WO2011139232A1

WO2011139232A1 - Automated identification of adipose tissue, and segmentation of subcutaneous and visceral abdominal adipose tissue

Info

Publication number: WO2011139232A1
Application number: PCT/SG2011/000171
Authority: WO
Inventors: Michael Chee; Vitali Zagorodnov
Original assignee: Nanyang Technological University; National University Of Singapore
Priority date: 2010-05-03
Filing date: 2011-05-03
Publication date: 2011-11-10

Abstract

Adipose tissue is identified within a three-dimensional abdominal MRI image set comprising a water-only image and fat-only image of the abdomen of a subject by forming a "fat-fraction" image, which has an intensity at each voxel equal to the intensity of the fat-only image divided by the sum of the intensity of the fat-only image and the water-only image. This fat-fraction image is thresholded to form a good first approximation of the position of the adipose image. The fat-fraction image is used to obtain a threshold to segment a fat-only image. The adipose tissue is segmented into SAT and VAT by defining a graph having vertices corresponding to voxels of the abdominal image, and edges connecting neighboring vertices. Each edge is given a weight. The graph is partitioned into background and SAT portions using a partition which minimizes the sum of the weights of the edges spanning the partition, and the adipose tissue to either side of the partition is respectively labeled as VAT and SAT. Experiments utilising the invention achieved excellent correlations ( r = 0.998 for SAT and r = 0.965 for VAT) with the manual measurements and reproducibility (CV=1.5% for SAT and CV=2.3% for VAT), which rivals that of CT-based measurements.

Description

Automated identification of adipose tissue, and segmentation of subcutaneous and visceral abdominal adipose tissue

Field of the Invention The present invention relates to methods for automated identification of adipose tissue in abdominal MRI images, and automated segmentation of subcutaneous abdominal adipose tissue (SAT) and visceral abdominal adipose tissue (VAT) in abdominal MRI images. It further relates to computational systems for performing the methods, and computer program products containing software for performing the methods.

Background of the Invention

Distribution rather than amount of adipose tissue across the body appears to influence cardiovascular and metabolic health. The initial observation that upper body fat.

accumulation carries higher metabolic risks compared to accumulation of fat in the lower part of the body dates back more than 60 years [1]. More recent studies have narrowed the focus to abdominal fat accumulation, distinguishing subcutaneous abdominal adipose tissue (SAT), which refers to the fat deposited under the abdominal skin, from the visceral abdominal adipose tissue (VAT), deposited within the abdominal cavity around the internal organs [2]. VAT has much stronger association (compared to SAT) with metabolic risk factors, such as insulin resistance, dyslipidemia, hypertension, and elevated cholesterol level [3-8]. These associations are possibly mediated through an increase in the plasma concentration of free fatty acids (FFA) [9, 10], which can influence hepatic glucose metabolism [11]. The importance of abdominal fat

accumulation has recently been acknowledged by the National Cholesterol Education Program - Adult Treatment Panel (NCEP-ATP III) [12] and the International Diabetes Federation [13], which include the waist circumference (rather than the body mass index) as a criterion to diagnose metabolic syndrome, a direct complication of obesity. Despite its importance, adipose tissue distribution cannot be directly assessed by existing anthropometric measures. Body mass index (BMI) has been shown to have high correlation with the SAT volume, r=0.94(f)/0.88(m), but a weaker correlation with the VAT volume, especially for men, r=0.80(f)/0.59(m) [14]. Waist circumference (WC) cannot distinguish SAT from VAT, muscles, and connective tissue, which makes it a suboptimal measure of visceral obesity. For example, both WC and visceral fat volume decrease following caloric restriction and/or exercise, but there appears to be no association between these variables [15]. Another study examining effects of exercise found significant decrease in visceral fat deposits but no change in body weight, body mass index, or the waist circumference [16]. Skinfold thickness, another commonly used anthropometric measure, provides an approximate measure of SAT, but has no direct relationship to VAT [10].

