresholds had to be defined manually. In Gronemeyer et al.’s study17 in 2000, although the thresholds were automatically set, VAAT had to be manually removed for the calculation of the subcutaneous abdominal adipose tissue (SAAT) area. Poll et al.’s study19 in 2002 took a similar approach, but the VAAT was separated by ROIs drawn manually. In this manuscript, we propose a fully automatic SAAT and VAAT separation method that allows abdominal fat distribution to be analyzed in a short time without any user intervention. Our method is accurate to within a few percent errors as compared with manual partition of subcutaneous and visceral abdominal fat distribution for nonobese to severely obese subjects. The method is very robust and can be applied to so-called ‘nonstandard’ cases, those with Table 1
imaging artefacts and unusual anatomy, with four relatively arbitrarily chosen MR pulse sequences.
Methods Subjects In total, 39 subjects (29 women and 10 men, aged 19–55 years (mean ¼ 42), BMI 23.4–39.4 kg/m2 (mean ¼ 29.0)) were enrolled and underwent abdominal MRI scans. The study was approved by the Institutional Human Subject Review Board of Wan Fang Hospital, Taipei, Taiwan. All subjects gave written informed consent before participating.
MRI protocols All subjects were scanned in a supine position in a 1.5 T Horizon scanner (General Electric, Milwaukee, WI, USA) using a standard body coil. Four MR pulse sequences were applied: the conventional T1-weighted spin echo (SE) and three relatively arbitrarily chosen protocols for which the adipose tissue was the brightest among various body tissues (1) T1-weighted fluid attenuated inversion recovery (FLAIR), (2) short inversion time inversion recovery (STIR), and (3) T2-weighted fast recovery fast spin-echo (FRFSE)). The machine settings for the four protocols are listed in Table 2. Subjects were asked to hold their breath while contiguous 10 mm axial slices were taken from 7.5 cm above to 7.5 cm below the umbilicus. Each display matrix was 512 512 with a pixel size of 0.86 0.86 mm. Most subjects were imaged on two occasions 1 month apart.
Automated image analysis The data obtained from the MR imaging sequences were transferred to personal computers in DICOM (Digital Imaging and Communications in Medicine) format. The platform for automated analysis was a personal computer equipped with an Intel Pentium 4 CPU 2.40 GHz and 1G RAM running RedHat Linux 9. A software program was developed at our institute by Dr C-H Chen under the IDL 6.0 (available from Research Systems Inc., Boulder, CO, USA)
Examples of previous adipose tissue assessment using MRI
Year
First author
Remarks
1990 1995 1997 1997 1999 2000 2002 2002 2002 2003 2003
Seidell Johannes Elbers Han Kamel Gronemeyer Pichiecchio Yang Poll Liu Donnelly
Thresholds manually set, SAAT and VAAT manually identified Automatic segmentation for different tissues in brains; not applicable for distinguishing SAAT from VAAT Seed point and the upper and lower intensity thresholds manually defined. A contour line was drawn to separate VAAT and SAAT Thresholds manually set; VAAT was encircled manually by an ovoid line Used ANALYZE image software; a line can be drawn manually to separate SAAT and VAAT Thresholds automatically set; VAAT manually removed, not fully automated. Used umbilical single slice only Fixed thresholds based on previous experimental results; SAAT and VAAT manually identified Automatic segmentation of different tissue in thighs; not applicable for distinguishing SAAT from VAAT Semiautomatic; thresholds automatically determined; ROIs were drawn manually Thresholds manually set; SAAT and VAAT manually identified For phantom only, clear separation between VAAT and SAAT; no details to tell if fully automated
International Journal of Obesity
Automated large-scale assessment of abdominal fat T-H Liou et al
846 Table 2
Parameters of MR pulse sequences
Sequence
TE (ms)
TI (ms)
TR (ms)
ETL
BW (kHz)
SE FLAIR STIR FRFSE
Min Minfull 20 85
F 760 300 F
440 1900 820 3300
F 8 10 20
15.63 20.83 31.25 31.25
Acquisition matrix
NEX
256 160 256 128 256 160 256 160
2 1 2 1
MR ¼ magnetic resonance; TE ¼ echo time; TI ¼ inversion time; TR ¼ repetition time; ETL ¼ echo train length; BW ¼ bandwidth; NEX ¼ number of excitations. SE ¼ T1-weighted spin echo; FLAIR ¼ T1-weighted fluid attenuated inversion recovery; STIR ¼ short inversion time inversion recovery; FRFSE ¼ T2-weighted fast recovery fast spin-echo.
