To determine whether a whole-body T1-mapping acquisition method improves the definition of adipose tissue (AT) and simplifies automated AT segmentation compared to an image-based method.
To determine whether a whole-body T1-mapping acquisition method improves the definition of adipose tissue (AT) and simplifies automated AT segmentation compared to an image-based method.
The study included 10 subjects. Two whole-body volumes were acquired from each subject using two different flip angles. Whole-body T1 maps were calculated from each pair of whole-body volumes. AT was automatically segmented from the T1 maps and from the original image slices. The results were evaluated using manually segmented slices as reference.
The T1-mapping method segmented more of the reference AT than the image-based method, with mean values (standard deviations (SDs)) of 87.7(5.1)% and 81.1(5.2)%, respectively. Compared to the image-based method, the T1-mapping method gives better histogram separation of AT in whole-body volumes. The suggested method also provides an output with smaller in-slice AT intensity variations.
The T1-mapping method improves the definition of AT. T1-based analysis is superior to analysis based on the original images, and allows fully automated and accurate whole-body AT segmentation. J. Magn. Reson. Imaging 2006. © 2006 Wiley-Liss, Inc.
If you can't find a tool you're looking for, please click the link at the top of the page to "Go to old article view". Alternatively, view our Knowledge Base articles for additional help. Your feedback is important to us, so please let us know if you have comments or ideas for improvement.
Research into the composition of the human body increases our knowledge of human biology. A goal of major importance is to increase our understanding of the relation between obesity (studied by assessing the adipose tissue (AT) distribution and other tissue components) and metabolic syndrome, with the aim of evaluating the risks of developing type 2 diabetes, dyslipidemia, hypertension, and cardiovascular disease (1–6).
A large number of methods for analyzing body composition are used in vivo (7). In their classic paper in 1992, Wang et al (8) pointed out that body composition should be analyzed on five different levels: I) atomic, II) molecular, III) cellular, IV) tissue system, and V) whole body. Each level has clearly defined components, and the sum of the weights of the components is equal to the total body weight. Methods based on dual-energy x-ray absorptiometry (DEXA), computed tomography (CT), and magnetic resonance imaging (MRI) are considered to give the most accurate body composition measurements (9–11).
DEXA makes determinations possible at both the atomic and molecular levels (levels I-II), and was originally developed for bone density analysis. Nowadays it is also used in body composition studies. The established DEXA technique provides information not only on the total and regional bone mineral content (BMC), but also on the content of soft tissue. The soft tissue is further divided into body fat (BF) and lean tissue mass (LTM). The three-compartment DEXA model (BMC + BF + LTM equals body weight) is attractive because of its low radiation dose and high degree of reproducibility. However, DEXA may yield invalid results for subjects with a body weight above 100 kg (12).
CT and MRI are both tomographic techniques that are used to analyze the body composition at the tissue-system level (level IV) (8) and allows accurate quantification of body tissues (11, 13). CT has the disadvantage of exposing the patient to ionizing radiation, whereas MRI has not been proven to have any adverse effects (14). In CT the Hounsfield value of the tissue in a voxel is mapped to the corresponding image pixel, whereas in MRI there is no absolute correspondence between tissue property and pixel value. MRI also has the drawback that it often yields an inhomogeneous signal intensity, which is more pronounced when a large field of view (FOV) is used.
CT and MRI have been used in multiple studies to measure abdominal AT. The accuracy of the result depends on the number of acquired slices and the slice locations used. The time required for the analysis, as well as the cost, increase with the number of slices used. Numerous studies have been performed to verify the validity of using single-slice CT and MRI to predict total AT volumes; however, no conclusion has been reached (11, 15–19). For a more accurate analysis, or when the effects of regional adiposity (20–24) are sought, the use of multislice imaging is required.
The most accurate results of body composition investigations are derived from whole-body analyses using contiguous slices. This setup, however, is rarely used because of the long acquisition and analysis times that are needed. When body-composition analysis is performed in cross-sectional or interventional studies, which may involve relocation or regional effects on AT, the importance of whole-body analysis is increased (25, 26). In whole-body CT the exposure to ionizing radiation is a limiting factor. For this reason a limited number of carefully chosen slice positions can be used, necessitating interpolation (13). Whole-body imaging requires a large FOV, which may induce artifacts in both CT and MRI. Many studies using whole-body MRI have been reported to require time-consuming analyses (11, 18, 19, 23, 27–35). The most frequently used analytical method is manual segmentation of T1-weighted spin-echo images acquired with interslice gaps.