Computerized Tomography (CT) and Magnetic Resonance Imaging (MRI) provide the most direct and validated methods of evaluating adipose tissue distribution [15, 17, 18]. While CT-based assessments are more accurate, MRI has the advantage. of not exposing subjects to ionizing radiation. However, one of the obstacles for wider acceptance of MRI in clinical practice is the substantial amount of labor involved in segmenting multislice or 3D MR images. Full manual tracing of fat masses may take up to an hour depending on the number of slices, making manual measurements impractical for routine clinical use. Most existing studies have involved small samples of 20-50 subjects (see recent review [11] of thirteen studies) and have often evaluated a single axial slice [19]. This is potentially less accurate than assessing a slab of abdominal tissue [20, 21]. Using single slices also results in poor reproducibility, with reported coefficients of variation (CV) ranging from 9.4% to 13.8%.

Several recently proposed algorithms for MRI-based assessment of abdominal tissue distribution [22-25] feature fully automated processing pipelines that make user intervention unnecessary. However, only one of these (HippoFat, [22]) has been independently validated, ranking lowest compared to several other semi-automated techniques. This is largely a result of frequent failure to differentiate VAT from

surrounding non-fat tissue. The reported test-retest reproducibility of VAT

measurements, quantified by coefficient of variation (CV), ranges from 7% for the method developed by Kullberg et al [23] to 13% for HippoFat [26].These results are similar to those obtained using manual single-slice MR-based methods and much poorer than that of CT-based measurements (2.8-4%) [27, 28]. Poor segmentation performance can be attributed to a) difficulty in selecting an appropriate threshold to separate adipose from non-adipose tissue. This is often due to the presence of intensity non-uniformity and irregular appearance of visceral adipose tissue, and b) difficulty in separating the adipose tissue into SAT and VAT compartments. This often results from 'connections' between these compartments.

Summary of the Invention The present invention aims to provide new and useful methods for automated segmentation of SAT and VAT, and systems and software products for performing the methods.

A first aspect of the invention proposes in general terms that an abdominal image (e.g. a three-dimensional image) of the adipose tissue of a subject (which is very likely to be a human subject, but could also be a non-human animal subject) can be segmented into SAT and VAT, by defining a graph (e.g. a three-dimensional graph) having vertices corresponding to voxels of the abdominal image, and edges connecting neighboring vertices. Each edge is given a weight. The graph is partitioned into inner and outer portions using a partition which minimizes the sum of the weights of the edges spanning the partition, and the adipose tissue on either side of the partition is respectively labeled as VAT and SAT. The image may be a three-dimensional image, or may alternatively be an image of a single slice of the abdomen (i.e. a two-dimensional image). In the latter case, the term "voxel" should be understood as equivalent to "pixel". However, the single slice of the abdomen may be one of a set of parallel slices of the abdomen (e.g. spaced apart in the height direction of the subject), so that each pixel of a single image constitutes a voxel of the set of slices considered as a whole.

The images are acquired by any imaging modality that yields volumetric or single transverse slice views of the abdomen. Currently this means MRI and Computerized Tomography (CT).

The weights can be selected to control the way in which the partition is generated. For example, they may be set higher for edges connecting two vertices corresponding to respective voxels representing adipose tissue which are far from any voxels which do not represent adipose tissue. This influences the partition to pass through narrow bridges of adipose tissue connecting larger masses of the same.