images. After all the slices were processed, the volumes of different adipose tissues were obtained, as described in the Manual Image Analysis section. Formation of TAAT mask. The initial threshold (body threshold) was empirically set at 20% higher than the average pixel value of an image slice. An approximate body mask was obtained by binarizing the image at that threshold. The mask was refined by a mathematical morphological operation28 ‘open’ to remove small dots outside the approximate body mask, and then ‘close’ was applied to fill in the gap inside the mask. The histogram of gray values covered by an approximate body mask typically showed two peaks belonging to adipose and nonadipose areas. The refined threshold (fat threshold) to separate the two tissues was calculated on the basis of the iterative clustering algorithm of Ridler and Calvard.29 The same threshold was also used in manual segmentation on the corresponding slice as described in the previous section. Once the fat threshold was determined, the TAAT mask (Figure 2a) was simply obtained by binarizing the image covered by an approximate body mask at the threshold.
Figure 1 Flow chart that depicts automated SAAT and VAAT segmentation. Four subprocedures would generate TAAT, nonfat, visceral, and SAAT masks. More details are described in the Automated Image Analysis section.
environment for analysis of MR images. The segmentation procedure consisted of four subprocedures as summarized in Figure 1. Each subprocedure corresponded to the formation of one of the body masks: TAAT mask, nonfat mask, visceral mask, and SAAT mask. The subprocedures are described below. They were performed slice by slice in multisliced MR International Journal of Obesity
Formation of nonfat mask. Approximate body masks do not coincide accurately with body contours, and the outer edges of a TAAT mask that represent body surface curves are not always contiguous, for various reasons such as imaging artefacts and deep skin folds common in obese subjects, as shown in the Results section. In order to have a body mask that better defines the outermost body contour on an image slice, the outermost edge pixels of the TAAT mask were first identified and presented using polar coordinates. The distance coordinates were then fitted against their angle coordinates as a polynomial of degree 8. The central distance difference for each edge pixel was also calculated. Those edge pixels that had differences larger than half of the maximal difference and smaller than half of the minimal difference were replaced by the fitted ones. This operation is equivalent to a selective smoothing of the edge. A refined body mask was then obtained by ‘shrinking’ the approximate body mask according to the smoothed contiguous edge. A Euclidean distance map for the refined body mask was generated by the morphological distance transformation.28 A ‘core’ nonfat mask was the collection of pixels that were in
Automated large-scale assessment of abdominal fat T-H Liou et al
847 ellipse fitting. The two initial visceral masks were then combined to form the visceral mask (Figure 2c) based on abdominal anatomy. In general, the abdominal side of the first initial visceral mask and most of the backside of the second initial visceral mask were taken into the visceral mask.
Figure 2 An example of the four masks described in Automated Image Analysis. The masks are painted in white and laid on the same MR slice. The MR pulse sequence was FRFSE. (a) TAAT mask, (b) nonfat mask, (c) visceral mask, and (d) SAAT mask. The Automated Image Analysis section has more details about these masks.
the refined body mask, with a distance larger than 10, and not in the TAAT mask. To obtain the nonfat mask (Figure 2b), conditional dilatation28 was applied on the ‘core’ nonfat mask to get rid of small nonfat areas outside the visceral area.
