Impact of the Analytical Approach on the Reliability of MRI-Based Assessment of Hepatic Fat Content.

MRI is a popular noninvasive method for the assessment of liver fat content. After MRI scan acquisition, there is currently no standardized image analysis procedure for the most accurate estimate of liver fat content. We determined intraindividual reliability of MRI-based liver fat measurement using 10 different MRI slice analysis methods in normal-weight, overweight, and obese individuals who underwent 2 same-day abdominal MRI scans. We also compared the agreement in liver fat content between analytical methods and assessed the variability in fat content across the entire liver. Our results indicate that liver fat content varies across the liver, with some slices averaging 54% lower and others 75% higher fat content than the mean of all slices (gold standard). Our data suggest that the entire liver should be contoured on at least every 10th slice to achieve close agreement with the gold standard.

Introduction

The global prevalence of nonalcoholic fatty liver disease (NAFLD) is estimated to be 25% among the adult population (1, 2). NAFLD is a spectrum of diseases that includes nonalcoholic fatty liver, defined as ≥5% liver fat content, and nonalcoholic steatohepatitis (NASH), which is characterized by excess fat accumulation accompanied by inflammation, cellular injury, and/or fibrosis (1). In the United States, NAFLD is recognized as an underlying cause for a significant proportion of chronic liver disease (1, 2). NASH may occur in up to 60% of individuals with NAFLD and significantly increases the risk for cirrhosis and hepatocellular carcinoma (1, 3).

Liver biopsy is considered the gold standard for measuring liver fat content, but there are risks associated with this invasive procedure (4, 5). There is also significant sampling variability among intraindividual liver specimens as a single biopsy is ∼1/50,000th of the total liver volume (6). Thus, noninvasive methods are regarded as essential tools to monitor liver fat content and are particularly useful in clinical trials given the limitations of biopsy. Imaging-based methods, including MRI and magnetic resonance spectroscopy (MRS), are available to estimate liver fat content (7, 8). While MRS is widely accepted as a highly accurate quantitative method, it has several limitations that preclude it from widespread use, as it is more time consuming, more complex to conduct, and less widely accessible compared with MRI (8, 9). Modifications of the original Dixon MRI methodology that use newer water-fat image-acquisition protocols have improved separation of the water and fat signals for more accurate liver fat quantification (10, 11). Together, these characteristics of MRI have increased its use as a reliable, noninvasive method for the assessment of liver fat content (7, 9, 12, 13).

After MRI scan acquisition, there is currently no standardized, image-analysis procedure for the most accurate estimate of liver fat content, and the total contoured area analyzed for liver fat content determination differs among studies (14–18). The total area of liver (area per slice and number of slices) covered by the regions of interest (ROIs) might influence measurement variability, test-retest reliability, and the amount of time required for analysis. Furthermore, variability in fat content throughout the entire liver has not been systematically assessed. This is important because variability across an individual slice or the entire organ may influence the total liver area or ROI size required for a reliable estimate of fat content. We determined intraindividual reliability of MRI-based liver fat measurement using 10 different MRI slice analysis methods in normal-weight, overweight, and obese individuals. We also compared the agreement in the determination of liver fat content between analytical methods. Finally, we assessed the variability in fat content across the entire liver. Taken together, these results may be used to inform the best analytical practice for the determination of human liver fat content using data from MRI scans.

Methods

Subjects

Study participants were recruited from the greater Seattle area. Subjects underwent telephone and in-person screening to determine eligibility to undergo MRI scanning. Eligibility criteria included BMI (kg/m2) ≥20 and age between 18 and 75 y. Individuals were excluded if they had contraindications to undergo MRI scan (i.e., claustrophobia, imbedded or implanted metal objects, large abdominal tattoos) or excessive alcohol intake (>14 drinks/wk). Written informed consent was obtained from all subjects. All scans were conducted between January 2016 and August 2017. This study was approved by the institutional review board at Fred Hutchinson Cancer Research Center.