A second aspect of the invention proposes in general terms, identifying adipose tissue from an abdominal MRI image set obtained by the Dixon technique, which generates an in-phase image, an out-phase image, a water-only image and fat-only image of the abdomen of a subject. The image set is used to form a "ratio" image, which has an intensity at each voxel representative of (e.g. proportional to) the intensity of the fat-ohly image at the same voxel divided by the intensity of a normalization image at the same voxel obtained from the image set. High intensity voxels of the ratio image can then be identified and labeled (e.g. by thresholding, or by an equivalent algorithm which is expected to give results of much the same quality, such as Markov Random Field labeling, a split and merge algorithm, or a level sets algorithm) to form a good first approximation of the position of the adipose tissue. In one possibility the ratio image may be a "fat-fraction" image, which has an intensity at each voxel equal to the intensity of the fat-only image at the same voxel divided by a normalization image at the same voxel, which is the sum of the intensity of a the fat-only image (F) and the water-only image (W). However, this is not the only possibility. For example, the ratio image may alternatively be simply F/W (that is, the normalization image is the water-only image, W). In another possibility, the ratio image may be (F- W)/(F+W), that is (F/W-1)/(F/W+1). In yet another possibility, the ratio image may even by F/l, where I is the in-phase image.

A first advantage of this idea is that the best threshold (i.e. the threshold which tends to give images most similar to ground truth images segmented manually) tends to be much the same for MRI image sets obtained from different scanners. This is because while the overall intensities of the fat-only and water-only images differ from scan to scan, the ratio of fat and water intensities in these images tends to be constant.

Another advantage of the second aspect of the invention is that if there is any non- uniformity in the intensity of the images of the image set, this tends to apply roughly equally to both the water-only and fat-only images, so that by using the ratio of images there is automatic compensation for non-uniform intensity.

Note that in the Dixon technique the water-only and fat-only images are derived from in- phase and out-phase images. As both water-only and fat-only images are acquired together, intensity non-uniformities are canceled out when the ratio image is formed. In other MRI techniques, water-only and fat-only images can also be acquired in two separate acquisitions, called water-suppressed and fat-suppressed acquisitions. The second aspect of the invention is superior to segmentation techniques based on such images, because if the images are acquired in two separate acquisitions, they may not be well registered and they do not have the same B1 transmit field inhomogeneity, meaning that intensity non-uniformities will not be completely canceled when the ratio image is formed.

A weakness in this approach is that a spuriously high ratio will arise in locations where the intensity of both the water and fat are very low. Thus, it is preferable to have a step of identifying "air" regions in the image in which there is likely to be neither water nor fat, and excluding these from the ratio image (either before or after the thresholding). A convenient way to identify such regions is using an in-phase MRI image which is part of the MRI image set, since regions of such images which have low intensity tend to correspond to air regions. However, there are other ways in which suppression of "air" voxels can be realized. For example, one can identify low intensity voxels in the F image, then identify low intensity voxels in the W image, and then label voxels appearing as low intensity in both F and W as air voxels.

The resultant thresholded ratio image may be used in a method according to the first aspect of the invention.

Alternatively, having obtained this approximate segmented image, it can be used to produce an enhanced segmented image. For example, it can be used to identify a threshold for segmenting the fat-only image itself. That is, the threshold for segmenting the fat-only image can be chosen such that the resulting thresholded fat-only image is most similar to the thresholded ratio image. This provides a very robust way of selecting a threshold for the fat-only image, compared to existing techniques.

The methods of the invention are performed "automatically" by which is meant that, although there may be human interaction in initiating the methods, there is no further human interaction until the methods are completed. The methods proposed here are motivated by a desire to simplify the intensity threshold selection through the use of images obtained by the Dixon technique [29, 30]. The concept of using a graph to perform adipose tissue partitioning was inspired by the use of such techniques in a very different biomedical imaging problem [31].

Experiments utilising both aspects of the invention have been found to achieve excellent correlations ( r = 0.998 for SAT and r = 0.965 for VAT) with the manual measurements as well as measurement reproducibility (CV=1.5% for SAT and CV=2.3% for VAT), that rivals CT-based measurements.

The invention may alternatively be expressed in terms of a computer system to perform the method. The computer system has a processor and a memory, and the memory stores program instructions which cause the processor to carry out a method as explained above. Alternatively, the invention may be expressed as a computer program product, such as a tangible recording medium (e.g. a CD or floppy disk), storing program instructions operative by a processor of a computer system to cause the processor to carry out a method as explained above.