Automation of the image analysis process is required in order to decrease the analysis time and to reduce inter- and intraobserver variations. The arbitrary unit (au) of the image intensity is a greater problem in automated analysis than in visual analysis. The interslice intensity offsets and the in-slice intensity inhomogeneities also make the automation process more demanding.
The purpose of this study was to investigate a method that hypothetically would improve the definition of AT, by comparing it with a method based on normal imaging. The consequences of the studied method for automated image analysis of whole-body volumes were also evaluated. Instead of using normal imaging, the proposed method calculates a whole-body T1 relaxation map from two imaged whole-body volumes acquired using a gradient-echo sequence with two different flip angles. AT was automatically segmented from the T1 maps and from the normal image slices. For evaluation of the results, manually segmented image slices were used as references.
After they gave informed consent, 10 volunteers were scanned using the T1-mapping protocol in a study approved by the ethics and radiation committees (see Table 1). The subjects were members of two nuclear families that were participating in a larger genetic obesity study (SOS Sibpairs) within the framework of the Swedish Obese Subjects (SOS) project, which aims to determine whether the mortality and morbidity rates among obese individuals who lose weight by surgical means differ from the rates associated with conventional treatment (36, 37). Families of the SOS Sibpair study were ascertained through an obesity discordant sibpair, identified in questionnaires completed by obese probands who volunteered to participate in various obesity or type-2 diabetes intervention studies.
|Age (years) F/M||Height (m) F/M||Weight (kg) F/M||BMI (kg/m2) F/M|
The whole-body image acquisition was performed on a 1.5 T clinical MRI scanner (Gyroscan NT, Philips Medical Systems, The Netherlands), using a spoiled T1-weighted gradient-echo sequence with repetition time (TR) = 177 msec and echo time (TE) = 2.3 msec. Axial slices were scanned with each subject in the supine position with the arms extended above the head to reduce the maximum patient width. A whole-body volume (denoted as Flip80) was acquired using a flip angle of 80°. This volume was used for the image-based analysis. The scan parameters RF power optimization, receiver optimization, and noise level determination were then turned off, which reduced the image quality but ensured a constant MR signal scaling. Two more whole-body volumes were acquired using two different flip angles: 80° and 30° (denoted as Flip80off and Flip30off, respectively). These volumes had suboptimal image quality and were used only to create the T1 relaxation map.
Because of the limitation of the table-top transition, the subjects were imaged in two half-body volumes. The following procedures, in chronological order, were carried out when the whole-body volumes were acquired: lower body part acquisition (Flip80, Flip80off, and Flip30off); subject repositioning; and upper part acquisition (Flip80, Flip80off, and Flip30off). Each half-body volume was scanned using six stacks, each containing 25 contiguous 8-mm-thick slices. Each stack was acquired in 18 seconds, and an interleaved slice order was used. An FOV of 530 mm was used, and three of the total 12 stacks were acquired with breath-holding. The total investigation time was about 45 minutes.
The signal equation for a spoiled, steady-state, gradient-echo sequence, the “FLASH equation” (38), is given as
Equation [ 1] shows how the signal intensity relates to the flip angle α, TR, TE, T1, and T2*. Acquisition of two images using two different flip angles means a two-point sampling of this function for each voxel. From these samples the T1 value (in msec) is calculated from
In Eq. [ 2], T1 is given as a function of TR and the intensities I1 and I2 from the two images acquired with use of the two flip angles α1 and α2, respectively.
The two half-body volumes were first rigidly registered using the software package ITK (39). To perform the registration, translations in three dimensions were used, where the sum of the squared image intensity differences was the optimization criterion.
The registration was performed to objectively position the MRI slices used in the evaluation. The positions were determined from CT slice coordinates, which were derived from a CT acquisition that was carried out in conjunction with the acquisition of the MR images. When the MRI positions were derived, the knee slice position was used as reference. The initial registration position in the head–feet direction was visually estimated and used as input to the registration algorithm.
The registration was performed on the Flip80 volume because of its higher signal-to-noise ratio (SNR) compared to the Flip80off volume and its superior tissue contrast compared to the Flip30off volume. The same registration parameters were used to register all whole-body volumes. During registration only the posterior half of the body was used in the optimization, in order to reduce the effects of abdominal motion.