MRI scan

Subjects were nonfasted when they underwent 2 same-day abdominal scans (referred to as “A scan” and “B scan”) on a Philips 3T Ingenia CX (Philips Healthcare) whole-body scanner at the Bio-Molecular Imaging Center at the University of Washington in Seattle, Washington. After the first scan, the subject was removed from the scanner and out of the scanner bed. After a short break of ∼10 min, the repeat scan (B scan) was started by placing the subject on the scanner bed, re-landmarking and rescanning the same sequences as the A scan. Images were acquired using the posterior coil integrated into the scanner table and a 32-channel anterior coil placed on the abdomen. After a 3-plane survey scan, a 3-dimensional (3D) 6-echo Dixon (mDixon) sequence with bipolar gradients was obtained with the following scan parameters: TR/TE1/delta TE (milliseconds), 5.8/1.00/0.7; flip angle 3°; matrix, 152 × 122; resolution (millimeters), 2.5 × 2.5 × 6.0; parallel imaging acceleration factor 2; and bandwidth, 2492 Hz, with a 13-s breath-hold. Seventy-five slices were obtained in each scan, capturing the entire liver. Proton density fat fraction (PDFF) maps were automatically calculated and stored in dicom format. As reference for segmentation, a higher-resolution 3D 2-point Dixon sequence (1.7 × 1.7 × 3.5 mm interpolated to 1 × 1 × 1.75 mm) was also obtained as a visual aid to be able to avoid including smaller liver vessels during segmentation, using a 17-s breath-hold.

Image processing

Dicom images were transferred to a custom segmentation and analysis program (CASCADE; University of Washington, Seattle, WA). Segmentation of the whole liver was conducted to set the gold standard for hepatic fat content in this study. The initial 5 and final 5 liver-containing slices were omitted due to minimal measurable area. Water images from both mDixon and higher-resolution anatomical reference Dixon sequence were displayed side-by-side. Contours were drawn on the mDixon sequence and projected to the reference Dixon sequence. The anatomical reference Dixon sequence was used as a visual aid to avoid including areas of vasculature within the contours. Contours covering the liver parenchyma were drawn to avoid vasculature and areas where partial volume averaging might occur, such as liver margins. More than 1 contour was used when discontinuous areas of the liver were present in an image. Examples of contours are shown in Supplemental Figure 1. Effort was taken to avoid large hepatic vessels in segmentation. Mean, SD, maximum, minimum, and range of PDFF corresponding to each contour were determined. Any contour that resulted in a negative value was considered to reflect an undetectable level of fat and a value of zero was input to replace negative values.

The mean PDFF of each slice was weighted by the slice contour area (millimeters squared) as a percentage of the total contoured liver area. For each scan, the weighted PDFFs of all liver slices were then summed. We set this analytical method as the study gold standard for percentage liver fat. To compare the study gold standard and smaller areas of contoured liver, we also determined PDFF based on every other, every third, every fifth, every 10th, every 20th, and middle slice. We also assessed PDFF based on a circular 3-cm2 ROI on the middle image, middle 3 images, and middle 5 images. One reader (MSB) contoured all scans and was blinded to both clinical characteristics and scan pairs.

Statistical analyses

Statistical analyses were performed using the Statistical Package for the Social Sciences for Macintosh (version 26.0; IBM) and GraphPad Prism (version 8.1; GraphPad Software). Distribution of variables was assessed by histograms and the Kolmogorov-Smirnov and Shapiro-Wilk tests of normality. All data were log transformed prior to intraclass correlation coefficient (ICC) and Pearson's correlation coefficient analyses. In the 3-cm2 ROI analyses, some participants had a PDFF value of zero. For these analytical approaches, a value of 1 was added to the PDFF values for all subjects prior to log transformation. A 2-way mixed effects model for ICC was performed for each image analysis method to determine the reliability of repeated abdominal MRI scans for the measurement of liver fat. ICCs were computed with all subjects together and also divided into 2 subgroups based on presence (≥5% fat) or absence (<5% fat) of fatty liver. ICC values <0.4 were considered poor, values between 0.4 and 0.75 were considered fair to good, and values >0.75 were considered excellent (19). Pearson's correlation coefficients were computed for each contour method to determine the association between PDFFs as determined by repeated MRI scans. Bland-Altman analyses were performed to assess the level of agreement between the gold-standard contour method and each of the alternative contour methods described above (20).

Results

Subject characteristics

Twenty individuals were enrolled into this cross-sectional study and all participants underwent 2 same-day abdominal MRI scans. Overall, the mean ± SD age was 49.9 ± 16.5 and 53.8 ± 16.3 y for women (n = 11) and men (n = 9), respectively. BMI was 30.3 ± 6.8 and 28.9 ± 4.3 for women and men, respectively. Subject characteristics by presence or absence of fatty liver disease are shown in ; the group with fatty liver (≥5% liver fat, n = 7) tended to be older and have a higher BMI than the group without fatty liver (<5% liver fat, n = 13).