Brief description of the drawings

Embodiments of the invention will now be described for the sake of example only with reference to the following figures in which:

Fig. 1 is a flow diagram of a second embodiment of the invention;

Fig. 2 is a diagram showing images during a part of the second embodiment of the invention;

Fig. 3, which is composed of Figs. 3(a) to 3(f), shows a comparative example; Fig. 4 is a diagram showing images during another part of the second

embodiment of the invention;

Fig. 5, which is composed of Figs 5(a) to 5(f), shows results of the second embodiment of the invention; Fig. 6, which is composed of Figs. 6(a) to 6(f) shows two types of segmentation error occasionally made by the second embodiment of the invention;

Fig. 7 shows correlation between manual ground truth results and the results of the second embodiment of the invention; and

Fig. 8 is Bland-Altman plots showing the difference between the manual and automated measurements.

Detailed description of the embodiments

We now discuss two embodiments of the invention. The input to each embodiment is an MRI image set, obtained from the Dixon technique [29, 30], in which two images (an "in- phase" and an "out-phase" image) are collected within a short interval of each other and in register. Water and fat signals are in-phase in the in-phase image and out-of-phase in the out-phase image. These in-phase and out-phase images are then post-processed to yield water-only and fat-only images. The four images (in-phase, out-phase, fat-only, and water-only) form the final image set that is stored on MR scanner console system. These are not the only forms of MRI images known in this field. MRI-based assessment of adipose tissue has been traditionally performed using T1-weighted acquisitions [17, 22, 23, 32, 33], and more recently using water-saturated MRI [34, 35]. Water-saturated techniques render tissues with low fat content as black rather than grey, making it easier to distinguish between adipose and non-adipose structures. Images obtained using the Dixon technique [29, 30] offer a further advantage over water-saturated MRI by acquiring two images (in-phase and out-phase) at different echo times (TEs), exploiting the difference in chemical shift between water and fat. The in- and out-phase images are further combined to obtain water-only and fat-only images, where the latter appear similar to water-saturated images. Water-only and fat-only Dixon images have so far been primarily used for generation of fat-fraction maps (ratio of fat-only and fat- only plus water-only images) to quantify the fat infiltration of non-adipose structures.

1. First embodiment

In a first embodiment of the method, the first step is the identification of a non-air mask. This is achieved by comparing the intensities of in-phase image voxels with a threshold equal to 0.05 times the maximum intensity of this image. Voxels with an intensity below the threshold are classified as air voxels and their intensity is set to 0 in the non-air tissue mask, while intensities of other voxels are set to 1. The threshold can be adjusted depending on the image noise. Alternatively, the threshold can be selected

automatically by deriving the location of a valley between two peaks in an intensity histogram.

An adipose tissue mask is formed by identifying voxels that have intensities greater than 0.8 in a fat fraction image formed by taking the ratio of the fat-only image to the sum of fat-only and water-only images, and at the same time have intensities equal to 1 in the non-air tissue mask. Intensities of such voxels are set to 1 in the adipose tissue mask, while the intensities of all other voxels are set to 0. Raising the threshold will decrease the false positive rate and increase the false negative rate. We are aware of at least one recent publication that used a similar approach [24], but to our knowledge this document was published only after the priority date of the present application.

The separation of the adipose tissue into subcutaneous and visceral compartments is accomplished by first converting the adipose tissue binary mask to a graph where each voxel becomes a vertex and undirected edges are formed between the vertices corresponding to neighboring voxels (note that in variants of the embodiment, the edges may be directed). The edges between adipose tissue voxels (voxels with intensity 1 in the mask) are assigned weight equal to the distance from such a voxel to its nearest non-adipose tissue voxel. Such a distance can be calculated by performing a distance transform operation on the adipose tissue mask. The weights of the edges between adipose and non-adipose tissue voxels are set to 1. The weights between non-adipose tissue voxels are set to infinity.

