Comparison of Methods for Assessing Abdominal Adipose Tissue from Magnetic Resonance Images

Transcription

1 Comparison of Methods for Assessing Abdominal Adipose Tissue from Magnetic Resonance Images Hassan Arif,* Susan B. Racette,* Dennis T. Villareal,* John O. Holloszy,* and Edward P. Weiss* Abstract ARIF, HASSAN, SUSAN B. RACETTE, DENNIS T. VILLAREAL, JOHN O. HOLLOSZY, AND EDWARD P. WEISS. Comparison of methods for assessing abdominal adipose tissue from magnetic resonance images. Obesity. 2007;15: Objective: To compare the inter-rater and intra-rater reliability and analysis time of two methods for quantifying visceral adipose tissue (VAT) and subcutaneous adipose tissue (SAT) volumes from magnetic resonance (MR) images. Research Methods and Procedures: Ten subjects (BMI, kg/m 2 ; 56 years of age 4 years) underwent MR imaging of the abdomen. Ten transverse T1-weighted images were selected from each scan and analyzed using two software packages that differ in principle. The first method, ANALYZE version 5.0, represents the manual threshold method, and the second, HIPPO version 1.3, is based on the fuzzy clustering approach. Inter-rater reliability for each method was assessed by comparing the intraclass correlation coefficients (ICCs) for VAT and SAT results from two evaluators, and intra-rater reliability for each method was assessed by comparing ICCs for VAT and SAT analyses performed 1 week apart by the same evaluator. The total time for analysis also was compared between methods. Received for review December 8, Accepted in final form February 12, The costs of publication of this article were defrayed, in part, by the payment of page charges. This article must, therefore, be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact. *Division of Geriatrics and Nutritional Sciences, Department of Internal Medicine, Washington University School of Medicine, St. Louis, Missouri; and Department of Nutrition and Dietetics, St. Louis University, St. Louis, Missouri. Present address: Department of Nutrition and Dietetics, St. Louis University, St. Louis, Missouri. Address correspondence to Edward P. Weiss, Department of Nutrition and Dietetics, St. Louis University, 3437 Caroline Street, Room 3076, St. Louis, MO eweiss4@slu.edu Copyright 2007 NAASO Results: The inter-rater reliability for VAT was greater with HIPPO than with ANALYZE (ICC vs ), whereas inter-rater reliability for SAT did not differ between methods (ICC and 0.987). The intra-rater reliability was equally high with HIPPO and ANALYZE for both VAT (ICC vs ) and SAT (ICC vs ). HIPPO required less than one-half as much analysis time as ANALYZE ( vs minutes, p ). Discussion: HIPPO software appears advantageous for the quantification of VAT from multislice MR images because inter-rater results are more reliable, and it is more timeefficient than less automated methods. Key words: abdominal obesity, adipose, adipose tissue, body fat distribution, central obesity Introduction Abdominal visceral adipose tissue (VAT) 1 has been implicated as a major risk factor for several diseases, including type 2 diabetes, coronary heart disease, hypertension, and stroke (1 5). The methods used for the measurement of abdominal VAT and subcutaneous adipose tissue (SAT) volumes include computed tomography (CT) and magnetic resonance (MR) imaging (6). MR imaging has advantages over CT in that it can be used to measure VAT and SAT without ionizing radiation and, therefore, can be used for multislice volume imaging at several time-points during the course of a study. In addition, the signal intensity of fat from CT scans is very low and is similar to that of air and bowel gas. The bright signal intensity of fat on T1-weighted MR images and the absence of signals from air and bowel gas permit accurate distinction among fat, other tissues, and gas. 1 Nonstandard abbreviations: VAT, visceral adipose tissue; SAT, subcutaneous adipose tissue; CT, computed tomography; MR, magnetic resonance; TT, total time; ICC, intra-class correlation coefficient; CI, confidence interval OBESITY Vol. 15 No. 9 September 2007

2 In general, most software packages for quantifying VAT and SAT produce similar results. However, the reproducibility of image analysis is sometimes less than optimal, especially among different evaluators. One technique that improves reproducibility is to quantify VAT and SAT in multiple slices rather than one slice (7,8). However, a costly consequence of analyzing multiple slices is the large time requirement. The software programs for analysis of MR images have improved in accuracy and efficiency in the last decade. A highly automated program, HIPPO (version 1.3; Pisa, Italy), was developed recently for quantifying VAT and SAT from MR images. This method has been established as valid and has the advantage of being more time-efficient than other analytic approaches (9,10). A possible additional benefit of using this more automated analytic approach is that interrater