Novel segmentation method for abdominal fat quantification by MRI

Authors

  • Anqi Zhou MS,

    1. Department of Radiology, University of Texas Health Science Center at San Antonio, San Antonio, Texas, USA
    Search for more papers by this author
  • Horacio Murillo MD, PhD,

    1. Department of Radiology, University of Texas Health Science Center at San Antonio, San Antonio, Texas, USA
    Search for more papers by this author
  • Qi Peng PhD

    Corresponding author
    1. Department of Radiology, University of Texas Health Science Center at San Antonio, San Antonio, Texas, USA
    2. Research Imaging Institute, University of Texas Health Science Center at San Antonio, San Antonio, Texas, USA
    3. Department of Radiology, Albert Einstein College of Medicine and Montefiore Medical Center, Bronx, New York, USA
    • Department of Radiology, Montefiore Medical Center, Albert Einstein College of Medicine, 111 East 210th Street, Yellow Zone, Bronx, NY 10467
    Search for more papers by this author

Abstract

Purpose:

To introduce and describe the feasibility of a novel method for abdominal fat segmentation on both water-saturated and non–water-saturated MR images with improved absolute fat tissue quantification.

Materials and Methods:

A general fat distribution model which fits both water-saturated (WS) and non–water-saturated (NWS) MR images based on image gray-level histogram is first proposed. Next, a novel fuzzy c-means clustering step followed by a simple thresholding is proposed to achieve automated and accurate abdominal quantification taking into consideration the partial-volume effects (PVE) in abdominal MR images. Eleven subjects were scanned at central abdomen levels with both WS and NWS MRI techniques. Synthesized “noisy” NWS (nNWS) images were also generated to study the impact of reduced SNR on fat quantification using the novel approach. The visceral adipose tissue (VAT) and subcutaneous adipose tissue (SAT) amounts of the WS MR images were quantified with a traditional intensity thresholding method as a reference to evaluate the performance of the novel method on WS, NWS, and nNWS MR images.

Results:

The novel approach resulted in consistent SAT and VAT amounts for WS, NWS, and nNWS images. Automatic segmentation and incorporation of spatial information during segmentation improved speed and accuracy. These results were in good agreement with those from the WS images quantified with a traditional intensity thresholding method and accounted for PVE contributions.

Conclusion:

The proposed method using a novel fuzzy c-means clustering method followed by thresholding can achieve consistent quantitative results on both WS and NWS abdominal MR images while accounting for PVE contributing inaccuracies. J. Magn. Reson. Imaging 2011;. © 2011 Wiley-Liss, Inc.

THE ASSOCIATION OF human abdominal fat and its body distribution with multiple metabolic syndrome abnormalities has fueled interest in accurate, fast and automatic methods of adipose tissue quantification (1–4). Of the current imaging methods that can be applied, computed tomography (CT) and MRI are the most promising tools (5, 6). Compared with CT, MRI can generate high quality cross-sectional images without ionizing radiation. Therefore, MRI is the preferable imaging modality particularly for longitudinal studies and studies that involve young subjects. Despite its straightforward appearance on MR images, abdominal fat tissue quantification has been challenging. Chief among these challenges are the disseminated distribution of fatty tissue and its inherent partial volume effects (PVE) in the visceral compartments of the abdomen and pelvis. Moreover, inherent visceral motion from peristalsis, vascular flow, pulsation, and breathing further contribute to limited accurate quantification.

Due to the highly complicated abdominal compartments and the disseminated nature of visceral fat, the ratio of the number of partial-volume (PV) voxels to full-volume (FV) voxels may be large (7). Its impact on VAT quantification may also vary significantly depending on body size, habitus, and adiposity. A recent study has shown that average PV and FV fat amounts are comparable in a group of subjects with an average BMI of 26, however, the magnitude of measurement uncertainty contributed by the PVE is greater in leaner individuals than those who are obese (8). Therefore, whatever the magnitude of the PVE, it needs to be accounted for to accomplish accurate and reproducible VAT measurements.

