| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Clinical Observations |
1 Department of Radiology, Cappagh National Orthopaedic Hospital, Finglas,
Dublin, 11 Ireland.
2 Vision Systems Group, Dublin City University, Dublin, Ireland.
Received July 6, 2004; accepted after revision October 6, 2004.
Address correspondence to S. J. Eustace.
OBJECTIVE. The purpose of this article is to determine the feasibility of using computer-assisted diagnosis (CAD) techniques to automatically identify, localize, and measure body fat tissue from a rapid whole-body MRI examination.
CONCLUSION. Whole-body MRI in conjunction with CAD allows a fast, automatic, and accurate approach to body fat measurement and localization and can be a useful alternative to body mass index. Whole-body fat analysis can be achieved in less than 5 min.
The accurate determination of a person's total body fat is an important issue in medical analysis because obesity is a significant contributing factor to a variety of serious health problems. The medical literature identifies a wide range of diseases that are closely linked to obesity. Current methods of fat assessment are largely inaccurate, and most current methods of fat determination cannot show regional fat distribution, which is important in defining disease risk. We introduce a method that combines computer-aided techniques with whole-body MRI techniques and enables accurate quantification and visualization of total body fat burden and regional fat distribution. This technique may be important in identifying and treating at-risk populations.
Materials and Methods
Patients
This study included 42 patients (21 men and 21 women) who were recruited
through local in-hospital advertising and a hospital-related sports clinic.
These volunteers included a cohort of international rowers, a group of elite
athletes in whom body fat estimation is of particular importance as
individuals are weight restricted. Local institutional board approval was
obtained. Informed consent was acquired from each subject, and his or her
weight and height were recorded. These were used to calculate the body mass
index (BMI) in each patient.
Full-Body MRI Acquisition
MR images were acquired on a 1.5 T imaging unit (Intera, Philips Medical
Systems), which was fitted with a tabletop extender and allowed automated
table movements. With the tabletop extender, caudocranial coverage of 200 cm
is achieved, which allows examination of all but the tallest subjects. Images
are acquired in six to seven fully integrated stacks with a small overlap. We
acquired coronal T1-weighted gradient-echo images (TR/TE,112/1.62; flip angle,
70°, number of signals averaged, 1; matrix, 256 x 256; 80%
rectangular field of view, 189 x 189). Automated voxel interpolation was
used to yield voxels of 2.02 x 2.02 x 8.00 mm3.
For the scanning process, patients are placed supine with their hands crossed over the abdomen. This enables the subjects' upper limbs to be imaged. We had initially placed the subjects' hands by the sides but found that this caused aliasing artifact. After imaging, the raw data are transferred, in DICOM format, to a workstation for analysis. At the workstation, complex sets of algorithms are sequentially applied to each data set to isolate fat.
Analysis of DICOM Images
In full-body MRI, the subject is imaged in a set of overlapping coronal
sections. The resulting series of MR images must be reconstructed into a
volumetric data set to facilitate analysis. There are two issues that require
particular attention in this process: spatial registration and gray-scale
matching. To achieve correct spatial registration, we use location and
orientation vectors stored in the DICOM headers. This allows us to accurately
position each image within a global coordinate system. In this way, we can
account for any overlap between adjacent coronal slices and correctly generate
the final volume.
Gray-scale matching is necessary because there can be significant intensity offsets between successive coronal sections due to the nature of the MRI acquisition process, particularly the close spatial relationship of the back to the receiver coils. It is necessary to minimize these effects to optimize the performance of the automated analysis procedure. We achieve this using histogram matching. An intensity histogram is constructed for each coronal section, the characteristic peak representing soft tissue is algorithmically identified in each case, and the set of peaks are aligned to match the gray-scale distributions across all sections.
|
To cope with these problems, we have devised a four-step segmentation algorithm. An initial threshold level is calculated based on an analysis of the data histogram. The peak representing soft tissue is located, and voxels with values that fall above the end of this peak are initialized as potential fat voxels. We then use a boundary-enhancement step to compensate for signal drop-off in some peripheral regions of the data. Next we apply a 3D region-growing procedure. Finally, we apply a region-refining process whereby the candidate voxels are grouped into connected regions [1]. Through this process, we arrive at a robust segmentation of the signal due to fat tissue within the data volume.
|
|
Results
Basic User Functions
Analysis results are presented in a number of ways. Simple orthogonal
review allows axial, coronal, and sagittal sections to be examined,
highlighting regions that have been classified as body fat. Color mark-up of
the images provides effective visual feedback, improving the readability of
the data (Fig. 1). This form of
review also allows detailed examination of the distribution of fat throughout
the body and facilitates the easy identification of areas of particular fat
concentration.
