To introduce and evaluate the performance of an automated fat quantification method for water-saturated magnetic resonance images.
To introduce and evaluate the performance of an automated fat quantification method for water-saturated magnetic resonance images.
A fat distribution model is proposed for fat quantification on water saturated magnetic resonance images. Fat from both full- and partial-volume voxels are accounted for in this model based on image intensity histogram analysis. An automated threshold method is therefore proposed to accurately quantify total fat. This method is compared to a traditional full-volume-fat-only method in phantom and human studies. In the phantom study, fat quantification was performed on MR images obtained from a human abdomen oil phantom and was compared with the true oil volumes. In the human study, results of the two fat quantification methods of six subjects were compared on abdominal images with different spatial resolutions.
In the phantom study, the proposed method provided significantly more accurate estimations of true oil volumes compared to the reference method (P < 0.0001). In human studies, fat quantification using the proposed method gave much more consistent results on images with different spatial resolutions, and on regions with different degrees of partial volume averaging.
The proposed automated method is simple, rapid, and accurate for fat quantification on water-saturated MR images. J. Magn. Reson. Imaging 2007;26:738–746. © 2007 Wiley-Liss, Inc.
HUMAN ADIPOSE TISSUE is a highly active metabolic and endocrine organ and is integrally involved in coordinating a variety of biological processes including energy metabolism, neuroendocrine function, and immune function. Excessive human body fat distribution in humans is closely correlated with increased risk of cardiovascular and metabolic diseases such as atherosclerosis, hypertension, and non-insulin-dependent diabetes mellitus (1–3). In particular, the amount of intra-abdominal fat (or visceral fat) is regarded as an important risk factor of these diseases. Therefore, traditional simple approaches to quantify obesity based on anthropometric parameters (e.g., body mass index [BMI], and waist-to-hip ratio) do not give the best correlated risk factors of certain diseases (4). Instead, human body fat imaging can be an effective diagnostic tool since both the absolute amount of fat and the anatomical fat distribution can be derived from images. Furthermore, monitoring the change of fat distribution in the human body longitudinally for patients with metabolic diseases, or cardiovascular disease, after pharmaceutical intervention or life style change, is also of great importance for drug development and disease treatment (5).
Currently, computed tomography (CT) and magnetic resonance imaging (MRI) are the commonly used imaging modalities for human body fat distribution measurement (6, 7). CT can generally lead to high-resolution images with very short scan duration. However, the ionizing radiation exposure to subjects limits its application in longitudinal studies and on young subjects. MRI has gained attention due to its capability of generating high-quality tomographic images without ionizing radiation (8). Furthermore, fat has a relatively short longitudinal relaxation time (T1) compared with most lean tissues. Therefore, bright fat signal is easily obtained using well-known MR sequences such as T1-weighted spin echo (T1W SE) or its fast variations (e.g., T1-weighted turbo spin echo [T1W TSE]). Despite these advantages, rapid and accurate fat quantification from typical MR images has been a challenge due to several factors. First, the image quality is usually compromised by a number of motion artifacts since imaging using MRI is slower than with CT. Patient bulk motion, respiratory motion, and involuntary peristalsis can all contribute to these artifacts. Second, images obtained with the traditional methods may have relatively low signal-to-noise ratio (SNR) and low contrast between fat and nonfat. This is particularly true for abdominal fat imaging wherein several images have to be acquired in a limited time window as breathhold is mandatory to minimize respiratory motion. Third, images obtained using traditional methods usually suffer from blood flow artifacts unless flow suppression is applied. Due to the nonideal image quality from MRI, manual contour drawing still seems to be a popular approach for fat quantification, although it is slow, and suffers greatly from inter- and intraobserver variations. In addition, fat volume influenced by partial volume effects (fat partially occupies a voxel) can not be accurately evaluated.
Recently, a novel MR imaging sequence, three-dimensional (3D) water-saturated balanced steady-state free precession (WS b-SSFP) has been proposed for rapid and high-quality fat imaging (9). This technique makes use of ultrafast b-SSFP (also known as balanced fast field echo [b-FFE], true fast imaging with steady precession [TrueFISP], and fast imaging employing steady state acquisition [FIESTA]) for rapid data acquisition, in combination with water saturation, to achieve high contrast between fat and water. Using this method, high-quality, fat-only images can be generated with a very short scan duration. The images it generates suffer from significantly fewer motion artifacts and no blood flow artifacts. The contrast between fat and nonfat tissue is also greatly improved compared with traditional approaches such as T1W TSE (9), and water-suppressed T1W TSE (10). The comparison between images obtained with T1W TSE and WS b-SSFP is demonstrated in Fig. 1. The non-WS T1W TSE image (Fig. 1a) has much lower contrast between fat and nonfat, which is also shown in the corresponding intensity histogram (Fig. 1b). On the WS b-SSFP image histogram (Fig. 1d), the fat signal peak is well separated from the background noise, as well as from suppressed water signal peak (also close to zero). Therefore, a simple threshold approach based on the image histogram might be feasible to offer an automated and rapid fat quantification. Additionally, since the majority of partial volume fat signal is delineated on the histogram, the threshold method might also achieve very accurate fat quantification with partial-volume fat volume signal incorporated.