Each body was segmented from the background using the region-growing algorithm illustrated in Fig. 1. One seed point was set inside the body in the Flip80 volume and a region was grown where pixel values were greater than a threshold. The threshold was chosen to include the body and exclude the motion artifacts outside the body. Since the region-growing yields an object with holes, i.e., void areas, from low image intensity areas, a single step of morphological closing (18-neighborhood) was applied to “close” the mouth of the subject and thus create an object surface that was free from holes. Another seed point was set outside the body and the background was grown until the first object surface segmented was reached. Thus a binary object was created from the body, which minimized the effects of motion artifacts outside the body. The T1 maps were generated by calculating T1 values inside the segmented bodies, using Eq. [ 2] and data from the Flip80off and Flip30off volumes.
Because of interslice intensity offsets in the Flip80 volumes, slice-wise thresholding of the axial slices from both the Flip80 and the T1-mapped volumes was performed for the segmentation of AT. To find the threshold value for use in AT segmentation, two Gaussian curves were fitted to the slice-wise histogram data. Maximum histogram coverage and minimum overlap between the Gaussian curves were used as optimization criteria for the Gaussian fitting algorithm. For each iteration in the algorithm a cost function was calculated and minimized. The cost function was calculated by adding the squared overlapping area of the fitted Gaussian curves to the sum of the squared differences between the true histogram and the fitted Gaussian curves. Initial and boundary conditions were used to enhance the performance. The threshold was chosen as the point of intersection between the two Gaussian curves (Fig. 2c).
The result from the registration of the two half-body volumes was compared, in the head–feet direction, with the manual result obtained by an experienced radiologist who visually optimized the anatomical correctness of the posterior half of the body.
The evaluation of the results in this study was based on a reference created from nine slices from each subject, manually segmented by an experienced radiologist. The nine slices were chosen to coincide with nine of the slices acquired with the CT scan method described by Chowdhury et al (13). The following slice positions, according to Chowdhury et al's annotation, were used: forearm; skull bending (denoted skull); sternoclavicular joint, lower border (shoulder); lowest sinus (liver); crista, upper border (L4 level); sacroiliac joint, lower border (hip); thigh; knee; and calf. The positions of the segmented slices in one subject are displayed in Fig. 3. The manually segmented AT was separated into four classes: subcutaneous AT (SAT), visceral AT (VAT), white bone marrow, and fat in the intestines. During the segmentation process the original whole-body Flip80 volume was available to aid identification of different tissues.
To assess the different in-slice AT variances in the Flip80 and T1-mapped slices, the reference areas, manually marked as AT, were used to mask AT from all Flip80 and T1-mapped volumes. AT histograms were then constructed from the Flip80 and T1-mapped “nine-slice volumes” for all subjects. A Gaussian curve was fitted to every AT histogram peak, using the method of least squares, and the standard deviations (SDs) of all Gaussian curves were measured. The means of all Flip80 slice histograms were adjusted to coincide with the mean of the summed slice histograms by adding a constant value to all pixels, in order to avoid effects of interslice intensity offsets when calculating the “whole-body” AT variance. The T1 values corresponding to the center of the T1 Gaussian curves were also measured.
The binary results from the slice-wise segmented AT, from the Flip80 and the T1-mapped nine-slice volumes, were compared with the reference. The numbers of true-positive (TP) and false-positive (FP) segmented AT pixels were counted and compared.
To investigate the origins of the FP pixels (displayed in red in Fig. 4d and e) from both AT segmentation methods, FP pixels were manually divided by an experienced radiologist into the following classes: SAT in the SAT-skin tissue interface, SAT in other tissue interfaces, VAT, white bone marrow, fat in the intestines, and erroneously segmented AT (denoted as “other”). Intermuscular AT was classified as SAT, and cardiac, thoracic, and intra- and retroperitoneal AT was classified as VAT. For the classification, the FP pixel positions were shown in the original images.
The software StatView (SAS Institute Inc., Cary, NC, USA; www.sas.com) was used for the statistical analysis, and Student's t-test was applied to generate all P-values presented. P-values below 0.01 were considered statistically significant.
In six of the subjects the automated registration result of the two half-bodies was in exact agreement with the manually performed registration. In the remaining four subjects a one-slice difference (i.e., ±8 mm in the head–feet direction) was found. The anterior half of the body showed the largest discontinuity effects, caused by soft-tissue motion resulting from the repositioning of the subject between the acquisitions of the two half-body volumes. Subjects with more abdominal AT were seen to be more affected by soft-tissue motion. A maximum area difference of 15.8% was seen across the registered border in one subject.
The calculated T1 maps yielded a better histogram separation of AT from lean tissue (LT) than the Flip80 volumes (Fig. 3).
The SDs of the Gaussian functions fitted to the AT peaks, expressed as the mean (SD), in the Flip80 and the T1-mapped volumes were found to differ significantly 327(86.4) and 86.9(12.0), respectively; P < 0.001). The T1 peak was found to be centered at 203(28.2) msec.