Formation of visceral mask. For the visceral mask, two initial visceral masks were generated by different approaches on the basis of anatomical knowledge of the abdomen. We assumed that the anterior boundaries that ran along the inner and outer sides of the SAAT on an axial MR slice were simple smooth curves convex ventrally except for the umbilical discontinuity on some slices. The posterior half of the inner border of the SAAT consisted of three parts: central paraspinal and bilateral posterolateral flank parts. The central part was relatively smooth but could be convex, either slightly dorsally or ventrally, with a relatively large curvature. The posterolateral parts might not be very smooth and might be relatively sharply concave inward on some image slices. On the other slices, the curve was smoothly convex outward with possible s-shaped transition zones. The first initial visceral mask was obtained in a somewhat similar way to the generation of the refined body mask, only this time the edge pixels were fitted as an ellipse. Missing edges and inappropriate edge pixels were replaced by the fitted ones with exceptions according to the previously mentioned anatomical knowledge. The so-called inappropriate edge pixels were determined as two standard deviations outside the mean edge distance. Morphological ‘open’ and ‘close’, to remove small dots outside a nonfat mask and to fill in the gap inside, respectively, were the operations used to obtain the second initial visceral mask. The kernel size for the two operations was adjusted according to body size obtained in
Formation of SAAT mask. The visceral mask obtained in the previous subprocedure was still an approximate mask. Parallel to the refining procedure, as described in the last two subsections, a Euclidean distance map and elliptical edge fitting were applied to refine the visceral mask. The collection of pixels in the refined visceral mask with distance larger than 32 and also in the nonfat mask made a core visceral mask. The SAAT mask (Figure 2d) was obtained by first selecting the pixels that were in the TAAT mask but not in the VAAT edges or ‘core’ visceral mask. Then again, conditional dilatation was applied to get rid of unwanted small VAAT areas outside of the SAAT area. For each image slice, the TAAT and SAAT areas were just the mask areas that bore the same names. The VAAT area equaled the SAAT area subtracted from the TAAT area.
Manual image analysis The DICOM data from the MR imaging sequences were also manually analyzed with a public-domain image processing program written in Java (ImageJ, release version 1.30v, 3 July 2003). For each axial MR slice, a ROI that roughly enclosed the whole body part was manually drawn, as shown in Figure 3a. The separation of adipose and nonadipose areas inside the ROI was performed by thresholding. The thresholds were supplied manually according to the results obtained in subsection formation of TAAT Mask in the Automated Image Analysis section. When the adipose tissue had been defined (Figure 3b), the TAAT area in mm2 on each slice could be calculated by ImageJ (Figure 3d). Another ROI was then drawn to include VAAT only (Figure 3c) and the VAAT area was estimated accordingly (Figure 3d). Like many other researchers listed in Table 1, we did not edit further inside VAAT and SAAT. Therefore, high-intensity nonfat pixels arising from bone marrow and fatty intestinal contents belong to VAAT. The VAAT area was then subtracted from the TAAT area to obtain the area of SAAT for each axial slice. Once all the slice areas of the TAAT, VAAT, and SAAT were calculated, the volumes of these were obtained simply by the summation of the slice areas and change of units. The adipose tissue volumes were presented in milliliters. Three graduate students were trained and supervised by a physician to perform these tasks. Coefficients of variation, defined as the standard deviation of a distribution divided by its mean, were calculated for inter- and intrareproducibility. International Journal of Obesity
Automated large-scale assessment of abdominal fat T-H Liou et al
848
Figure 3 Sample results of the manual technique. The image processed is the same as the one presented in Figure 2. The ROIs are drawn in yellow. (a) An ROI roughly covers the whole abdominal area of a cross-section. This could exclude body structures other than the abdomen, such as arms. (b) After thresholding. The threshold of the adipose tissue for each slice is supplied by the automated technique. The justification of using the same threshold is stated in the text. The TAAT area is calculated by ImageJ as shown in (d), first line. The unit is mm2. (c) The ROI for visceral area. Note that there is no further editing to exclude pixels such as nonfat bone marrow or intermuscular adipose tissue. The VAAT area is calculated by ImageJ at this step as shown in (d), second line. (d) Areas of TAAT and VAAT. The SAAT is just the difference between TAAT and VAAT.
successfully overcame many difficulties encountered in automatically identifying SAAT and VAAT using MR images. They included, but not exhaustively, heterogeneous image gray level due to heterogeneous MR field, significant variety of the shape and size of VAAT and SAAT, either in the same subject or different subjects, respiratory motion artefacts, and reconstruction aliasing. Since some subjects missed some scans and/or sequences, there were 4290 images in total that were analyzed with both the manual and the proposed fully automatic methods. Example results of the automatic segmentation on different subjects with different MR pulse sequences are given in Figures 4–7. White contour lines enclose the SAATs in Figures 4 and 6 and the VAATs on Figures 5 and 7. Results in Figures 4 and 5 show that, even in normal conditions, the anatomical variety of subjects under consideration could be larger than one might think. However, in spite of different brightness of the adipose tissues, different thicknesses and shapes of SAAT and VAAT, the umbilical discontinuities that disrupted SAAT, and isolated low-signal areas inside SAAT, the proposed algorithm had performed the segmentation correctly. The segmentation results for the so-called unusual cases are depicted in Figures 6 and 7. Although small segments of either VAAT or SAAT were erroneously assigned or not assigned, respectively, as SAAT, the proposed method handled conditions of unusual abdominal anatomies, imaging artefact caused by respiratory movements, and
Algorithm evaluation All the segmented results were visually inspected slice by slice by a physician to check the appropriateness of the automatic segmentation. The volumes of SAAT and VAAT measured by manual image analysis were treated as the true ones. They were compared with those obtained automatically. The accuracy of the proposed method was then evaluated by three errors: mean percentage error, root mean squared error (RMSE), and percentage root mean squared error (%RMSE), and Bland–Altman plots.30 Note that letting the thresholds be the same in both manual and automated fat segmentation methods focuses the validation on the separation of SAAT and VAAT.