Three-dimensional volume-rendered views (Fig. 2) provide an excellent overview of the data, and when used with data space clipping, can effectively visualize the body fat distribution within a volume of interest, providing a more complete view and comprehensive breakdown of the distribution of fatty tissue within the body.
In addition to these visual tools, numeric results are presented to the user. Figure 3 shows an example of the typical results generated by the system. Estimates are made of the subjects' height and weight, and measurements are performed to calculate values for the actual and percentage body fat detected, measured by volume and by weight. All these calculations are performed automatically without any initialization or subsequent intervention required on the part of the user. The entire analysis process, from raw DICOM data to final results, takes less than 2 min.
A database of 42 data sets (Table 1) containing 21 men and 21 women in the age range 18–56 years was used to assess the segmentation algorithm. The BMI was calculated directly from the patients' mass and height and indirectly using the segmentation results. The directly measured BMI values of these data sets range from 18 to 35. Comparison of the direct and indirect methods was used to validate the segmentation procedure, and a good correlation was found (Fig. 4). Much of the data spread in the graph in Figure 4 is due to incomplete body acquisitions in the MRI scanner. Restricting the data to the 21 most complete data sets, where data loss at the extremities, especially at the feet of tall subjects, is kept to a minimum, results in a far superior straight-line fit with greatly reduced deviations from the fitted line in the restricted data subsample. This is shown in the broken line and its associated subset of data points indicated in Figure 4. The sample points representing incomplete acquisitions all fall above the broken line and pull the full-sample trend line (the solid line) in an upward direction. This reflects the relative effect of the missing data on the BMI and percentage body fat calculations. The missed data represent a relatively small volume and, therefore, weight (the feet and ankles), but can result in a disproportionately large variation in the estimated height of the subject. With down-pointing toes, the change can be as much as 30 cm in some cases. These two factors result in little difference in the estimated percentage body fat but can cause a significant elevation in the estimated BMI value since the subject is estimated to be shorter but not much lighter than would be the case were the measurements based on a complete data set. Note that in the case of complete data sets, the automatic method consistently estimates a BMI less than the measured BMI. This is due to an overestimation in the height caused by down-pointing toes. An alteration in the acquisition protocol supporting the subject's feet vertically could correct for this. Correspondingly, calculated BMIs based on incomplete data sets tend to be higher than the equivalent measured values for the reasons given. These observations support the assertion that the automatic measurements achieved represent accurate results for the data under examination. Avoiding errors in height estimation can be easily performed by supporting the feet in a neutral position and by recognizing that 200 cm is the maximum obtainable coverage. With careful technique, all but the tallest patients can be included.
|
|
|
Discussion
Obesity is increasingly recognized as a major health problem in the developed world. An emotive issue that is firmly in the public eye, fat deposition is a complex process that is poorly understood. Genetic, environmental, and hormonal influences are all known to have an effect. It is known from diseases such as growth hormone deficiency and hypercortisolism, which predispose to increased fat deposition, that abnormal fat distribution often occurs in a central or visceral distribution. This visceral adiposity is a known risk factor for hypertension and cardiovascular disease [3]; cerebrovascular disease [4]; insulin resistance and type 2 diabetes [4]; endometrial, colon, and other forms of cancer [5]; and depression and other psychological dysfunctions [6–8]. Despite extensive research interest in obesity, estimation of body fat content in a noninvasive manner is difficult. Widely used anthropometric measurements are crude markers of body fat distribution and also allow no breakdown of regional distribution of fat. In addition, they are subject to extensive inter- and intraobserver error and have poor reproducibility. Many authors have questioned their validity in defined populations such as the young [9] and in different ethnic groups [10]. The derivation of body fat from anthropometric means is difficult and requires the use of equations, more than 3,000 of which have been validated in different age, sex, and ethnic populations.