In addition to a needed accurate and reproducible method of fat quantification, such method must be easy to perform and have no significant untoward effects by its repetition in the same subject. Non–water-saturated (NWS), T1-weighted (T1W) turbo spin-echo (TSE) images have been traditionally used for subcutaneous abdominal tissue (SAT) and visceral abdominal tissue (VAT) quantification (9–11). Although fat signal is usually significantly higher than nonfat tissues, several factors may complicate the image quality and, therefore, accurate quantification of VAT. First, the traditional methods may have relatively low signal-to-noise ratio (SNR) and/or low contrast-to-noise ratio (CNR) between fat and nonfat, because multiple images have to be acquired within a limited time window during breath hold to minimize respiratory motion. Second, images obtained using traditional methods usually suffer from blood flow artifacts. Moreover, the image quality is usually compromised by the above mentioned motion artifacts and PVE.

Another factor that contributes to the difficulties is the inherently different Tmath image and T2 for the different abdominal organ tissues leading to signal intensity differences. In this context, usually the largest visceral organ parenchyma signal becomes dominant (most often the liver). All these have contributed to the difficulties when it comes to accurate abdominal fat (VAT and SAT) quantification on MR images. Therefore, laborious manual contour drawing or region-of-interest (ROI) editing approaches are still being used in many studies to accomplish accurate and reproducible measurements despite their time consuming and shortcomings of large inter- and intra-observer variability.

The intermediate signal intensity of PV voxels makes it challenging to separate abdominal voxels into either a water or other nonfatty tissue voxel and a fat voxel in a NWS image. Alternatively, water-saturation (WS) MR imaging or more advanced techniques such as “iterative decomposition of water and fat with echo asymmetry and least squares estimation” (IDEAL) can be applied to generate MR images with suppressed nonfatty tissue signal, which may help identify and, therefore, estimate PV fat amount due to greatly improved contrast (12–15). However, these techniques are not yet widely used because additional advanced MRI scans dedicated to fat quantification are needed.

Here, we propose a simplified fat tissue quantitative method based on image histogram for NWS or WS MR images that incorporates spatial information. This model takes both FV and PV fat voxels into consideration. Based on this model, we further describe an automated segmentation approach, which uses a new fuzzy c-means (FCM) clustering method followed by a simple thresholding step to quantify total VAT (FV and PV fat). The method was tested in WS, NWS, and synthesized noisy NWS images to evaluate its feasibility and its sensitivity to image noise while accounting for PVE contributions.

MATERIALS AND METHODS

Simplified Fat Distribution Model

In an ideal NWS spin-echo image (neglecting noise, and other imaging imperfections such as B0 and B1 inhomogeneities), signal intensity from a voxel consisting solely of adipose tissue would give a signal maximum of the image (Sf), and the signal intensity from a voxel consisting solely of nonfatty tissue would give a signal of Sw which is lower than Sf (neglecting the proton density and T1 and T2 differences among nonfatty tissues, and possible air in the intestines). All other voxels that are partially filled with adipose tissue (the other part is filled with nonfatty tissue) would have signal between Sw and Sf, depending on the volume ratio of fat in that voxel. For voxels with FV fat or water, they will ideally give a single signal intensity of Sf and Sw, respectively. For PV fat voxels, the signal is determined by:

equation image(1)

where ρ is the volume percentage of fat within that voxel. It is worth noting that this relationship is not necessarily true for all MR images. It can only be applied to MR images acquired with spin echo (SE), TSE, and gradient recalled echo (GRE) techniques using in-phase TEs. GRE images using other TEs may suffer from compromised signal summation of water and fat signal in PV voxels when water and fat spins are not fully in-phase.

Ideally, the number of FV fat voxels (Nf) can be demonstrated on the image intensity histogram as shown in Figure 1a. FV fat voxels are mainly from bulk fat, such as subcutaneous fat, and PV fat voxels are located mainly at the water–fat tissue interfaces. Due to the complexity of fat distribution, the distribution of the PV fat signal within the histogram is not predictable. However, if the number of PV voxels in an image is large enough, the probability of a PV fat voxel to have a certain ratio of fat should be close to the probability of another ratio. Therefore, PV voxel signal distribution is nearly constant over all signal intensity between Sw and Sf on the histogram. Thus, a uniform (or rectangular) distribution as shown in the gray area of Figure 1a can be used to approximate the PV fat signal distribution. The voxel density (Np) in the histogram is image-specific and can be derived from the image histogram. This has been shown previously in WS MR images, and the distribution should also work for NWS images.