TABLE 1

Characteristics of study participants, grouped by absence (<5% fat) or presence (≥5% fat) of fatty liver

	Liver fat <5% (n = 13)		Liver fat ≥5% (n = 7)
	Mean ± SD	Range	Mean ± SD	Range
Age, y
Women (n = 11)	45.7 ± 18.7	23–70	57.3 ± 9.6	44–67
Men (n = 9)	50.5 ± 16.7	21– 65	60.3 ± 16.5	42–74
BMI, kg/m2
Women	27.3 ± 2.9	23.4–30.9	35.7 ± 8.7	26.8–43.2
Men	27.8 ± 3.3	24.1–32.3	31.2 ± 5.9	27.6–38.0

Variability in liver fat content across the entire organ

The whole liver was contoured on every slice to determine the study gold standard for the PDFF of each scan for all participants. The mean number of contoured liver slices across all scans was 51, with a range of 35 to 67 slices. Liver fat content ranged from 0.69% to 18.5% among the A scans and 0.58% to 18.4% among the B scans based on the study gold-standard PDFF method. Assessment of PDFF on a slice by slice basis revealed variability in fat content across the liver for nearly all study participants (). For each scan of all subjects, we calculated the degree to which the slices with the lowest and highest PDFF values of each individual deviated from the mean PDFF across all slices (denoted by the dashed lines in Figure 1). We then calculated the overall mean minimum and maximum deviations across all scans. We found that, on average, the lowest PDFF values were only 46% of the mean PDFF while the highest PDFF values were 175% of the mean (Supplemental Figure 2). Together, these results demonstrate the substantial variability in fat content throughout the liver in both the presence and absence of fatty liver.

FIGURE 1

Individual variation in hepatic fat content. Abdominal MRI scans were conducted in nonobese and obese adults to assess liver fat content. Participants are grouped by absence (<5% liver fat) (A) or presence (≥5% liver fat) (B) of fatty liver disease. Note the different scales of the y-axes in A and B panels; this is to allow for better visualization of variability in fat across the liver among subjects with low liver fat content. The whole liver was contoured on every slice. Slice number refers to transverse liver sections going in the caudal to cranial direction. Dashed lines represent mean percentage liver fat for scan A (shown in red) and scan B (shown in black). PDFF, proton density fat fraction.

Reliability of liver fat content by repeated MRI scans

ICCs for the 10 analytical methods for PDFF determination revealed excellent (>0.75) reliability when all subjects were included (). To assess whether reliability varies by high or low liver fat content, we also performed ICC analyses for the 10 methods after categorizing the subjects based on the absence (<5% liver fat) or presence (≥5% liver fat) of hepatic steatosis as assessed by the study gold-standard method. ICCs for the PDFF in the group without fatty liver were fair to excellent, ranging from 0.50 to 0.79 (Supplemental Table 1). ICCs were consistently higher for the methods that contoured the whole liver as compared with methods that relied only on 3-cm2 ROIs (Supplemental Table 1). ICCs for all contour methods were excellent for the fatty liver group, with a range of 0.86 to 0.95 (Supplemental Table 2).

TABLE 2

ICCs for PDFFs between 2 successive same-day MRI scans, all participants included

Contour method	ICC (95% CI)
All slices	0.95 (0.88–0.98)
Every other slice	0.95 (0.87–0.98)
Every third slice	0.95 (0.87–0.98)
Every fifth slice	0.94 (0.86–0.98)
Every 10th slice	0.95 (0.87–0.98)
Every 20th slice	0.95 (0.88–0.98)
Middle slice	0.88 (0.72–0.95)
Middle 5 slices ROI	0.90 (0.76–0.96)
Middle 3 slices ROI	0.90 (0.76–0.96)
Middle slice ROI	0.89 (0.74–0.95)

n = 20. ICC, intraclass correlation coefficient; PDFF, proton density fat fraction; ROI, region of interest.