The result from the automatic AT segmentation in one of the volunteers is presented in Fig. 4. The figure shows the nine analyzed slices from the Flip80 volume, the T1-mapped volume, and the reference. It also shows the automated segmentation results from the Flip80 and the T1-mapped volume.
The T1-mapping method gave higher TP values (P < 0.001), but it also gave higher FP values (P < 0.002). The mean differences in TP and FP between the Flip80 and the T1-mapping technique were 5.2% and 13.8%, respectively. The total amount of AT segmented in the reference was 18120 cm2 (14962 cm2 SAT, 3158 cm2 VAT). The results from the automatic AT analysis and the evaluation of the origins of the FP pixels are displayed in Table 2. The evaluation showed that the largest proportions of FP from the T1-mapped volumes came from SAT in the SAT-skin interface, SAT in other interfaces (mainly SAT-muscle), and bone marrow (26.2, 33.2, and 21.5%), respectively, whereas the major proportion (70.0%) of the Flip80 FP came from true erroneously segmented tissue. The T1-mapping method increased the segmented amount of AT in bone marrow and intestines by 119% and 140%, respectively. True erroneously segmented AT (denoted as FP* in Table 2) accounted for 0.3% of the total amount of AT in the T1-mapped volumes, but for 4.5% in the Flip80 volumes.
|Measurement (SAT, VAT)||Evaluation of FP (% of FP)|
|Flip80||83.3 (4.43)||5.45 (11.2)||4.55 (9.35)||18.8 (30.9)||3.91 (5.85)||1.95 (4.26)||8.40 (9.93)||6.18 (12.7)||70.0 (27.8)|
|T1||88.5 (5.10)||19.3 (5.20||0.29 (0.20)||26.2 (5.52)||33.2 (5.89)||10.1 (2.39)||21.5 (5.07)||7.57 (2.24)||1.48 (0.97)|
|Slice||Reference total AT||Flip80 (cm2)||T1 (cm2)|
|Total AT||TP||FP*||Total AT||TP||FP*|
|1 Forearm||24.7 (7.62)||11.6 (7.74)||10.5 (4.87)||0.04 (0.06)||21.9 (8.33)||15.0 (4.33)||0.79 (0.65)|
|2 Skull||60.9 (20.3)||34.8 (15.7)||34.3 (13.1)||6.70 (20.0)||63.8 (27.9)||42.1 (11.3)||4.06 (3.82)|
|3 Shoulder||263 (114)||219 (93.1)||218 (23.3)||30.1 (94.6)||273 (121)||230 (18.7)||0.47 (1.21)|
|4 Liver||273 (117)||219 (113)||214 (35.0)||69.9 (147)||261 (131)||217 (32.8)||0.00 (0.00)|
|5 L4||469 (174)||426 (177)||418 (27.0)||0.11 (0.29)||453 (179)||419 (26.0)||0.09 (0.29)|
|6 Hip||376 (126)||320 (118)||318 (13.9)||0.06 (0.12)||385 (129)||346 (12.8)||0.00 (0.00)|
|7 Thigh||248 (87.6)||220 (92.9)||215 (19.4)||0.90 (0.51)||276 (91.4)||238 (9.19)||0.00 (0.00)|
|8 Knee||65.7 (23.6)||53.2 (22.5)||52.1 (8.27)||0.58 (0.66)||77.3 (24.6)||61.2 (3.33)||0.00 (0.00)|
|9 Calf||32.3 (14.1)||22.7 (14.9)||19.8 (7.46)||0.36 (0.48)||39.3 (15.3)||28.1 (2.72)||0.00 (0.00)|
|Subject (M/F)||Reference/Flip80/T1 (cm2)|
|SAT||VAT||White BM||AT in intestines|
We have shown that for segmentation of AT the analytical method based on T1 mapping is superior to that based on Flip80 images. The T1-mapping method shows a more distinct histogram separation of AT from LT in whole-body volumes and gives an output with less pronounced tissue intensity variations. The output is also provided on an absolute scale (in msec) instead of au.
Jin et al (31) reported a higher proportion of TP values, about 92%, from their hybrid AT segmentation method. However, their method analyzes only six slices of the abdomen and requires user interaction (as opposed to our automated approach), which results in a small degree of operator variability.
There are many sources of error in the calculation of the T1 map, including image noise, nonoptimal breath-holding, patient movement between the acquisitions of the different volumes, repositioning of the patient, and the fact that only two sampling points are used to calculate the T1 value. Despite all these sources of error, the T1-based method gives better results, probably because it uses a relative signal intensity change as opposed to the absolute signal intensity used in the image-based method, and also because it uses more information than the image-based method.