Figure 1.

Histogram model of MR images. a: Ideal image histogram when no imaging noise exists. b: Imaging noise kernel. c: Theoretical (dashed curve) and predicted image histogram (solid curve) of a NWS MR image. d: Theoretical (dashed curve) and predicted image histogram (solid curve) of a WS MR image.

Considering both FV and PV fat signal, the image histogram is shown as the dashed lines in Figure 1c if an ideal imaging method is used, as predicted by the fat distribution model. With imaging noise, the resultant histogram will be as shown in the solid curve in Figure 1c, which is derived from the ideal histogram, convolved with a normalized Gaussian kernel representing Gaussian noise (with standard deviation = σ) (Fig. 1b). When Sf - Sw is much larger than σ, the two peaks will not overlap, and fat quantification can be performed using a simple threshold method as follows.

Overall, the total fat volume of an image can be calculated as:

equation image(2)

where Vtotal is the total volume of fat on the image, Vfull is the total volume of fat in all FV fat voxels, and Vpartial is the total volume of fat in PV fat voxels. Because on average, PV voxels are half filled, theoretically fat volume can be calculated as follows:

equation image(3)

where Vvoxel represents volume size of a single image voxel. Therefore, the signal threshold which best separates fat from nonfat is:

equation image(4)

By combining Equations (3) and (4), fat volume is determined to be:

equation image(5)

Note that Nf + (Sf − Sth) × Np is the integration of the histogram curve above Sth, representing the total number of voxels with signal intensity higher than Sth (Fig. 1c). Therefore, a fat-only binary image can be generated by setting the voxels with intensity higher than Sth to 1.0, and the other voxels to 0. Total fat volume can be calculated by the total number of fat voxels times Vvoxel. Using this approach, only Sth is used in the final fat quantification process, and the accuracy of Nf and Nw from curve-fitting does not significantly influence the final fat quantification. To include both FV and PV voxels in fat volume measurement, Sth should be best set to (Sf + Sw)/2.

The same simplified model works for both WS and NWS MRI images. When effective water suppression is achieved, signal of nonfat tissues will be close to noise level (Sw ≈ 0) and the water peak will deviate from Gaussian shape. Under such conditions, it will be close to Rician distribution as shown in Figure 1d and Sth will not change. This general model will, therefore, degenerate to the model for WS MR images, as described previously (14).

General Quantification Method

Ideally, the fat and water peaks are well separated (i.e., CNR between fat and nonfat is higher than 4) and a simple thresholding method as described above should result in accurate fat quantification. However, when the two peaks have some overlap, the quantification based on histogram analysis will become more difficult. This is because a voxel that has intermediate signal intensity between Sw and Sf may be one occupied by FV fat, by FV water, or by a mixture of both water and fat (PV voxel). To categorize these voxels, more features other than voxel signal intensity are needed. As previously shown, visceral PV fat voxels are mainly located adjacent to bulk fat voxels in human abdomen (14). This spatial information can, therefore, be used in combination with voxel signal intensity to label PV voxels, which will ultimately lead to more accurate fat quantification and less variability between studies (16).

This can be accomplished with a two-step procedure to first eliminate FV water voxels, so that a “pseudo-water-saturated” (pWS) MR image can be generated before a thresholding step similar to that of a WS MR image. In the first step, a FCM clustering method is used in combination with spatial information to separate the voxels into two fuzzy clustering centers (water-only voxels and non–water-only voxels). The first clustering center contains voxels that are occupied by FV nonfatty tissues only, and the second clustering center contains voxels that are fully or partially occupied by fat. For an image with N pixels, the FCM clustering method used an automated algorithm to minimize the following objective function (17):

equation image(6)

where, m is the fuzzy exponent that controls the amount of fuzziness of the classification which is two in this study; X = {x1,x2,…xk,…xN} represents a data set in a finite-dimensional vector space; μik is the degree of membership of object xk to cluster i; dik = ∥xkvi∥ is the distance between an object xk and the cluster center vi. The most commonly used feature (i.e., the data set X) is the intensity value of image pixel. To exploit the spatial information, we extend our data set as (yk ,sk) where, yk is the intensity value at the k-th location and sk, is a spatial function at the k-th location which is defined as:

equation image(7)

where equation image is a square mask centered on k-th voxel, and yl represents the intensity values within the mask. We used 5 × 5 square mask in this study. The membership function and cluster centers can be obtained through an iterative process:

equation image(8)
equation image(9)

The water-only image will be generated using a threshold value of 0.3 to include only FV water voxels. The pWS MR image will be generated by subtracting the water-only image from the original image.