reliability might be better than with less automated approaches; however, this has not been previously assessed. The purpose of the present study was to test the hypothesis that inter-rater and intra-rater reliability for VAT and SAT volumes, measured from MR images, are greater when using HIPPO than they are when using other software. We also compared the HIPPO-derived VAT and SAT volumes with values derived using another, less automated software package. Lastly, we quantified the image analysis time for both methods to determine whether HIPPO is a more timeefficient method than the other method. ANALYZE Direct 5.0 software (Mayo Clinic, Rochester, MN) is MR image analysis software that is based on the same principles as most other MR image analysis programs. We used ANA- LYZE in the present study to represent other, less automated software. Research Methods and Procedures MR Image Acquisition Proton MR images were obtained on 10 healthy, nonobese women (n 7) and men (n 3) (age, 56 4 years; range, 52 to 61 years; weight, kg; range, 67.7 to 96.6 kg; BMI, kg/m 2 ; range, 23.3 to 29.7 kg/m 2 ) who were part of a larger study (11). Ten 10-mm-thick axial images with no intersection gap were acquired for each subject using a 1.5-T superconducting magnet (Siemens, Iselin, NJ) and a T1-weighted pulse sequence. The first slice was at the level of the inferior border of the more inferior kidney, which typically coincides with the L4 to L5 interspace, and the remaining slices were acquired superiorly. A multislice spin-echo inversion-recovery pulse sequence was used and set to optimize bright (adipose tissue) vs. dark (aqueous signal) contrast from adjacent tissue (repetition time, 550 ms; echo time, 12 ms; field of view, mm). The images were taken with 134 phase-encoding steps to form images that were stored in a 16-bit format. Image Analysis Two copies of the computerized MR images were made, coded, and subsequently analyzed by two different evaluators (A and B). Each scan was analyzed using both HIPPO and ANALYZE at two time-points 1 week apart. MR images consist of pixels (a pixel is the smallest part of a digital image). Each analysis program has user-friendly tools for assignment of the pixels according to their intensity, with light pixels representing fat and dark pixels representing non-fat. The pixel height, pixel width, and slice thickness are multiplied to determine the volume of each voxel (i.e., volume pixel, the smallest distinguishable boxshaped part of a three-dimensional digital image). Fat volume in the region of interest is calculated as the product of the number of fat voxels and the voxel volume. ANALYZE (Analyze Direct, version 5.0) is capable of calculating volumes and surface areas of different tissues of the body, including brain, muscle, and fat. For VAT and SAT analyses, the evaluator manually designates areas deemed to be visceral or subcutaneous fat by placing a seed in the appropriate location and then moving a slider to highlight the areas corresponding to a specified pixel intensity. The ANALYZE program sums the values for all of the slices selected and expresses the result as mean slice surface area or volume for both VAT and SAT. HIPPO (version 1.3; Pisa, Italy) (9) was specifically designed for the measurement of abdominal adipose tissue and was custom-written in the IDL 6.0 base. The image files are loaded into the HIPPO software and are then analyzed using the segment slice function. This software function initiates the automatic segmentation of the abdominal image into fat and non-fat components by designating each pixel as fat or non-fat. The principle used by HIPPO was originally described by Lancaster et al. (12). In brief, HIPPO is based on the fuzzy c-mean approach (i.e., fuzzy logic) (13,14) and is able to make unsupervised classification of data into a number of clusters by identifying different tissues in an image without the use of an explicit threshold. HIPPO then automatically divides the pixels into light (fat) and dark (non-fat) and expresses them as a twin-peak histogram. The area under the peak representing fat is expressed as a volume (centimeters cubed). The output result is in the form of a computer spreadsheet. Total time (TT) for image analysis was quantified as the time from loading the files into the software to the generation of the output files with the calculated VAT and SAT volume data. TT was compared between the ANALYZE and HIPPO methods. Statistical Analyses Intra-class correlation coefficients (ICCs) and associated 95% confidence intervals (CIs) for each method were used to assess inter- and intra-rater reliability for VAT and SAT volumes. Inter-rater reliability measures were based on the OBESITY Vol. 15 No. 9 September