A "foreground seed" (or equivalently "SAT seed") is then defined as follows. First, holes in the non-air tissue mask are filled by estimating all connected components of voxels with intensities 0. The intensities of voxels in all such components but the largest are then changed to 1. This filled mask is then morphologically eroded by 3 voxels using an arbitrary structuring element. The foreground seed is defined as the set of voxels that have intensity 1 in the filled mask, intensity 0 in the eroded mask, and intensity 1 in the adipose tissue mask. A "background seed" is defined as the set of all non-adipose tissue voxels, i.e. voxels with intensity 0 in the adipose tissue mask. Note that some of the background seed voxels may be extra-corporeal.

The graph is then split by finding the cut that has the smallest cost among all cuts separating the foreground and the background seeds. The cost of the cut is defined as the sum of the weights of all removed edges. The voxels that appear in the foreground portion of the cut are declared as SAT. The adipose tissue voxels that appear in the background portion of the cut are declared as VAT.

The threshold of 0.8 which was used to identify the adipose tissue will have wrongly classified partially filled voxels with fat portions ranging between 0.5 and 0.8 as non- adipose tissue. To correct this, a post-processing step is performed on the VAT, such that all voxels with intensities between 0.5 and 0.8 (or other fat fraction threshold(s) chosen by the user) in the fat fraction images that are in the neighborhood of voxels classified as VAT are re-classified as VAT. 2. Second embodiment

We now turn to discussing in detail a more sophisticated embodiment.

2.1 Automated Image Analysis The proposed algorithm consists of two main portions, a) fat-fraction based

adipose/non-adipose tissue separation, and b) a graph-cut based algorithm that cuts connections formed between subcutaneous and visceral compartments. Fig. lis a flow diagram of the overall algorithm. Fig. 2 shows a typical image at each step of the first part of the process. Fig. 4 shows a typical image at each step of the graph-cut process. The input to the method is a Dixon MR sequence image set as explained above, but in fact only uses three out of the four images (fat-only, water-only and in-phase images) as shown at the top of Fig. 2, in the portion marked 2A, which will be referred to here as Fig. 2A.

Most existing solutions follow a similar framework, i.e. thresholding to obtain initial mask followed by a cut, although our implementation of these steps is unique. In computer vision, this framework has been broadly referred to as mask segmentation [36].

The steps of the method are as follows, where the numbering corresponds to that of Fig. 1.

Fat-fraction based adipose/non-adipose tissue separation

1 ) Form the fat-fraction image, by dividing the fat-only image by the sum of the fat-only and the water-only images: FF =——— (1)

F + W This gives an image such as shown in Fig. 2B. This division results in a noisy appearance of voxels outside the abdomen. These voxels are filled with air and have low intensity in both water-only and fat-only images, which can makes their ratio arbitrary large. 2) Identify and remove air voxels by selecting the voxels in the in-phase image with intensity less than aJ_imx , where I_mas is the maximum image intensity and is the air- voxel threshold, empirically fixed at a = 0.05. This gives ah image as shown in Fig. 2C.

3) Remove the air voxels identified in step 2 from the output of step 1 (giving an image as shown in Fig. 2D). This step is not separately indicated in Fig. 1 , but is included in the step marked 1.