In the second step, a simple signal intensity thresholding will be applied to the pWS MR image to generate a binary “fat-only” image. A threshold value of (Sf + Sw)/2 will be used so that PV fat voxels with signal above this threshold can be compensated by the fat amounts in voxels with signal below this threshold value. The total fat amount can, therefore, be calculated by the number of fat voxels on the fat-only image multiplied by the voxel volume size.

MRI Acquisition

Eleven healthy volunteers (8 males and 3 female) with mean age of 41 ± 9 years old, mean body mass index (BMI) of 25.1 ± 2.7 kg/m2 participated in this study with approval of the Institutional Review Board. Informed written consent was obtained from each volunteer before the exams. These subjects underwent abdomen imaging on a 1.5 Tesla (T) clinical MR scanner (Intera, Philips Healthcare, Netherlands) using a standard whole-body quadrature coil. After survey imaging, each subject underwent both NWS T1W TSE and WS T1W TSE. In each sequence, 6 axial slices centered at L2–L3 level were acquired in two consecutive expired breathholds, with 14 s each. The imaging FOV varied from 400 mm to 500 mm depending on the size of each volunteer. Imaging parameters included: TR/TE/FA=500 ms/5 ms/90°, turbo factor = 7 and readout bandwidth = 128 kHz. Image pairs were visually inspected to exclude images with severe fat amount or distribution changes between WS TSE and NWS TSE images due to peristalsis. Total 60 WS and NWS image pairs were included in this study.

Postprocessing

Automated computer algorithms were developed to perform fat quantification within a software package named “QFAT” developed using Interactive Data Language (IDL, Research Systems, Inc., Boulder, CO). This algorithm used several steps to quantify both VAT and SAT amounts for each WS and NWS image. First, automated intensity correction was performed on each original slice, which iteratively applied two-dimensional polynomial field correction to find the minimal entropy of the image grayscale histogram (18). This procedure reduced the signal nonuniformity due to B0 and B1 inhomogeneities (19). Second, automatic regional segmentation was performed to generate the following three contours using a chain code algorithm for boundary detection (20). Details of the automatic regional segmentation algorithms will be published later in another article. Briefly, the first contour was the external one surrounding the abdominal region and with arms excluded (red contour in Fig. 2aWS). The second contour was located at the internal SAT boundary covering the abdominal muscles (green contour in Fig. 2bWS). The third contour covered VAT but carefully excluded intra-muscular and vertebral fat (yellow contour in Fig. 2bWS). The three automated contours were manually corrected if needed to cover the intent region of interest of each. In the third step, voxels within the first contour underwent FCM algorithm and signal of FV water voxel were set to zero to generate a pWS MR image, as described in the “General Quantification Method” section. Lastly, this pWS MR image was further thresholded with a value of (Sf + Sw)/2 to exclude voxels with fat volume less than half of the voxel size. The number of fat voxels of SAT and VAT was calculated as voxel size times the number of fat voxels within the corresponding anatomic regions. Fat amount of each region was then automatically calculated from the corresponding number of fat voxels multiplied by the volume size of the voxel.

Figure 2.

Representative images and histograms to show the fat quantification steps for WS (upper left), NWS (upper right), and nNWS (lower left) image datasets. a: Original abdominal MR images after intensity correction (corresponding to the black curve in the histogram). b: “pseudo-WS” image with signal of water-only voxels (blue curve) set to zero (red curve in histogram). c: Final fat-only images when Sth thresholding was applied. d: Histograms of images or tissue compartments.

To compare the validity of the proposed fat quantification approach for WS images, QFAT results were compared with results from a simple thresholding method applied to the WS MR images using ImageJ, a public domain image processing program (ImageJ, version 1.42). The threshold value used to generate the “fat-only” image was manually set to Sf/2 (close to (Sf + Sw)/2 for WS images), and the contours were manually drawn to match the corresponding contours used in QFAT. Fat amount of each region was calculated based on the “fat-only” image the same way as in QFAT.