3 Table 1. Inter-rater reliability for quantification of abdominal VAT and SAT volumes using two different analytical methods Method Variable N ICC (95% CI) ANALYZE VAT (0.477, 0.954) SAT (0.950, 0.997) HIPPO VAT (0.986, 0.999) SAT (0.908, 0.994) Data are ICCs with 95% CIs in parentheses. ANALYZE, MR images analyzed using ANALYZE Direct, version 5.0 software; HIPPO, MR images analyzed using HIPPO version 1.3 software. first analysis by each evaluator for each method. Intra-rater ICCs were based on analyses performed 1 week apart by the same evaluator. Between-method comparisons were performed with Deming regression and Bland-Altman plots using the first reading of each evaluator as an independent observation. Deming regression assesses the linearity between methods, with statistical significance demonstrated if the CI for the slope does not contain 1 and if the CI for the intercept does not include 0. Bland-Altman plots are presented for visual representation of the agreement between methods. Independent Student s t tests were used to compare the TT for image analysis for ANALYZE and HIPPO. Statistical tests were two-tailed, with significance accepted at p All statistical analyses were performed using SAS software, version of the SAS System for Linux (SAS Institute, Inc., Cary, NC) and MedCalc for Windows, version (MedCalc Software, Mariakerke, Belgium). Results Inter-rater Reliability Inter-rater reliability coefficients for quantification of VAT and SAT by ANALYZE and HIPPO are shown in Table 1. The inter-rater ICC for VAT with ANALYZE was satisfactory; however, the ICC for VAT analysis with HIPPO was better, with only minimal overlap in the 95% CIs between the two methods. Inter-rater reliability for SAT was excellent for both ANALYZE and HIPPO, and these ICCs were comparable. Intra-rater Reliability ICCs for intra-rater reliability were high for both evaluators A and B using ANALYZE and were comparably high using HIPPO (Table 2). The intra-rater ICCs were similar for the ANALYZE and HIPPO methods for both evaluators. Agreement between Methods Table 3 shows the agreement between the VAT and SAT results acquired using HIPPO and those acquired using Table 2. Intra-rater reliability for quantification of abdominal VAT and SAT volumes using two different analytical methods Evaluator Method Variable N ICC (95% CI) A ANALYZE VAT (0.977, 0.998) SAT (0.942, 0.996) HIPPO VAT (0.984, 0.999) SAT (0.974, 0.998) B ANALYZE VAT (0.968, 0.998) SAT (0.996, 1.000) HIPPO VAT (0.999, 1.000) SAT (0.992, 0.999) Data are ICCs with 95% CIs in parentheses. ANALYZE, MR images analyzed using ANALYZE Direct, version 5.0 software; HIPPO, MR images analyzed using HIPPO version 1.3 software. ANALYZE. For VAT, the mean between-method difference of cm 3 ( %) reflects a bias for somewhat lower VAT results from HIPPO. Likewise, there was a bias toward lower SAT results from HIPPO as reflected by a mean difference of cm 3 ( %) between methods. Results from the Deming regression analysis indicate that the slope and y-intercept for the relationship between VAT results from the two methods were not statistically different from 1 and 0, respectively. Regression analysis of the SAT data indicated that although the y-intercept for the relationship between results from the two methods was not statistically different from 0, there was a tendency for the slope to be less than 1, indicating that HIPPO tends to yield lower SAT values than ANALYZE as SAT volume increases. The agreement between the two methods is also expressed as Bland-Altman and Deming regression plots in Figures 1 and 2. Total Analysis Time The mean analysis time for ANALYZE ( minutes) was more than twice as long (p ) as that for HIPPO ( minutes). Discussion Our study demonstrated that HIPPO, a relatively new and more automated software analysis program, had better interrater reliability for VAT analysis than the more manual ANALYZE method. This is especially important in the context of long-term longitudinal studies because the original evaluator may not be available for follow-up VAT analyses at the end of the study. Intra-rater reproducibility (i.e., repeat analyses done by the same evaluator) was high 2242 OBESITY Vol. 15 No. 9 September 2007

4 Table 3. Agreement between ANALYZE- and HIPPO-determined abdominal VAT and SAT volumes Raw differences (cm 3 ) Differences (%) Regression results Mean SD, volume (cm 3 ) Mean SD 95% LOA Mean SD 95% LOA Slope Intercept VAT ANALYZE, , , HIPPO, (0.75, 1.13) ( 247, 149) SAT ANALYZE, , , HIPPO, (0.74, 1.01) ( 199, 309) Statistics are based on data from the first reading of each scan from both evaluators. Differences are calculated as HIPPO ANALYZE. Percentage differences are calculated as percentage of averages. Slope and intercept data are based on Deming regression analyses using ANALYZE as the independent variable and HIPPO as the dependent variable and are presented as parameter estimates with associated 95% CIs in parentheses. Selected VAT and SAT statistics are presented graphically in Figures 1 and 2, respectively. LOA, limits of agreement (mean difference 1.96 SD). for both ANALYZE and HIPPO, with no difference between these methods. HIPPO required