Results Figure 4 Some results of the automatic segmentation in ‘typical’ cases. The
It took approximately 7 min and 15 min to complete the 15-slice volume estimation of the three adipose tissue classes using automated and manual methods, respectively. The coefficients of variation for inter- and intrareproducibility of the manual technique are from 0.009 to 0.026 and from 0.006 to 0.022, respectively. Since the errors are about the same for both SAAT and VAAT, and SAATs are usually more than VAATs, the coefficients for manual evaluation of VAAT are usually larger than those for SAAT. The proposed method International Journal of Obesity
SAATs are enclosed by white contour lines. The MR pulse sequences were STIR, SE, FLAIR, and STIR for subfigures (a)–(d), respectively. (a) The thickness of SAAT could be small compared with the abdominal diameters and arms can be seen in the field of view. (b) The backside inner boundary of SAAT was ‘spiky’ and the cross-section of an abdomen could look ‘flat’. (c) There was an umbilical discontinuity that made the inner and outer boundaries of the SAAT one. The thickness of SAAT could vary significantly with different locations on the same image slice. The cross-section of an abdomen could be like an inverted ‘heart’. (d) There might be low-signal areas inside SAAT. These were real nonadipose connective tissues such as fascia and not one of the imaging artefacts.
Automated large-scale assessment of abdominal fat T-H Liou et al
849
Figure 5 The same results of the automatic segmentation on ‘typical’ cases as in Figure 4. However, instead of SAAT, it is the VAATs that are enclosed by white contour lines.
Figure 6 Some results of the automatic segmentation in ‘unusual’ cases. The SAATs are enclosed by white contour lines. The MR pulse sequences were SE, FLAIR, SE, and FRFSE for subfigures (a)–(d), respectively. Small segments of VAAT in the left-lower quadrant were erroneously included in SAAT in subfigures (a)–(c). (a) The image has no left–right symmetry and almost no distinction between SAAT and VAAT over some parts of the anterior abdomen. (b) Part of a ‘halo’ can be seen above the anterior abdomen, which was considered due to an imaging artefact caused by respiratory movements. The artefact not only blurred the true borders of the anterior SAAT and made the thickness of it uneven, but also created a false border that passed through it. (c) There is a large ‘island’ of SAAT completely separated from the main SAAT because of the deep thick folds of the skin and adipose tissues. In addition, for the same reason, the boundary contours of the SAATs have very sharp turns. (d) Small parts of the SAAT to the left were not included because of lower brightness caused by a heterogeneous magnetic field during MR imaging.
heterogeneous magnetic fields with satisfying accuracy. The figure legend has more details. In addition to examples that show some of the qualitative results in the figures, quantitative results are summarized in
Figure 7 The same results of the automatic segmentation in ‘unusual’ cases as in Figure 6. However, instead of SAAT, it is the VAATs that are enclosed by white contour lines.