To test the effectiveness of the fat quantification method on NWS images when water and fat peaks overlap on histogram, a synthesized “noisy” NWS (nNWS) image dataset was generated. Random noise with Guassian distribution was added in each NWS image to reduce overall image SNR by half. Exactly the same fat quantification steps as NWS images were followed to quantify SAT and VAT amount on nNWS images. The QFAT results of nNWS images were then compared with those from NWS images.

Statistical Analysis

Fat quantification results of SAT and VAT obtained from WS images using ImageJ, WS, NWS, and nNWS images using QFAT were statistically compared on a slice-by-slice basis. The mean and standard deviation of the measured SAT and VAT amounts were calculated from each measurement. The differences of VAT and SAT using different methods were detected by analysis of variance (ANOVA). The intra-operator reliability of each method and image techniques were estimated by intra-class correlation (ICC) and 95% confidence intervals (CIs). The consistency of fat quantification results was studied using linear regression analysis and the Pearson correlation coefficient (r). All analyses were performed using IBM SPSS Statistics 18 (SPSS Inc., Chicago, USA) for Windows. Bland-Altman plot was also generated to test if the two methods can be used interchangeably. A P value < 0.05 was taken to be statistically significant.

RESULTS

The same automated algorithms were applied to WS, NWS, and nNWS image datasets. Representative images and corresponding histograms are shown in Figure 2. In all histograms, only the voxels within the skin contours (red) are included (the background noise in air voxels were excluded from consideration). In the WS image (Fig. 2aWS), muscles, bowel, and nonfat intra-abdominal organs are effectively suppressed compared with the corresponding NWS image (Fig. 2aNWS). This is also demonstrated in the histogram (Fig 2dWS) with a peak close to zero (SW) and a high signal fat peak centered at Sf. For the NWS MR image (Fig. 2aNWS), the nonfat voxels have signals much larger than zero but less than Sf. These voxels are represented by the large water peak at Sw between zero and Sf in the gray-level histogram (Fig. 2dNWS). In both cases, the segments between Sw and Sf of the black curves in the histograms are close to be constant/flat, validating that fat occupancy of the PV voxels have a random distribution. The histograms for the WS and NWS images are close to the corresponding ones as predicted by the fat distribution models described in the “Simplified Fat Distribution Model” section.

Similar to the NWS image, the nNWS image had a large water peak at Sw between signal zero and Sf (Fig. 2dnNWS). The water and fat peaks have some overlap due to increased peak width when random noises are added to the NWS image. The FC step effectively excluded the FV water voxels for all WS, NWS, and nNWS images, leading the corresponding pseudo-WS images (Fig. 2b) comprised of only FV and PV voxels (others are set to zero). The signal distribution of the excluded full-volume water voxels in each histogram is shown as the blue curve in Figure 2d. The blue curves on the NWS and nNWS histograms are close to Gaussian distribution, as expected. The final fat-only images (Fig. 2c) were generated from the pseudo-WS images after thresholding with the corresponding Sth. The signal intensity distribution of VAT and SAT voxels used for quantification are in yellow and red on the histograms, respectively.

SAT and VAT quantification results from WS images using manual thresholding method in ImageJ, WS, NWS, and nNWS images using QFAT method are compared in Figure 3. All four results of SAT strongly correlate (r > 0.999, P < 0.0001) and there is no statistical difference between them (P > 0.54) (Fig. 3a). For VAT, the results are similar with correlation coefficient of r > 0.994, P < 0.0001 and P > 0.21 for the ANOVA test (Fig. 3b).

Figure 3.

Comparison of fat amount quantification results using ImageJ and QFAT. a: SAT. b: VAT.

Correlation of VAT results from WS images quantified using ImageJ and QFAT methods is shown in Figure 4a. There was strong correlation with slope = 0.992 and R2 = 0.995 between these two quantification methods. In addition, the measurement differences are almost all within the 95% limits in the Bland-Altman plot (Fig. 4b), confirming that the two methods can be used interchangeably. Similarly, the correlation plot of VAT results from WS images and NWS images quantified using QFAT is shown in Figure 5a. There is strong, linear correlation with slope = 1.03 and R2 = 0.989. VAT mean from NWS images was approximately 7.8% higher (134.2 ± 64.4 cm3 versus 144.7 ± 62.3 cm3) but not significantly different (P = 0.36). The ICC is 0.981 which shows strong intra-class correlation. The corresponding Bland-Altman plot also shows consistency between the two methods as shown in Figure 5b. Lastly, the comparison between VAT results of NWS and nNWS using QFAT is shown in Figure 6. There was strong correlation with slope = 0.975 and R2 = 0.988 between these two quantification methods.