less than half as much analysis time as ANALYZE, which is a significant advantage, especially when multiple slices are analyzed at multiple time-points during a study. Even though our evaluators were trained together to use both analytic techniques, they failed to identify the same VAT regions when using ANALYZE. This discrepancy might reflect the inherent difficulty with visual identification and manual selection of visceral fat in abdominal MR images. On the other hand, the principle used in calculating visceral fat in the fuzzy-logic program, HIPPO, presumes the histogram peak to be symmetrical for both fat and non-fat curves, which may not be the case in subjects with relatively little visceral fat (9). Although HIPPO was biased to produce somewhat lower ( 11%) values for both VAT and SAT than ANALYZE, the Deming regression analyses indicated that abdominal adipose tissue stores quantified using HIPPO and ANA- LYZE were not significantly different for either VAT or SAT. Previous studies have also indicated that HIPPO tends to underestimate VAT (9,10) when compared with less automated methods. However, our finding that SAT tended to be underestimated is not in agreement with findings from previous studies in which HIPPO tended to overestimate SAT (9,10). In summary, the more automated HIPPO software method for quantifying abdominal VAT and SAT volumes from MR images appears to have distinct advantages over less automated methods because there is less variability between different evaluators, and the analysis time is greatly reduced. Acknowledgments We thank the study participants for their cooperation, Karen Steger-May (Division of Biostatistics, Washington Figure 1: (A) VAT volumes presented as a Bland-Altman plot. The solid line represents the mean bias, with 95% CI defined by the dotted lines. (B) VAT Deming regression line (solid line) and the line of identity (x y, dotted line). OBESITY Vol. 15 No. 9 September

5 Figure 2: (A) SAT volumes presented as a Bland-Altman plot. The solid line represents the mean bias, with 95% CI defined by the dotted lines. (B) SAT Deming regression line (solid line) and the line of identity (x y, dotted line). University School of Medicine, St. Louis, MO) for performing the statistical analysis and for assisting in the interpretation of the results, and Morgan Schram for technical support in the handling and analysis of the MR images. This study was supported by the NIH (Cooperative Agreement AG20487 and Clinical Nutrition Research Unit Grant DK56341). E.P.W. was supported by Institutional National Research Service Award NIH AG References 1. Kissebah AH, Freedman DS, Peiris AN. Health risks of obesity. Med Clin North Am. 1989;73: Lapidus L, Bengtsson C, Larsson B, Pennert K, Rybo E, Sjöström L. Distribution of adipose tissue and risk of cardiovascular disease and death: a 12 year follow up of participants in the population study of women in Gothenburg, Sweden. Br Med J (Clin Res Ed). 1984;289: Despres JP, Moorjani S, Ferland M, et al. Adipose tissue distribution and plasma lipoprotein levels in obese women: importance of intra-abdominal fat. Arteriosclerosis. 1989;9: Björntorp P. Portal adipose tissue as a generator of risk factors for cardiovascular disease and diabetes. Arteriosclerosis. 1990;10: Larsson B, Svardsudd K, Welin L, Wilhelmsen L, Björntorp P, Tibblin G. Abdominal adipose tissue distribution, obesity, and risk of cardiovascular disease and death: 13 year follow up of participants in the study of men born in Br Med J (Clin Res Ed). 1984;288: Seidell JC, Bakker CJ, van der KK. Imaging techniques for measuring adipose-tissue distribution: a comparison between computed tomography and 1.5-T magnetic resonance. Am J Clin Nutr. 1990;51: Elbers JM, Haumann G, Asscheman H, Seidell JC, Gooren LJ. 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: Terry JG, Hinson WH, Evans GW, Schreiner PJ, Hagaman AP, Crouse JR III. Evaluation of magnetic resonance imaging for quantification of intraabdominal fat in human beings by spin-echo and inversion-recovery protocols. Am J Clin Nutr. 1995;62: 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: Demerath EW, Ritter KJ, Couch WA, et al. Validity of a new automated software program for visceral adipose tissue estimation. Int J Obes (Lond). 2007;31: Racette SB, Weiss EP, Villareal DT, et al. One year of caloric restriction in humans: feasibility and effects on body composition and abdominal adipose tissue. J Gerontol A Biol Sci Med Sci. 2006;61: Lancaster JL, Ghiatas AA, Alyassin A, Kilcoyne RF, Bonora E, DeFronzo RA. Measurement of abdominal fat with T1-weighted MR images. J Magn Reson Imaging. 1991;1: Bezdek JC, Hall LO, Clarke LP. Review of MR image segmentation techniques using pattern recognition. Med Phys. 1993;20: Udupa JK, Samarasekera S. Fuzzy connectedness and object definition: theory, algorithm and application in image segmentation. Graph Models Image Process. 1996; OBESITY Vol. 15 No. 9 September 2007