Pearson's correlation coefficients revealed strong relations between PDFFs from the 2 scans (). The correlation coefficients for the analytical methods that included contouring of whole-liver slices (Figure 2A–F) were slightly greater (all ≥0.95) than those that contoured only 1 slice or relied on 3-cm2 ROIs, which ranged from 0.88 to 0.90 (Figure 2G–J). We also conducted Pearson's correlations separately for subjects grouped according to absence or presence of fatty liver (Supplemental Tables 3 and 4). Among subjects without fatty liver, correlation coefficients for methods that contoured whole-liver slices ranged from 0.74 to 0.79 (Supplemental Table 3), although the method that relied only on the middle slice had a correlation coefficient of 0.49. For methods that used 3-cm2 ROIs, the correlations were also weaker, with r values between 0.51 to 0.55 (Supplemental Table 3). The correlation coefficients for the repeat scans were strong for all analytical methods among the subjects with fatty liver, ranging from 0.92 to 0.98 (Supplemental Table 4).

FIGURE 2

Associations between hepatic fat content estimated by same-day abdominal MRI scans (scans A and B) using 10 different liver slice analysis methods. Pearson's correlation coefficients between the 2 scans were calculated for analyses when the entire liver was contoured on all slices (A), every other slice (B), every third slice (C), every fifth slice (D), every 10th slice (E), every 20th slice (F), and middle slice only (G). Correlations were also conducted between the 2 scans when hepatic fat content was assessed using a 3-cm2 ROI on the middle 5 slices (H), middle 3 slices (I), and middle slice only (J). Light blue indicates participants without fatty liver (<5% fat), dark blue indicates participants with fatty liver (≥5% fat). All data were log transformed prior to analyses. P < 0.0001 for all correlation coefficients. PDFF, proton density fat fraction; ROI, region of interest.

Agreement between analytical methods

Bland-Altman analyses were performed to assess the agreement between the study gold standard and the alternative analytical methods. There were no obvious patterns in the absolute differences between the gold standard and any method, with similar absolute deviation across the range of PDFFs within any given method comparison in the A scans (). However, there was a clear trend for lower agreement with the gold standard as the total area of contoured liver in the comparison method declined (i.e., a greater spread of the data points and wider CIs). The level of bias (mean difference; blue dashed line in Figure 3) between the gold standard and the alternative methods ranged from −0.005 to 0.22. The absolute difference in PDFF between the gold standard and every other, every third, every fifth, and every 10th slice was <1% for all subjects (Figure 3A–D). However, the absolute difference increased up to 2% when agreement between every 20th slice, the middle slice, or a 3-cm2 ROI on the middle slice, middle 3 slices, or middle 5 slices with the gold standard was assessed (Figure 3E–I). Overall, results were similar for the set of B scans (data not shown), with the bias for each of the method comparisons comparable to that of the A scans. However, the maximum absolute difference between the gold standard and each of the methods using ROIs was 4% for 1 participant in the B scans.

FIGURE 3

Bland-Altman plots of agreement between different analytical methods of abdominal MRI images to estimate hepatic fat content. The whole liver was contoured on every slice to set the study gold standard for liver fat content. Plots show the difference between the study gold standard (all slices) and whole-liver contours on every other slice (A), every third slice (B), every fifth slice (C), every 10th slice (D), every 20th slice (E), and middle slice (F). Plots of agreement are also shown for the study gold standard and a 3-cm2 ROI on the middle 5 slices (G), the middle 3 slices (H), and the middle slice (I). The blue dashed line represents the bias (mean difference) between methods; the numeric value of the bias is shown in blue text. The black dotted lines represent the 95% CI of the limits of agreement. The x-axis mean PDFF % is the mean of the 2 methods compared in each plot. All results shown were generated from the A scans. PDFF, proton density fat fraction; ROI, region of interest.

We then constructed Bland-Altman plots based on percentage difference between methods [e.g., 100 × (study gold standard % fat − every other slice % fat)/mean % fat] to assess relative differences in agreement across the range of PDFFs. The level of bias differed among the analysis methods and ranged from −0.07% for the method that included every third slice up to ∼30% for the methods that included only 3-cm2 ROIs (Supplemental Figure 3).