Figure 4.

Comparison of VAT quantification results of WS images using ImageJ and QFAT. a: Linear Correlation plot. b: Bland-Altman plot. The upper- and lower-limits represent 95% of differences (mean ± 1.96 standard deviation).

Figure 5.

Comparison of VAT quantification results of WS and NWS images using QFAT. a: Linear Correlation plot. b: Bland-Altman plot.

Figure 6.

Comparison of VAT quantification results of NWS and nNWS images using QFAT. a: Linear Correlation plot. b: Bland-Altman plot.

Statistical results of SAT and VAT amounts quantified from ImageJ and QFAT on WS and NWS images are summarized in Table 1. All SAT and VAT quantification results were relatively consistent with each other. Consistent and comparable results by the three methods provide further reassurance of their validity.

Table 1. Inter-rater Reliability of Abdominal Fat Quantification (n = 60)
MethodSATVAT
ImageJ vs. QFAT (WS images)r = 0.9995, p < 0.0001r = 0.9974, p < 0.0001
ICC (95% CI): 0.999 (0.992,1.000)ICC (95% CI): 0.996 (0.977,0.998)
WS vs. NWS images (QFAT)r = 0.9997, p < 0.0001r = 0.9946, p < 0.0001
ICC (95% CI): 0.997 (0.564,0.999)ICC (95% CI): 0.981 (0.548,0.995)
NWS vs. nNWS images(QFAT)r = 0.9997, p < 0.0001r = 0.9942, p < 0.0001
ICC (95% CI): 1.000 (0.999,1.000)ICC (95% CI): 0.994 (0.990,0.996)

DISCUSSION

Despite progress in abdominal fat tissue quantification by MRI, challenges remain for improved accuracy and speed given the inherent technical and physical factors from the MR techniques and the visceral environment. Fast, fully or semi-automated quantification of abdominal fat on standard T1W MR images have been a major challenge. The contrast between fat and nonfat in NWS techniques is generally poor and image quality is easily compromised by various motion artifacts.

A few automated fat segmentation methods have been developed. Lancaster et al proposed a thresholding approach based on the image histogram and studied two adaptive procedures to seek a threshold signal intensity to separate fat and nonfat voxels (11). However, the accuracy of both procedures, namely the center and valley methods, is compromised by the fact that water and fat signal peaks are overlapping, and the estimations are prone to bias depending on the size of water and fat peaks. A manual thresholding method has been proposed by Liu et al to separate fat from nonfat using a manually selected threshold (21). Liou et al used a fully automated thresholding algorithm to segment both the VAT and SAT (22). In their approach, a threshold value was automatically calculated using a traditional iterative clustering algorithm and applied to the whole image. Positano et al used an approach to perform Gaussian distribution curve fitting for adipose tissues within the abdomen, which generally underestimates VAT amount (23), and may fail when the number of FV fat voxels is small (24). In addition to these limitations, all these approaches lack careful consideration of the large fat amount in PV fat voxels. They are, therefore, subject to large systematic errors and/or subjectivity (23). None of these methods have been cross-validated using a WS image dataset as a reference to study the agreement between them. Our approach is based on a model with both FV and PV fat voxels considered. Large fat amounts within PV voxels are included in the quantification regardless of the fat-to-water contrast. This essentially improves quantification accuracy, and greatly reduces VAT quantification variability.