Table 3. The mean volumes (7standard deviation, s.d.) of SAAT obtained by automated analysis from different MR pulse sequences ranged from 3772.9 (71065.5) to 3852.2 (71105.0) ml for FLAIR and FRFSE, respectively. Those of VAAT were from 1704.4 (7688.7) to 1908.3 (7678.7) for STIR and FRFSE, respectively. Both the mean volumes (7s.d.) of SAAT and VAAT obtained by the manual technique are also listed in Table 3 for a side-by-side comparison. They are extremely close to those obtained by the automated method. The Pearson’s correlation coefficients between the two methods ranged from 0.998 to 1.000 for different adipose tissue locations and MR pulse sequences. Percentage error was defined as the difference between the volumes obtained by the manual and automatic image analysis methods divided by the volume obtained by the manual one. They ranged, for SAAT of all subjects, from 3.2% (STIR) to 7.6% (FRFSE) with mean (7s.d.) for each MR pulse sequence 0.9% (70.5%) for FLAIR, –0.2% (71.4%) for SE, –0.7% (70.6%) for STIR, and 0.0% (71.1%) for FRFSE. For VAAT, they ranged from –13.1% (SE) to 10.2% (STIR) with mean (7s.d.) for each MR pulse sequence 2.1% (71.6%) for FLAIR, 0.7% (72.9%) for SE, 2.0% (72.1%) for STIR, and 0.1% (71.6%) for FRFSE. The positive and negative percentage errors canceled out so that the means were close to zero. For better assessment of the performance of the proposed algorithm, RMSE was used. It is defined as the square root of the average over the squared differences between the volumes obtained by the manual and automatic analysis methods. Its unit is ml. The %RMSE was calculated as RMSE divided by the corresponding mean volume obtained by the manual method. RMSEs of SAAT are 38.1 (FLAIR), 47.1 (SE), 38.7 (STIR), and 41.1 (FRFSE). Those of VAAT are 37.4 (FLAIR), 47.4 (SE), 39.5 (STIR), and 32.8 (FRFSE). Percentage RMSEs of SAAT are 1.0% (FLAIR), 1.2% (SE), 1.0% (STIR), and International Journal of Obesity
Automated large-scale assessment of abdominal fat T-H Liou et al
850 Table 3 Sequence
The adipose tissue volume estimation and its errors Location
Manual
Automated
%Error
%Error
RMSE
%RMSE
Mean (ml)
s.d. (ml)
Mean (ml)
s.d. (ml)
Mean (%)
s.d.(%)
(ml)
(%)
r
FLAIR
Subcutaneous Visceral
3805.9 1795.9
1072.6 695.3
3772.9 1827.3
1065.5 696.8
0.9 2.1
0.5 1.6
38.1 37.4
1.0 2.1
1.000 1.000
SE
Subcutaneous Visceral
3786.8 1787.2
1085.7 684.7
3775.3 1795.3
1070.3 677.3
0.2 0.7
1.4 2.9
47.1 47.4
1.2 2.7
0.999 0.998
STIR
Subcutaneous Visceral
3829.1 1675.6
1106.2 683.8
3799.7 1704.4
1094.6 688.7
0.7 2.0
0.6 2.1
38.7 39.5
1.0 2.4
1.000 0.999
FRFSE
Subcutaneous Visceral
3853.8 1905.4
1114.4 680.1
3852.2 1908.3
1105.0 678.7
0.0 0.1
1.1 1.6
41.1 32.8
1.1 1.7
0.999 0.999
s.d. ¼ standard deviation; %Error mean ¼ mean percentage error; %Error s.d. ¼ standard deviation of the percentage error; RMSE ¼ root mean squared error; %RMSE ¼ percentage root mean squared error; r ¼ Pearson’s correlation coefficient between manual and automated techniques for each MR sequence. FLAIR ¼ T1weighted fluid attenuated inversion recovery; SE ¼ T1-weighted spin echo; STIR ¼ short inversion time inversion recovery; and FRFSE ¼ T2-weighted fast recovery fast spin-echo.
Figure 8 Bland–Altman plot depicting the difference between the manual
Figure 9 Bland–Altman plot depicting the difference between the manual
and automated estimation of visceral abdominal adipose tissue (VAAT) in the T1 STIR sequence. The solid line indicates the mean differences between the two methods and the dotted lines mark two standard deviations away from the mean differences.
and automated estimation of VAAT in the T1 SE sequence. The solid and dotted lines represent the mean differences and the standard deviations, respectively, as in Figure 8.