It is the goal of this work to discuss a novel automated fat quantification method on WS b-SSFP abdominal fat images. A simple fat distribution model is first discussed, which takes into account both full- and partial-volume fat pixels in an image. An automated method to evaluate fat volume using the proposed fat distribution model is tested and compared with a traditional full-voxel only quantification method in both phantom and human studies.
Due to the limited spatial resolution of any imaging modality, adipose tissue can either occupy a full voxel (full-volume fat), or coexist with nonfat tissue (water), air, or both, in a voxel (partial-volume fat). This is particularly true for MRI, which is well-known to generally have much lower spatial resolution than CT. The spatial resolution of MRI is mainly limited by SNR and imaging duration. In an ideal water-saturated image (neglecting residual water signal, 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, and the signal intensity from a voxel partially filled with adipose tissue (the other part is filled with water, air, or both) would be linearly proportional to the volume ratio of fat in that pixel. Thus, a simple fat distribution model is defined as follows: Fat in an image can be in either full-volume fat voxels, or partial-volume fat voxels. For voxels with full volume fat, they will ideally give a single signal maximum (Smax, with the number of pixels N1). For partial-volume fat voxels, the signal is linearly proportional to the volume fraction of fat (0–Smax).
Fat in full-volume fat voxels is mainly the static bulk fat, such as subcutaneous fat. Ideally, the number of full-volume fat voxels (N1) can be demonstrated on the image intensity histogram as shown in Fig. 2a. Partial-volume fat pixels are located mainly at the adipose tissue interface, and the lower signal can be due to partial volume fat filling, or due to intestinal motion. Due to the complicated fat distribution and the possibility of intestinal motion, the distribution of the partial-volume fat signal within the histogram is not predictable. However, if the number of pixels in an image is large enough, the probability of a partial volume fat pixel to have a certain ratio of fat should be close to the probability of another ratio. Therefore, partial-volume pixel signal distribution is almost constant over all signal intensity less than Smax on the histogram. Thus, a uniform (or rectangular) distribution as shown in Fig. 2b can be used to approximate the partial volume fat signal distribution. The pixel density (N2) in the histogram is image-specific, and should be derived from the image histogram. This assumption has to be validated experimentally.
Considering both full- and partial-volume fat signal, the image histogram is shown as the dashed lines in Fig. 2d 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 Fig. 2d, which is derived from the ideal histogram, convolved with a normalized Gaussian kernel representing Gaussian noise (with standard deviation [SD] = σ) (Fig. 2c). The background noise and suppressed water signal is also illustrated as the peak with signal intensity close to zero. Assuming the SNR of fat (defined as SNR = Smax/σ) is much greater than one, the histogram density (N2) of partial-voxel pixels can be derived directly from the experimental histogram. Also, N1, which represents the number of full-volume voxels, can be determined via curve fitting.
Overall, the total fat volume of an image can be calculated as:
where Vtotal is the total volume of fat on the image, Vfull is total volume of fat in all full-volume fat voxels, and Vpartial is the total volume of fat in partial-volume fat voxels. Because on average, partial-volume voxels are half filled, theoretically fat volume can be calculated as follows:
where Vvoxel represents volume size of a single image voxel. Therefore, the signal threshold which best separates fat from nonfat is:
Note that N1 + (Smax – Sth) × N2 is the integration of the histogram curve above Sth, representing the total number of voxels with signal intensity higher than Sth (Fig. 2d). Therefore, a fat-only binary image can be generated by setting the voxels with intensity higher than Sth to 1, 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 N1 and N2 from curve-fitting does not greatly influence the final fat quantification. To include both full- and partial- volume fat in fat volume measurement, Sth should be best set to Smax/2.