Two main observations emerged from this analysis. First, the relative agreement with the study gold standard declined as the total area contoured declined. While the agreement between the gold standard and every other, every third, every fifth, and every 10th slice was good, the agreement declined noticeably starting with every 20th slice and declined further when only 3-cm2 ROIs were contoured (Supplemental Figure 3E–I). The bias (denoted by the blue dashed line in each panel) of ∼30% for the 3 methods that relied on ROIs (Supplemental Figure 3G–I) reveals that these methods tend to underestimate fat content relative to the study gold standard, although this problem occurred predominantly among the subjects with low liver fat. Indeed, the second main observation from these analyses revealed that the relative differences between methods among measurements in subjects without fatty liver (i.e., when liver fat content was <5%) had the greatest impact on the bias. This was especially problematic among the methods that relied on the 3-cm2 ROIs, with >100% deviation from the mean PDFF for several individual subjects (Supplemental Figure 3G–I). Results for the set of B scans were nearly identical (data not shown).

Discussion

We conducted this study to assess reliability of liver fat quantification by the Philips 3T Ingenia whole-body MRI scanner and to determine whether reliability is impacted by the analytical approach of the MRI images. Manual contouring of MRI slices for quantification of liver fat is a time-intensive endeavor. This may be especially true for longitudinal research studies in which subjects undergo MRI scans at multiple time points to determine the effect of a given intervention on liver fat content. In addition, with several dietary components contributing to the development of NAFLD, the assessment of liver fat content is central to clinical nutrition research studies or nutritional therapy related to the prevention or management of fatty liver disease (21, 22). For example, the effect of various therapies that target the gut microbiome as well as a wide range of dietary supplements on liver fat content using MRI-PDFF measurements have been reported (23, 24). Currently, the field lacks a standardized method for the analysis of MRI slices for liver fat quantification. We questioned whether image analysis methods that do not require contouring the entire liver on every MRI slice would agree with fat content estimates by contouring the entire liver on every slice, which we defined as the study gold standard. To address this, we assessed the performance of 9 different post-scan image analysis methods relative to the gold standard. Our analyses revealed that the degree of agreement between the gold standard and the alternative methods declined as the total area of contoured liver declined. We found that contouring the whole liver on every other, every third, every fifth, or every 10th slice provided close agreement with PDFFs based on contouring every slice. However, the agreement with the gold standard declined when fewer than every 10th slice was contoured or when 5, 3, or 1 3-cm2 ROIs were used.

MRI image analysis methods currently used vary by total liver area contoured, number of slices contoured, location of slices within the liver, and regional placement of ROIs on liver slices. Examples from the literature confirm such variability in methodology performed and include the following: single ROIs on each of the 9 liver segments on an image (25, 26), 5 ROIs on an image (14), 3 ROIs on each of the 9 liver segments (27), a single ROI on 3 slices (28), and an ROI on the right and left liver lobes (29). Few studies report which liver slice(s) are contoured, or whether the contoured liver slice was chosen using a standardized approach. There are a limited number of reports comparing MRI analysis methods for the best estimate of liver fat content. Vu et al. (30) compared the reliability of 17 ROI sampling methods that ranged in total contoured area from 0.8 to 45 cm2 with varying coverage of the liver lobes. Although most sampling techniques yielded excellent correlation against whole-liver contouring, the authors found that it was attenuated as the total area contoured declined. In addition, Vu et al. (30) reported that reliance predominantly on the right liver lobe for ROI contouring resulted in a significantly greater liver fat content estimation as compared with whole-liver segmentation.

ICCs of liver fat quantification showed excellent reliability when all subjects, representing a wide range of liver fat content from ∼1% to 18%, were included in the analysis. Our results are consistent with others that have also reported high reliability of MRI-based measurement of liver fat content (17, 26, 28, 31). Excellent reliability was reported in a study of obese children and adults, with a liver fat range of ∼1% to 35%, who underwent 3 same-day MRI scans. In this study, 1 slice was analyzed per MRI scan in which liver fat was assessed in the 9 Couinaud liver segments and then averaged to determine right lobe, left lobe, and whole-liver fat content; ICCs for all regions assessed were >0.99 (31). Wu et al. (17) conducted a study of 15 individuals who underwent a total of 4 MRI scans, 2 scans each in 2 different scanners. Excellent reliability was reported for both intra- and interscanner measurements, with ICCs >0.83 across both hepatic lobes. Using a single ROI on 3 consecutive slices, Tyagi et al. (28) reported excellent inter-examination reliability, with an ICC of >0.99 in a cohort of overweight and obese subjects who underwent 3 same-day MRI scans with liver fat content ranging from 0.4% to 34%.