1.1% (FRFSE). Those of VAAT are 2.1% (FLAIR), 2.7% (SE), 2.4% (STIR), and 1.7% (FRFSE). Figures 8 and 9 are Bland– Altman plots for the comparison of VAATs estimated by manual and automated techniques. MR sequences STIR (Figure 8) and SE (Figure 9) were chosen because they are the two sequences with the largest RMSEs and %RMSEs. Note that in both Figures 8 and 9 the mean difference lines are almost horizontal. This implies that the performance of our algorithm does not change much with the size of VAAT. For manual assessment, the volumes of the three adipose tissues obtained from the four MR pulse sequences were not statistically different. The F values of ANOVA31 on the manual assessment of TAAT, SAAT, and VAAT for different pulse sequences were 0.48, 0.05, and 1.37, respectively, whith P-values of 0.70, 0.99, and 0.25. There were, again, no statistical differences in the squared errors of the proposed automatic assessment of SAAT and VAAT obtained from
different pulse sequences. The F values were 0.32 (P ¼ 0.81) and 1.77 (P ¼ 0.15), respectively. However, for percentage errors, significant differences could be found for both SAAT and VAAT by single-factor ANOVA: F values were 14.54 and 13.77 for SAAT and VAAT, respectively (both Po0.001). Further Tukey tests32 revealed that, for percentage errors of both VAAT and SAAT, the four MR pulse sequences could be divided into two groups. Pulse sequences STIR and FLAIR were in one group (IR group), and SE and FRFSE (SE group) were in another. There were statistically significant differences in mean percentage errors between groups but not within each group. This can also be seen by comparison of Figures 8 and 9. The mean difference line in Figure 9 (SE) is higher than that in Figure 8 (STIR). This implies that our technique tends to slightly but statistically significantly overestimate VAAT in T1 STIR compared with T1 SE.
International Journal of Obesity
Automated large-scale assessment of abdominal fat T-H Liou et al
851
Discussion Most of the previous studies have defined VAAT within a range from 5 cm below the lumbar spine (L4–5) to the image slice corresponding to the superior border of the liver. The dominant part of the VAAT would reside between the levels L2 and S1. Since determining spinal levels was not convenient in an ordinary clinical MR setting, we imaged VAAT by 15 contiguous slices, 7.5 cm above and below the umbilicus (L4–5) level, which covered approximately the L2–S2 levels. Although the amount of VAAT measured in this study was not the total VAAT volume, it should be close to and have high correlation with the total amount of VAAT. Another issue is that the small amount of intermuscular and paravertebral adipose tissues may influence the accuracy of the measurement of VAAT. However, it is still controversial whether such compartments of adipose tissue should be included in VAAT. As in many other studies, we included these adipose tissues in VAAT. Since the automated segmentation method we present is aimed at the analysis of large amounts of image data such as in epidemiological studies, the simplified adipose tissue classification and ignoring small numbers of high-intensity nonfat pixels could be justified. The fact that statistical tests did not show significant differences in the volumes of various adipose tissues obtained from different MR pulse sequences by manual image analysis supports our hypothesis that large-scale epidemiological studies related to the distribution of abdominal adipose tissues could be performed on MRI data not originally collected for assessing adipose tissues. Although Tukey tests show that the performance of the automated technique is statistically significantly different in SE and IR groups if evaluated by percentage error, this does not mean the SE group has less error than the IR group. Since there is no statistically significant difference in squared error between the two groups, our proposed fully automatic method worked equally precisely but not equally accurately for the two different pulse sequence groups. However, the results of the automated method were less biased when applied to sequences SE and FRFSE than to sequences FLAIR and STIR. As shown in Table 1, most previous studies of the assessment of abdominal adipose tissue were not really automatic. Recently, Positano et al.33 proposed a method for unsupervised assessment of VAAT by MRI. The fuzzy c-mean approach and an active contour algorithm were applied to identify and estimate the volumes of VAAT and SAAT. Besides our proposed method, this is the only published method with full automation that we know of. However, they provided no discussion or description of how their algorithm could handle the ‘nonstandard’ cases with umbilical discontinuity, low-signal areas inside SAAT, and fragmented SAATs such as those in Figures 4–7. Although it is possible, if appropriate parameters are set, for the active contour algorithm to separate VAAT from SAAT when there is no
clear boundary between parts of the two components or to go around sharp turns, we suspect that it would be difficult to set the parameters for general conditions with significant anatomical and other variations. In addition, overestimation of SAAT and underestimation of VAAT were reported with percentage error means (7s.d.) of 6.4% (75.8%) and –7.9% (716.4%). Although the pulse sequences used in their and our studies were different, our proposed method seems to be more accurate and less biased. Some of our parameters, such as the initial thresholds in the formation of the TAAT mask, the degree of fitting a polynomial in the formation of the nonfat mask, the criteria for the distance between the visceral core and the visceral edge in the formation of SAAT mask, and the kernel size for ‘open’ and ‘close’ in the formation of various masks, were somewhat empirical. However, judging from the results and the number of images processed, they were quite robust and worked in many difficult cases. For further improvement of the current algorithm, normalization or standardization both for a heterogeneous magnetic field and anatomical variety should be considered. Knowledge-based gap filling and edge modification in the formation of several masks actually served as constraints in an active contour algorithm, but with more flexibility. In conclusion, we have shown that the proposed automated analysis of visceral and subcutaneous abdominal fat volumes using axial MR images is sufficiently rapid, reliable, and accurate to be a surrogate measure to replace laborious manual measurement. This technique may serve as a methodological basis for many clinical and scientific applications, particularly those dealing with a large amount of data such as epidemiological studies.