4) Threshold the result of step 3 at a fixed intensity value T = 0.6 to separate the adipose from the rest of the image (giving an image as shown in Fig. 2E), thereby obtaining an "adipose tissue mask". This step represents a key difference between our approach and previous work. Most existing approaches used global intensity thresholding [14, 23, 25, 34, 37, 38] or fuzzy clustering [22, 39] to outline adipose structures. Because MR image intensities are arbitrarily scaled, the threshold cannot be set prior to image acquisition, and is estimated based on the image histogram, for example, at the deepest point in the valley between the peaks corresponding to adipose and non-adipose tissue [37]. This is illustrated in Fig. 3, where Fig. 3(a) is a Dixon's fat-only image, and Fig. 3(b) is an intensity non-uniformity corrected image. Non-uniformity correction removes the smooth variations of image intensities, but preserves regional signal differences that are biologically relevant. Fig. 3(c) is an intensity histogram based on Fig. 3(a) and Fig. 3(d) is an intensity histogram based on Fig. 3(b). As illustrated, intensity non-uniformity and heterogeneity in the appearance of the abdominal structures can cause this approach to fail by causing the image histograms to become unimodal, with a single peak corresponding to non-adipose voxels and a long tail corresponding to adipose voxels. The unimodal histogram complicates the determination of intensity threshold, found to be 0.2 and 0.26 for Fig. 3(a) and 3(b) respectively. These result in thresholded images Fig. 3(e) and 3(f) respectively. The results of thresholding are similar regardless of whether non-uniformity correction has been applied, assuming a proper choice of the threshold value.

By using the fat fraction (FF) images, the second embodiment greatly simplifies the process of finding the appropriate threshold, by reducing the intensity nonuniformity and removing arbitrary MR scaling. Since the water-only and fat-only images are acquired together, they are influenced by the same B1 transmit field inhomogeneity, which can be modeled using multiplicative bias field B , i.e. F = F₀B , where F₀ is the fat-only image that is free from intensity nonuniformity. This leads to cancelation of the bias field when the images are divided:

FF -J— ^F°^B — A_

F + W F₀B + W₀B F_A +W₀

Similarly, since the same arbitrary MR scaling is applied to water-only and fat-only images, it is eliminated when the images are divided. As a result, the fat-fraction image intensities range from 0 to 1 , with 0 corresponding to pure non-fat voxels (voxels that have zero fat content) and 1 to pure fat voxels (voxels that are fully filled with adipose tissue). This allows the threshold to be fixed prior to image acquisition. In our implementation we used T = 0.6 , which was determined empirically.

5) Intensity thresholding of the fat fraction image results in somewhat noisier appearance compared to the thresholding of the corresponding fat-only image (compare Fig. 2E with Fig. 3(a) and 3(f)). This occurs because taking the ratio of images tends to amplify motion artifacts (ghosting) and brighten the voxels that have low intensities in both water-only and fat-only images but which have not been removed by air voxel masking. As shown later in the Results section, this increased noise causes poor reproducibility of the segmented visceral adipose tissue. To avoid this problem, we use the thresholded fat-fraction images to estimate the appropriate threshold value for the corresponding fat-only images. In other words, we make a selection from the various thresholded versions of the fat-only image which could be generated (shown as Fig. 2F). This is implemented by finding the intensity threshold T that maximizes the similarity between the fat-only image F , thresholded with T , and the corresponding fat- fraction image FF , thresholded with 0.6:

T = arg max JS{F > T,FF > 0.6) (2)

T

Here JS stands for Jaccard Similarity, which is a measure of overlap (or similarity) between two binary masks defined as

A J B

SA TA A T Separation Using Graph Cuts

The second embodiment's cutting approach is shown in a schematic form in the lower portion of Fig. 1 and is further illustrated in Fig. 4. The input consists of the appropriately thresholded fat-only image (as shown in Fig. 2G and Fig. 4A) and the mask generated by the thresholding of the in-phase image (as shown in Fig. 2C and Fig. 4B).

6) Remove upper limbs using 3D connected component labeling and selection of the largest connected foreground component. This procedure will work only if the hands appear disconnected from the abdomen in the MR image, which must be ensured during the acquisition. The resulting abdominal mask is likely to contain holes, which can be filled by conducting another connected component labeling process and reclassifying all connected background regions (except the largest one) as foreground (giving an image as shown in Fig. 4C).