The goal of the phantom study was to compare accuracy of fat quantification methods on WS b-SSFP images. This was achieved by making use of a human abdomen phantom filled with known volume of vegetable oil. Briefly, the human abdomen phantom had dual-layered concentric cylinders with internal/external oil volumes of 3.16 liters/6.34 liters, simulating an intraabdominal fat to subcutaneous fat ratio of 0.498. In addition, the inner cylinder had an intravenous (IV) fluid bag, a bottle, and a tube, each filled with water fluids with T1, T2 similar to abdominal organs. MR experiments were performed on three different days using a 3D WS b-SSFP pulses sequence as described elsewhere (phantom study in Ref.9). On each day the scan was repeated with the phantom relocated, leading to six datasets. The internal, external, and total oil volumes were then measured using a semiautomated computer software package (Q-Fat) developed with IDL (Interactive Data Language; Research Systems, Boulder, CO). In brief, a signal threshold (Sth) was determined to generate a binary fat-only image. Two user-defined contours were then drawn to enclose the whole abdomen phantom, and to separate internal and external oil. The total oil volume and the internal oil volume were then calculated automatically by counting the number of fat voxels enclosed by each contour times the voxel size. External oil volume was calculated by subtracting internal volume from the total volume.
Two different methods were used to determine the signal intensity threshold for binary fat-image generation. The first method used a threshold of Smax – 2.35σ (where σ is the SD of the fat peak). It is approximately the minimum fat signal of the main fat peak in the histogram; therefore, only full-volume fat voxels were included in fat quantification. The threshold determined using this method is referred to as Thf, and the method is called the Thf method in the rest of this work. In the second method, the fat signal threshold of Smax/2 was used, and Smax was determined automatically via curve-fitting based on the model described earlier in the “Quantification Method” section. Therefore, both full- and partial-volume fat voxels were included for fat quantification. The threshold determined using this method is referred to as Thfp, and the quantification method is called the Thfp method. A single factor analysis of variance (ANOVA) was used to determine whether the two approaches predicted significantly different oil volumes and ratios. A P value of less than 0.05 was considered statistically significant.
The aim of the human study was to compare fat quantification differences of the two methods on images with different degree of partial volume effect. Six subjects (four males, and two females, BMI range from 22 to 30 kg/m2) were recruited. Written informed consent was obtained from all participants before MR examinations. All experiments were performed on a 1.5-T whole-body clinical MR scanner. Eight slices were planned centered at the L2–L3 level using a similar WS b-SSFP fat imaging sequence as described (9). In addition, higher spatial resolution images were acquired to achieve more accurate fat quantification. The measured matrix size was 400 × 400 reconstructed to 512 × 512, and slice thickness was 3 mm. Other parameters included: TR/TE/FA = 3.6 msec/1.36 msec/40°, TFE factor = 200. The scan duration was 17 seconds, and eight images were acquired with expired breathhold. Only the central four slices (referred to as S1–S4) were used for data processing. To simulate partial volume effect, images with 6-mm-thick, and 12-mm-thick slice datasets were reconstructed from the original four slices. The two 6-mm-thick slices were generated by averaging S1 and S2, and by averaging S3 and S4, respectively. The 12-mm-thick slice scan dataset had only one image, which was generated by averaging all four 3-mm-thick slices (S1–S4). The reason to reconstruct thicker slices from thinner slices instead of performing extra scans is to avoid true fat volume variations between several different scans due to motion (e.g., bulk, respiratory, and peristaltic motion). Since water signal is effectively suppressed in WS b-SSFP, no destructive fat–water signal addition effect is expected. The simple averaging method therefore should lead to images which are close to images generated by independent scans with thicker slices.
Fat quantification for each image (four 3-mm-thick slices, two 6-mm-thick slices, and one 12-mm-thick slice) was then performed using the same Thf and Thfp methods described in the phantom study. The same two predefined contours were applied to images with different slice thickness to avoid operator-induced inconsistencies. Intra-abdominal fat (IAF, i.e., visceral fat) and subcutaneous abdominal fat (SAF) were then calculated using the same procedures as in the phantom study. The resultant total fat volume for the 12-mm-thick abdomen region in each dataset group was calculated by summing up the area of fat of each slice times the corresponding slice thickness. Fat quantification results on images of different slice thicknesses, and with different postprocessing methods were compared to investigate the performance of the two fat quantification approaches.