Acknowledgements We thank J-R Liao and the staff in Department of Radiology, Taipei Medical University-Wan Fang Hospital, for their work on MR image acquisition. We also thank Y-F Tsai, Y-W Sun, and L-M Chen for their assistance in performing manual image analysis. This research is partly supported by Grant NSC 93-2213-E-038-005 from the National Science Council, Taiwan.
References 1 Caprio S. Relationship between abdominal visceral fat and metabolic risk factors in obese adolescents. Am J Hum Biol 1999; 11: 259–266. 2 Gastaldelli A, Miyazaki Y, Pettiti M, Matsuda M, Mahankali S, Santini E et al. Metabolic effects of visceral fat accumulation in type 2 diabetes. J Clin Endocrinol Metab 2002; 87: 5098–5103. 3 Smith S, Lovejoy J, Greenway F, Ryan D, deJonge L, de la Bretonne J et al. Contributions of total body fat, abdominal subcutaneous adipose tissue compartments, and visceral adipose tissue to the
International Journal of Obesity
Automated large-scale assessment of abdominal fat T-H Liou et al
852 4
5
6
7
8
9 10
11
12
13
14
15
16
17
metabolic complications of obesity. Metabolism 2001; 50: 425– 435. Wajchenberg BL. Subcutaneous and visceral adipose tissue: their relation to the metabolic syndrome. Endocrine Rev 2000; 21: 697– 738. Miyazaki Y, Glass L, Triplitt C, Wajcberg E, Mandarino LJ, DeFronzo RA. Abdominal fat distribution and peripheral and hepatic insulin resistance in type 2 diabetes mellitus. Am J Physiol Endocrinol Metab 2002; 283: E1135–E1143. Thomas E, Bell J. Influence of undersampling on magnetic resonance imaging measurements of intra-abdominal adipose tissue. Int J Obes Relat Metab Disord 2003; 27: 211–218. Armellini F, Zamboni M, Rigo L, Todesco T, Bergamo-Andreis IA, Procacci C et al. The contribution of sonography to the measurement of intra-abdominal fat. J Clin Ultrasound 1990; 18: 563–567. Liu K, Chan Y, Chan W, Kong W, Kong M, Chan J. Sonographic measurement of mesenteric fat thickness is a good correlate with cardiovascular risk factors: comparison with subcutaneous and preperitoneal fat thickness, magnetic resonance imaging and anthropometric indexes. Int J Obes Relat Metab Disord 2003; 27: 1267–1273. Campbell LV, Carey DG, Chisholm DJ. Measurement of central adiposity. A bet each way? Diabetes Care 1996; 19: 1033–1034. Kamel E, McNeill G, Han T, Smith FW, Avenell A, Davidson L et al. Measurement of abdominal fat by magnetic resonance imaging, dual-energy X-ray absorptiometry and anthropometry in nonobese men and women. Int J Obes Relat Metab Disord 1999; 23: 686–692. Snijder M, Visser M, Dekker J, Seidell JC, Fuerst T, Tylavsky F et al. The prediction of visceral fat by dual-energy X-ray absorptiometry in the elderly: a comparison with computed tomography and anthropometry. Int J Obes Relat Metab Disord 2002; 26: 984–993. Greenfield JR, Samaras K, Chisholm DJ, Campbell LV. Regional intra-subject variability in abdominal adiposity limits usefulness of computed tomography. Obes Res 2002; 10: 260–265. Kobayashi J, Tadokoro N, Watanabe M, Shinomiya M. A novel method of measuring intra-abdominal fat volume using helical computed tomography. Int J Obes Relat Metab Disord 2002; 26: 398–402. Seidell JC, Bakker CJ, van der Kooy K. Imaging techniques for measuring adipose-tissue distribution – a comparison between computed tomography and 1.5-T magnetic resonance. Am J Clin Nutr 1990; 51: 953–957. Elbers J, Haumann G, Asscheman H, Seidell J, Gooren L. Reproducibility of fat area measurements in young, non-obese subjects by computerized analysis of magnetic resonance images. Int J Obes Relat Metab Disord 1997; 21: 1121–1129. Han T, Kelly I, Walsh K, Greene R, Lean M. Relationship between volumes and areas from single transverse scans of intraabdominal fat measured by magnetic resonance imaging. Int J Obes Relat Metab Disord 1997; 21: 1161–1166. Gronemeyer S, Steen R, Kauffman W, Reddick W, Glass J. Fast adipose tissue (FAT) assessment by MRI. Magn Reson Imag 2000; 18: 815–818.