It has been shown that partial volume fat voxels exist mainly at bulk fat boundaries in abdominal MR images (14). This is expected because of the disseminated nature of intra-abdominal fat, which can be simplified as lean-tissue organs immersed in adipose tissue. Therefore, voxels adjacent to adipose tissue but with intermediate signal intensity are most likely to be PV fat voxels rather than nonfatty tissue voxels. This highly correlated spatial information is an important characteristic that can be used to aid the segmentation procedure. Our novel FCM algorithm used the signal intensity of each voxel as well as the signal intensity of the neighboring voxels to compute the membership function. Owing to the incorporation of spatial information, most of the PV fat voxels are readily identified. Furthermore, the new FCM technique is less sensitive than other methods to noise. This approach is thus particularly advantageous if CNR between fat and nonfat tissues is not high, and/or SNR of the image is poor, leading to overlapping water and fat peaks in the image histogram.

We have validated the feasibility of this approach by comparing the fat quantification results of WS and NWS MR images with the result of WS MR images quantified using a simple thresholding method with ImageJ. It has been reported that fat quantification on WS MR images suffer less from VAT amount under-estimation and variability since both FV and PV fat voxels can be easily included in the quantification with minimal nonfatty voxel contamination (14). In this study, our approach resulted in statistically the same quantification in WS MR images as ImageJ. In addition, it also resulted in slightly larger VAT quantification (7.8%) without statistically significance (P = 0.36) on NWS images compared with WS images. This overestimation is expected, however, because blood within vessels as well as bright artifacts due to blood flow may be quantified as FV or PV fat voxels using current methods. In addition, short T1 water-components within bowel having bright signal on NWS images are quantified as fat, which is suppressed on WS images. Residual motion artifacts due to respiratory, peristalsis, and bulk motion tend to be more severe on NWS images, which all contribute to larger VAT amount from NWS images.

The combination of the novel FCM method with threshoding has additional merits compared with traditional methods. First, this method does not presume Gaussian peak distribution of either the water or the fat peak on the gray-level histogram, and does not require curve fitting for fat quantification. This may be potentially important for images with small amounts of VAT and, therefore, small number of FV fat voxels. In these images, the fat peaks on the histograms may not exist and cannot be easily identified using automatic algorithms. It may also be important when lean tissue organs have different signal intensities and Gaussian distribution does not represent the nonfat tissue signal distribution. Second, the method proposed here is a fast and fully automated procedure that can be used in unsupervised, cost-effective abdominal fat quantification. The current implementation of this method takes only approximately 1 s to finish the segmentation of ten NWS images on a regular personal computer. Third, the proposed procedure can be readily applied to both WS and NWS MR images without further parameter optimization. Relatively consistent quantification results can be obtained using the same algorithm. Although not directly validated in this work, it is speculated that this method is less sensitive to quantification variability due to different fat-to-water contrast compared with traditional methods because of the built-in FCM that better accounts for PVE contributions. Overall, this method can potentially be applied to multi-site and/or longitudinal clinical or epidemiological studies when consistent fat quantification is needed for MR images with a wide range of image quality.

In this work, we focus on presenting the methodology and overall feasibility of a VAT quantitative and automated method that extend and improve on prior work on adipose tissue segmentation. A limitation of the proposed method is that the signal differences from different abdominal organs are not considered, although its impact is largely averaged out when multiple images to cover more abdominal organs are included for fat tissue quantification. This new method provides increased speed, automation, and most importantly the needed accuracy of the measurement. However, more subjects with a larger range of age and BMI distribution need to be included in future studies to evaluate the robustness of this automated fat quantification approach. The potential to reduce quantification variability from MR images of different MR imaging techniques, fat-to-water contrast, and noise level also need to be further validated. Future studies to compare the performance of this method to other existent methods are also warranted.

In conclusion, the proposed two-step procedure to first generate pWS MR images followed by an automated signal intensity thresholding is a feasible approach for fast and automated abdominal fat segmentation that takes into consideration PVE contributions. The new FCM clustering method, which combines voxel signal intensity with spatial information, has advantage over traditional methods because most PV fat voxels can be identified to reduce VAT quantification variability. It is, therefore, a nonsubjective, reproducible, and fully automated segmentation method for abdominal fat quantification on standard NWS and WS abdominal MR images. This will ultimately make it a cost-effective approach to perform large-scale fat quantification on clinically acquired, WS or NWS human abdominal MR images. Potential applications for such a method include longitudinal studies where fat quantification and distribution changes need to be monitored; pharmaceutical interventions in metabolic syndrome research and life style modifications; and, improved drug development/testing and disease treatment where body fat can serve as a biomarker.

Ancillary