The BMI, which infers body fat content from height and weight alone is convenient for mass screening, but again, its validity has been questioned in different ethnic groups and age populations [11–13]. It is also known to be inaccurate in athletic individuals [14, 15] and elderly individuals [14, 16] due to extremes of muscle content. In addition, it gives no insight into the regional distribution of fat.
Complex laboratory-based methods of total body fat measurement include techniques such as hydrodensity, radioactive dilution, and air displacement (Bod Pod, Life Measurement). While water displacement is recognized as the gold standard, it is time consuming, extremely inconvenient, and not widely available. The validity of some of these techniques has also been questioned [17–19].
More recently, new techniques have been developed, and among the most important are dual-energy X-ray absorptiometry (DEXA), near infrared interactance (NIR), and total body electrical conductivity (TOBEC). DEXA is probably the most widely available and appears to be reasonably accurate, although it does require ionizing radiation exposure. TOBEC, segmental, and total body impedance infer body fat distribution from resistance to an electric current and have their detractors [18, 20, 21]. In most cases, these techniques are accurate but the equipment is dedicated and expensive, and this is a deterring factor for their application in current medical investigations. In addition, DEXA requires exposure to ionizing radiation.
In our study, we compared our investigational method for fat analysis with BMI alone. While a broad correlation was obtained, we think that most of the disparity between the two techniques is based on a flawed reference standard. Comparison with a better method of fat analysis such as water displacement would have increased the validity of our results, but we did not have access to such techniques. Furthermore, institutional review board restrictions precluded us from using any methods that included ionizing radiation. Finally, even if we had an absolute reference method available for comparison, no gold-standard test exists today for quantifying regional fat distribution, which we think is one of the more important outcomes of our research methods.
Previous authors have used cross-sectional imaging techniques to estimate body fat content [22–24]. CT is widely available, quick, and, because of the unique reproducibility of fat attenuation, suitable for automated image analysis. However, the prohibitive dose of radiation involved precludes its use in a general population. In addition, whole-body imaging using a helical or newer generation MDCT would necessitate extensive image interpolation, thus potentially introducing bias.
MRI, although not as widely available as CT, is becoming more widely used and does not entail ionizing radiation. Previous authors have determined the accuracy and reproducibility of MRI for fat analysis when compared with cadaveric dissection [25] and in animal models [26]. On T1-weighted imaging, fat returns high signal because of a high concentration of relatively immobile protons. Only paramagnetic substances such as iron or melanin, highly viscous fluids, and moving spins (e.g., flowing blood) also return high signal using this sequence. Iron and melanin are not present in the body in any meaningful quantities. In addition, because we acquired coronal images oriented parallel to the axes of most major blood vessels, flow-related enhancement was not a major issue.
Although MRI for fat analysis has already been used by previous authors, a review of the literature reveals that a wide spectrum of techniques has been used, often consisting of single or selected slices, with subsequent extrapolation to the remainder of the body [27–30]. Other early studies on whole-body MRI for fat analysis used gaps of 1–3 cm between axial slices, which require interpolation and thus inevitably introduce bias. Indeed, previous authors have shown that subsampling and limited scanning do introduce bias and increased uncertainty into recorded fat measurements [29]. In our study, we achieved whole-body coverage without slice gaps, thus eliminating any potential for bias and ensuring accurate and reproducible results. In addition, developments in hardware, field homogeneity, and the use of gradient-echo sequences have reduced our imaging time to approximately 140 sec.