A representative result from the phantom study, obtained using the WS b-SSFP sequence, is shown in Fig. 3. Figure 3a is an original phantom image. Figure 3b–d are the derived, signal intensity-segmented images, demonstrating the voxel distribution of different signal intensities shown in the histogram (signal intensity range b–d in Fig. 3e). It is seen that the corresponding water signal on the histogram is close to zero (noise level), validating the expectation that the water signal is effectively suppressed (Fig. 3b). Partial-volume fat (oil) voxels are mainly located at the edge of bulk volume of fat (Fig. 3c). Bulk fat voxels also correspond to the main fat signal peak on the histogram (Fig. 3d). The curve fitting result based on the fat distribution model is also illustrated in Fig. 3e. The main (full-volume voxel) fat signal peak also agrees well with a Gaussian-shape peak. Partial-volume fat signal between the fat peak signal and zero agrees well with a uniform distribution. This is consistent with the previous uniform distribution assumption for partial-volume fat voxels.
The phantom oil volume measurement results using the two different threshold methods are presented in Table 1. It is apparent that the two processing methods resulted in significantly different accuracies on phantom oil volume estimation. The method using simple full-volume voxel number counting method (Thf method) led to an average underestimation of 10.8% for the internal oil volume, and 4.8% for the external oil volume measurement. The fat quantification method proposed herein which considered both full- and partial-volume fat (Thfp method) generated estimates which were much closer to the true volumes. The mean deviations of internal and external oil volumes are only 0.1% and 2.6%, respectively. The relatively higher overestimation (2.6%) for the external oil volume is most likely due to B0 inhomogeneity and gradient nonlinearity, since a large 3D volume (27 × 27 × 20 cm3) was included in the same scan without table motion.
It is clearly demonstrated that partial-voxel oil volume comprises a considerable portion of the total oil volume, and significant volume underestimation will be introduced if this effect is overlooked. For human abdominal fat quantification, it is speculated that this effect is enhanced due to the more complicated anatomic structure, much larger voxel size, and severe motion-induced signal averaging.
A representative result from the human study, using the WS b-SSFP sequence, is shown in Fig. 4. Figure 4a is an original MR image. Figure 4b and c are the corresponding resultant fat-only images using Thf and Thfp methods, respectively. Bulk fat voxels also correspond to the main fat signal peak in the histogram (Fig. 4b). Figure 4c also shows more partial-volume fat voxels mainly located at the edge of bulk fat compared to Fig. 4b. In Fig. 4d, two predefined contours are overlaid on the binary, fat-only image for regional fat quantification. The curve fitting result (Fig. 4e) based on the fat distribution model is also similar to that found in the phantom study. However, a much larger number of partial-volume fat voxels are found compared to the phantom image. This is mainly due to the more disseminated fat distribution in human abdomen. The good agreement of curve fitting is also consistent with the previous uniform distribution assumption for partial-volume fat voxels.
The representative original images with high spatial-resolution from a volunteer, and the reconstructed images with lower spatial resolution, are shown in Fig. 5. The original images suffer much less from blurring and partial-volume effects than the reconstructed, thicker slices. The fat quantification results, using the two methods on images with different spatial resolution, are shown in Table 2. Consistent with phantom study results, the Thf method provided much lower mean fat volumes compared with Thfp method, for both IAF and SAF, and for images with different slice thicknesses. The underestimation rises as slice thickness increases for both IAF and SAF. In addition, IAF was even more severely underestimated than SAF with the Thf method, most likely due to the different degrees of partial volume effect of the two fat populations. The difference reached –27.5% when the slice thickness was 12 mm.
|Subject number||Thickness = 3 mm||Thickness = 6 mm||Thickness = 12 mm|
Figure 6 demonstrates the intrasubject IAF quantification variations on images with different slice thicknesses for the Thf and Thfp quantification methods. Generally, measured IAF volumes are lower if thicker slices are used for Thf method, and much more consistent fat quantification can be achieved using Thfp method. The average coefficients of variation (CV, defined as the SD of the three measurements as a percentage of the mean) of IAF measured with Thf and Thfp are 7.8% and 1.5%, respectively. Therefore, the new processing method is more robust when spatial resolution of an image is reduced.
It has been demonstrated that the proposed fat quantification method based on a more accurate fat distribution model offers an easy and accurate way for fat quantification. This method takes advantage of the superior image quality obtained using WS b-SSFP (9). Therefore, fat quantification difficulties due to lowfat-to-nonfat contrast are minimized. Furthermore, very accurate fat quantification can be achieved since the partial-volume effect inherent in any imaging modality is also incorporated.