However, none of these studies assessed the reliability of liver fat measurements separately for individuals grouped by the presence or absence of fatty liver. Our results address this gap and demonstrate that, although reliability is excellent when assessed across a wide range of liver fat percentages, the low end of the range suffers from reduced test-retest reliability when a limited area of the liver is contoured. Despite this, MRI-based quantification of liver fat sufficiently distinguishes between absence and presence of fatty liver. Furthermore, given the very wide range of liver fat content that would likely be seen in most Western populations, particularly those with obesity or metabolic disease, the relative inaccuracies in individuals with low liver fat may be less concerning than they seem.

Heterogeneity in liver fat distribution is a well-known feature and may take various forms among individuals (32). The most common pattern for liver fat is diffuse distribution, but in some cases, it does not occur homogenously (32, 33). For example, other distribution patterns may include localized regions of accumulation—for example, near the gallbladder or specific liver segments, localized regions of fat sparing, or several distinct fatty lesions spread throughout the organ (32, 33). It is possible that at least some of the variability in fat content observed by liver biopsy may be explained by these heterogenous fat distribution patterns (6). In ex vivo livers, Bannas et al. (26) demonstrated lower agreement between repeated liver fat measurements using histological assessment of biopsies when compared with quantification by MRI across a wide range of liver fat content.

Although we did not investigate heterogeneity within slices to describe fat-distribution patterns, our slice by slice PDFF assessment revealed a heterogenous distribution throughout the liver for nearly all subjects. Others have reported variability in hepatic fat content by MRI across the 9 Couinaud segments that make up the liver, with significant interindividual variability in the degree of liver fat heterogeneity (26, 30, 34). Hui et al. (35) conducted a study to assess heterogeneity of liver fat distribution in adults and children with or without NAFLD. The authors placed an ROI on the superior right lobe, superior left lobe, and the inferior right lobe to assess fat within different regions. Fat content differed significantly among the 3 regions in both NAFLD groups but not in the groups without NAFLD, with the highest fat content in the superior right lobe and the lowest content in the superior left lobe. Regional variation in fat content has been hypothesized to be at least partially due to differential insulin exposure or contribution of blood sources (mesenteric vs. splenic) across the hepatic lobes (32, 34, 35).

Our study had several limitations. First, as we had only 1 reader contour the MRI slices, we do not have an assessment of interobserver reliability. Whether our findings of strong agreement between analytical methods and high variability in fat content throughout the liver would be consistent with other observers is not known. In addition, because all of our scans were conducted on a single-vendor MRI scanner and in a single research facility, our results may not be broadly generalizable to other scanners and techniques. Our study did not include any participants with known liver disease, which may limit the generalizability of the results to a broader patient population. Another limitation is that we did not assess the reliability of the panel of analytical approaches using currently available automatic or semi-automatic segmentation software. With increased availability of such automatic software programs, the reliance on manual segmentation for liver fat assessment may be declining.

The prevalence of fatty liver disease has reached epidemic proportions in adults and children (21, 36). Thus, the availability of reliable, noninvasive methods for liver fat quantification is critically important, not only to assess liver fat content but also to assess responses to interventions (37). However, it is also important to note that it remains unclear whether changes in liver fat are informative of potential changes in histological characteristics of NASH, including lobular inflammation, hepatocyte ballooning, and fibrosis. A recent longitudinal study of NASH patients demonstrated no association between improvements in liver fat content and histological changes (38), suggesting that the assessment of liver fat content alone may not be a robust biomarker for NASH histological features.

Overall, our results are consistent with the existing literature that demonstrates excellent reliability of liver fat quantification by MRI and the ability to distinguish between the absence or presence of fatty liver. It should be emphasized, however, that the greatest reliability occurred among individuals with hepatic fat content ≥5%. Together with the evidence discussed earlier, our study results demonstrate that total area and location of contour sampling affect the hepatic fat estimate as determined by MRI. These findings may be useful in clinical trials and the clinical assessment of liver fat content that rely on MRI-PDFF and lack a standardized post-scan processing method. The results from our assessment of different analytical methods indicate that, for the most reliable estimate of liver fat content, it is beneficial to contour larger areas, which may be especially important in longitudinal studies or studies that include a low-liver-fat group. Specifically, our data suggest that the entire liver should be contoured on at least every 10th slice to achieve close agreement with the gold standard.

Click here for additional data file.