Our technique effectively divides the body into voxels of 2.02 x 2.02 x 8.00 mm3. By dividing the body into such small 3D voxels and subsequently using computer-assisted diagnostic (CAD) techniques as outlined, a definite representation of fat distribution is obtained, and potential inaccuracies due to partial volume effects are kept to a minimum. A degree of controversy exists in the literature as to the exact biochemical consistency of the tissue detected as fat by MRI. However, most authors accept that adipose tissue is composed of 84.67% triglyceride, 12.67% water, and 2.66% protein, giving a density of 0.9196 kg/L [31]. In addition, this biochemical consistency appears to be homogeneous throughout the human body, so that confounding variables that underlie the difficulties with other body composition methods such as age, sex, and ethnic origin are removed. Thus, the use of equations, apart from the automated total body fat calculation mentioned previously, is removed. In our cohort of patients, we examined a group with a wide spectrum of body fat levels and distributions without obvious detriment or difficulty, including athletes who are ill served by other available methods.
In this investigation, we illustrate how the use of CAD techniques based on the methodologies of advanced image processing and analysis can be used to quantify the fat distribution within the body in sequences of full-body MR images. The outcome of this research effort is a system that assesses a subject's full-body MRI scan, providing numeric and visual feedback to illustrate its findings. This system generates results in a matter of minutes, allowing for an initial assessment to be performed immediately after the completion of an MRI scan. By highlighting areas where body fat is concentrated, the system allows radiologists to quickly identify and examine regions of interest in the scan.
The system's numeric outputs also provide an accurate measurement of body fat as a percentage of whole-body mass. This is an important metric, which is difficult and time consuming to derive by alternative means. As alluded to previously, the distribution of actual fat tissue in the body is an important measurement of health and overall fitness and is not well quantified by BMI. In this regard, the assessment of body fat in athletes involved in programs of intensive training was an area of particular importance in our study. The ability to localize fat distribution and to show an athlete exactly where it is on the body is of great interest because such information can be used to help shape the training schedule of the athlete. This has particular import for many weight-restricted athletes, such as rowers, boxers, and jockeys. The entire body is covered, without gaps, in an imaging time of approximately 2 mins 20 secs, and 32 coronal slices of 8 mm thickness are acquired for each of the six to seven stacks.
Although we had a cohort of athletes in our study, this technique could be applied to obese patients with equal facility. To account for greater body mass, the slice thickness can be increased, which slightly reduces the specificity of the technique, or the number of slices can be increased, which slightly increases the time of acquisition. All MRI tables have tabletop weight restrictions (usually 300–350 lb) and limited bore size (60 x 60 cm on the Intera). These restrictions preclude analysis of extremely large patients on the system we used. However, the offline analysis program we have developed could be applied to data from any MRI machine, and because it realigns information according to information obtained in the DICOM header, whole-body analysis could be applied even without a moving tabletop, as long as meticulous patient positioning was observed. Hence, the imaging units that are currently used to image the largest patients in clinical practice today could be adapted to measure body fat content with this software.
In conclusion, we have developed a technique that, by combining MRI and CAD, makes it possible to improve the ease, efficiency, and effectiveness with which the analysis of body fat distribution can be conducted on a routine basis.
References
CiteULike 
Complore 
Connotea 
Del.icio.us 
Digg 
Reddit 
Technorati What's this?
This article has been cited by other articles:
|
|
 |
|
  D. D. Brennan, P. F. Whelan, K. Robinson, O. Ghita, J. M. O'Brien, R. Sadleir, and S. J. Eustace Reply Am. J. Roentgenol., December 1, 2007; 189(6): W385 - W385. [Full Text] [PDF]
|
|
|
|
|
|
M. Mantatzis and P. Prassopoulos Total Body Fat, Visceral Fat, Subcutaneous Fat, Bone Marrow Fat? What Is Important to Measure? Am. J. Roentgenol., December 1, 2007; 189(6): W386 - W386. [Full Text] [PDF]
|
|
|
|
 |
|
 M. J. Siegel, C. F. Hildebolt, K. T. Bae, C. Hong, and N. H. White Total and Intraabdominal Fat Distribution in Preadolescents and Adolescents: Measurement with MR Imaging Radiology, March 1, 2007; 242(3): 846 - 856. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
= best
21 data sets, • = rest of full 42 data sets, broken trend line = best 21
data sets, solid trend line = full 42 data sets.