For more robust T1 mapping, more points from the T1 relaxation curve can be sampled. However, this will increase the total scanning time required. A modified implementation in which the same stack is acquired without a table-top transition in between the acquisitions of the different flip angles might reduce the artifacts from patient motion, as well as the relative acquisition time required.
The need for a fast whole-body imaging sequence, limited by necessary breath-holding, led us to choose a gradient-echo sequence. The flip angles of 80° and 30° were chosen based on the following criteria: a large separation in flip angle, a high SNR, and a large difference in intensity variations in LT and AT between the images acquired with the different flip angles.
If a multiplicative bias field is assumed to affect a position in both the Flip80off and Flip30off volumes with the same factor, it can be seen in Eq. [ 2] that this will not affect the calculated T1 value. This study also shows that the in-slice inhomogeneities were smaller in the T1-mapped volumes than in the Flip80 volumes. The presence of T1 inhomogeneities is believed to be due to inhomogeneous B0 and B1 fields and patient-induced inhomogeneities.
The automatic thresholding, with fitting of Gaussian curves to histogram data, was more successful for the T1 volumes due to the better histogram separation of the AT. In one Flip80 volume the algorithm failed to separate the AT pixels from the liver pixels. In one subject the algorithm also segmented the brain as AT in one Flip80 volume. These problems might be approached by automatic segmentation and deletion of the liver and brain before the automated thresholding. The manual analysis of the L4 slice in subject 3 was difficult. As shown in Table 4, this affected the white bone marrow figures measured in the Flip80 volume.
The automatic thresholding of the T1 volumes can probably be performed volume- or stack-wise with a similar result. In this study it was carried out slice-wise to increase the similarity to the Flip80 analysis. It is possible that in future analyses a whole-body threshold can be used, after its accuracy is verified, to reduce the analysis time. It is also possible that automated segmentation of the muscle, liver, and brain can be achieved with the use of T1-mapped volumes.
When the registration results were evaluated, mismatches were observed in the anterior parts of the trunk. These can be attributed to differences in the abdominal shape before and after the change between “feet first” to “head first” on the patient table. This will cause an error when calculating total tissue volumes. The largest effects can be expected in the abdominal SAT (and VAT) volumes since these compartments are most likely to change shape due to body movement. In a single-slice study, these effects of motion will influence the result. In a whole-body analysis, it is recommended that the whole body be scanned without patient repositioning or, when this is not possible, that the overlap between the volumes be positioned in a region less likely to relocate tissue (e.g., the thighs).
The generation of the reference is not optimal, since only nine slices per volunteer are segmented, and it relies on the assumption that the manually-segmented images are correct and free from artifacts. However, the positions of the slices cover the whole body, using a well defined and verified scheme, and the total sum of 90 slices will probably give a reliable result. Furthermore, the use of the Flip80 volume when creating the reference will most likely benefit the automated Flip80 method more than the T1 method. T2* effects will decrease the image intensity in bone marrow and intestines, while the calculated T1 value will be less affected, thus leading to systematic FP segmentation of bone marrow and intestines. The FP* values, as shown in Table 2, are derived from the evaluation of FP pixel origins and represent the number of segmented pixels that clearly do not contain AT, e.g., those containing only muscle, liver, or brain. With the T1-mapping method the position of the FP pixels that are judged to belong to the classes SAT and VAT lie on the borders of the AT segmented in the reference, thus causing a systematic difference in the delineation of AT. In body composition investigations, the skin, bone marrow, and intestinal fat are of limited interest and might therefore have to be removed from the analysis. The skin can be removed from the analysis by morphological erosion of the binary body. The bone marrow can be removed on the basis of prior knowledge and image analysis tools. Fat in the intestines can be reduced, for example, by instructing the subjects concerning their food intake.
Results of previous studies have motivated the use of automated segmentation of a large number of slices for an accurate body composition analysis (29, 40). Gong et al (29) reported manual analysis coefficients of error to be about 10%, 5%, and 3% in estimates of total muscle and fat volume in muscular dystrophy patients and controls when they analyzed 15, 20, and 35 axial slices, respectively, which justifies the use of large numbers of slices for an accurate analysis. Positano et al (40) reported manual intraobserver variabilities of 0.2% and 11% in segmentation of SAT and VAT, respectively, supporting the use of automated analysis.
In conclusion, T1-mapping gives a better definition of AT than does the use of image intensity, and thus simplifies automated whole-body AT segmentation.