International Journal of Obesity
18 Yang GZ, Myerson S, Chabat F, Pennell DJ, Firmin EN. Automatic MRI adipose tissue mapping using overlapping mosaics. Magn Reson Materials Phys Biol Med 2002; 14: 39–44. 19 Poll LW, Wittsack HJ, Koch JA, Willers R, Scherer A, Kapitza C et al. Quantification of total abdominal fat volumes using magnetic resonance imaging. Eur J Med Res 2002; 7: 347–352. 20 Donnelly LF, O’Brien KJ, Dardzinski BJ, Poe SA, Bean JA, Holland SK et al. Using a phantom to compare MR techniques for determining the ratio of intraabdominal to subcutaneous adipose tissue. Am J Roentgenol 2003; 180: 993–998. 21 Ross R, Shaw KD, Martel Y, Guise JD, Avruch L. Adipose tissue distribution measured by magnetic resonance imaging in obese women. Am J Clin Nutr 1993; 57: 470–475. 22 Ross R, Shaw KD, Rissanen J, Martel Y, Guise JD, Avruch L. Sex differences in lean and adipose tissue distribution by magnetic resonance imaging: anthropometric relationships. Am J Clin Nutr 1994; 59: 1277–1285. 23 Ross R, Goodpaster B, Kelley D, Boada F. Magnetic resonance imaging in human body composition research. From quantitative to qualitative tissue measurement. Ann NY Acad Sci 2000; 904: 12–17. 24 Lancaster JL, Ghiatas AA, Alyassin A, Kilcoyne RF, Bonora E, DeFronzo RA. Measurement of abdominal fat with T1-weighted MR images. JMRI 1991; 1: 363–369. 25 Pichiecchio A, Uggetti C, Egitto MG, Berardinelli A, Orcesi S, Gorni KO et al. Quantitative MR evaluation of body composition in patients with Duchenne muscular dystrophy. Eur Radiol. 2002; 12 (11): 2704–2709. 26 Shen W, Wang Z, Punyanita M, Lei J, Sinav A, Kral JG et al. Adipose tissue quantification by imaging methods: a proposed classification. Obes Res 2003; 11: 5–16. 27 Johannes B, Jurgen R, Thomas T. Differentiation of normal and pathologic brain structures in MRI using exact T1 and T2 values followed by a multidimensional cluster analysis. In: Greenes R, Peterson H, Protti D (eds). Medinfo 95 – Proceedings of the 8th World Congress on Medical Informatics. Amsterdam: North Holland, 1995. pp 395–399. 28 Mathematical morphology. In: Haralick RM, Shapiro LG (eds). Computer and robot vision vol. I. Addison-Wesley: New York, 1992. pp 157–261. 29 Ridler TW, Calvard S. Picture thresholding using an iterative selection method. IEEE Transactions on Systems, Man and Cybernetics 1978; (SMC-8): 630–632. 30 Bland JM, Altman DG. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet 1986; 1: 307–310. 31 Multiple hypotheses: The analysis of variance. In: Zar JH (ed). Biostatistical Analysis 2nd edn. Prentice-Hall: Englewood Cliffs, NJ, 1984. pp 162–184. 32 Multiple comparisons. In: Zar JH (ed). Biostatistical Analysis, 2nd edn. Prentice-Hall: Englewood Cliffs, NJ, 1984. pp. 185–205. 33 Positano V, Gastaldelli A, Sironi A, Santarelli M, Lombardi M, Landini L. An accurate and robust method for unsupervised assessment of abdominal fat by MRI. J Magn Reson Imag 2004; 20: 684–689.