This method is therefore superior to the traditional methods for fat quantification. Most of those approaches were developed for non-water-suppressed T1W images which suffer from low contrast, motion and flow artifacts. Despite the difficulties, several automated or semiautomated approaches have been proposed. Lancaster et al (11) proposed a threshold method based on the image histogram and studied two adaptive procedures to seek a threshold signal intensity to separate fat and nonfat pixels. However, the accuracy of the two procedures they studied, 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, particularly for subjects with very low or high visceral fat volume compared to nonfat volume. Positano et al (12) have recently proposed an unsupervised method for the assessment of abdominal fat volume. They performed Gaussian curve fitting for the fat signal distribution on the histogram using those pixels with signal intensity higher than the average fat signal. This method works well if there is no contamination from water pixel signal in the region of curve fitting, but is still inaccurate if there is relatively low visceral fat volume compared to nonfat volume. Furthermore, the existing methods do not consider the influence of the partial-volume effect on fat volume quantification. Therefore, they are inherently subject to high systematic error.
Due to the highly complicated anatomic components and the disseminated nature of intra-abdominal fat, partial-pixel fat volume may contribute greatly to the total abdominal fat. The partial-volume effect is further confounded by blurring from the intestinal motion or the artifacts in the human abdomen (13). It is particularly problematic for area/volume measurement using MRI, as MRI generally has much lower spatial resolution than CT (14). Therefore, a quantification method that can accurately evaluate partial volume fat will greatly improve the overall fat volume measurement accuracy, especially for normal or under-weight subjects.
Improved MR techniques such as 3D WS b-SSFP can improve the spatial resolution which leads to less severe partial volume effect. However, the ultimate spatial resolution of MRI is inherently limited by SNR and scan duration. Due to the large field of view (FOV) associated with human abdomen imaging and the need for data acquisition during patient breath-hold (to minimize motion artifacts), image voxel size of MR is usually large compared with CT. The simple fat distribution model includes fat from both full- and partial-volume fat voxels; therefore pixel signal distribution can be correlated with the fat volume based on an image histogram analysis. Fat quantification using this method can be much easier, and more accurate, compared with the traditional methods where only full-volume fat pixels can be considered. The accuracy of this method on fat volume measurement is validated in the phantom study. In the human study, much lower fat volume estimation is obtained using the full-volume-fat-only method compared with the new method, mainly due to the influence of partial-volume voxels. This is particularly true for IAF due to the more complicated intra-abdominal fat distribution. In addition, fat quantification using the new method, when applied to images with different slice thicknesses, is much more consistent than that obtained using the traditional method. These results are most easily explained by the insensitivity to partial-volume effects of the new method.
There are several limitations of this study. First, although partial-volume fat contribution is considered in the novel model, truncation artifacts (Gibbs effect) due to the limited in-plane spatial resolution are not considered in this fat quantification method (15). Truncation artifacts will generally introduce slight overestimation of fat volume. However, in the human study, in-plane resolution was kept the same for all three slice thicknesses. It therefore could not be the reason for differences between the quantification methods. The second limitation of this study is that the point-spread function depending on the specific imaging pulse sequence (e. g., turbo spin echo, water-saturated b-SSFP) may lead to image blurring, and is not considered here. Another limitation of this study is that, there are other factors which may lead to signal non-uniformity which are not considered. Magnetic field and radiofrequency (i.e., B0 and B1) inhomogeneities can both introduce image signal nonuniformity. These effects may lead to a distorted, non-Gaussian fat peak on the intensity histogram, and therefore an inaccurate Sth estimation. To correct for these effects, post-processing steps such as the ones described in (16–18) can be taken to generate uniform full-volume fat signal. Some of these correction algorithms have been implemented on some clinical MR scanners and may improve fat quantification accuracy. Image histogram may also deviate largely from what is hypothesized in this work for images with fatty liver. Therefore, more sophisticated image processing steps have to be taken to avoid quantification errors. Additionally, further investigation is warranted to validate the robustness of this method on MR images obtained using more clinically available sequences (such as water-saturated turbo spin echo sequences (19), and water-saturated spoiled gradient echo sequences (20)), as those images may have relatively poorer image quality than that of WS b-SSFP images.
In conclusion, the proposed fat quantification method has been validated and offers the following benefits: 1) fat quantification is rapid, nonsubjective, and can potentially be fully automated; 2) fat quantification is much more accurate compared with traditional methods; 3) fat measurements suffer much less from variations due to spatial resolution changes. The improved fat quantification method can thus lead to a much easier and more accurate quantitative monitoring of human body fat distribution. The method is also applicable for use in clinical trials of new drugs, to study fat redistribution under pharmacological influence (e.g., antiobesity drugs or drugs that interact with adipose tissue). Due to the greatly improved fat measurement accuracy and reduced measurement variations, it is likely that fewer subjects will be needed in a study to achieve a certain statistical significance. Therefore, clinical trial of a new drug can be completed with greatly reduced expense and in a much shorter time.