7) Define a SAT seed region (a region fully contained inside SAT) by selecting a 5mm outer layer on the abdominal mask. To obtain this layer, the abdominal mask was eroded by 5mm, and then subtracted from the non-eroded mask. The in-phase thresholded image (with hands removed and holes filled) was used as the abdominal mask, rather than the thresholded fat image, because the latter might not have continuous outer border due to umbilical discontinuity. The layer thickness (5mm) was chosen so that the resultant region has no overlap with VAT. This assumes that VAT is located deeper than 5mm from the external abdominal border. This assumption is reasonable, because VAT is separated from the external abdominal border by layers of skin, fat and muscle. The result is as shown in Fig. 4D.

8) To separate subcutaneous from visceral adipose tissue, we used an in-house developed graph cut technique, which has recently been validated on the brain skull stripping problem [31 ]. In this approach, the fat-only thresholded image (Fig. 2A) is converted into a graph, with voxels as vertices, edges connecting neighboring vertices, and weights of the edges assigned as follows,

where D(v,) and D(v_j) are distances from voxels v, and v_y to the nearest adipose tissue boundary. The weights of the edges between adipose and non-adipose tissue voxels are set to 1. The weights between non-adipose tissue voxels are set to infinity.

A background seed is defined as the set of all non-adipose tissue voxels, i.e. voxels with intensity 0 in the adipose tissue mask. Note that some of the background seed voxels may be extra-corporeal.

The graph is then partitioned into background and SAT compartments, the background seed and the SAT seeds being in the respective compartments, so that the value of the cut (sum of the weights of all cut-off edges) is minimized (to give a final image such as shown in Fig. 4E). The distance transform weight assigned by Eqn. (3) forces the cut to go through the narrow connections between the compartments, because the weights of the edges within these connections are smaller compared to weights of the edges deeper inside SAT or VAT. Finding the minimum cut in itself is a well-known combinatorial optimization problem, and can be solved using a variety of available rnin- cut/max flow algorithms [40]. Finally, the adipose tissue voxels that appear . in the background portion of the cut are declared as VAT. Our cutting approach offers several advantages over existing solutions, such as snakes initiated at an external SAT border and grown to an internal SAT border [22], or convex hull [24]/morphological closing [25] procedures applied to non-adipose tissue mask. Snakes may overgrow beyond the desirable position if the image edges are too soft, or stopped too early by false edges, which can be formed with within SAT by various connecting tissue and fibers. A convex hull approach might not be appropriate because the internal SAT surface is not always convex [25]. Morphological closing may fail when the connections between subcutaneous and visceral adipose tissue compartments are wider than the size of the structural element used [31]. Our approach overcomes the last problem by allowing cutting even wide connections, as long as the connection width is smaller than the sizes of the regions connected by it.

2.2 Results

Subjects and scans

The algorithm was tested on 289 elderly Chinese subjects (144 males, mean age 68 years), each scanned once on a 3T Tim Trio MRI scanner (Siemens, Erlangen,

Germany). All participants provided informed consent prior to undergoing evaluation and this study was approved by the National University of Singapore Institutional Review Board. A small subset (N=9, 5 males, age 65±5, BMI 24.5+4.7) of these subjects' scans was used for quantitative validation of our algorithm. These scans were manually segmented at every 10^th slice (8 traced slices per scan) by a board certified radiologist who was blinded to the algorithm performance. Analyze 9.0 (Mayo Clinic, MN, USA) was used to perform the tracings. Ten additional young volunteers (4 males, age 25±3, BMI 21.6±2.2) were each scanned three times. After each acquisition, the subjects were taken out of the scanner and repositioned. The purpose of these repeated measurements was to investigate the influence of repositioning on the reproducibility of the measured adipose tissue volumes. All MRI acquisitions were performed using a 6-channel body coil and 2-point Dixon sequence (TR = 4.20ms, TE1 = 1.225ms, TE2 = 2.45ms, flip angle = 10°, bandwidth 850 Hz/Px, FOV 380 x 285mm, 80 axial slices, 256 x 192 matrix, 2.5mm slice thickness, 16s acquisition time). The volumes were centered over L2-L3.

Experiment 1. Qualitative Evaluation

For qualitative evaluation we applied the second embodiment to the full dataset (289 scans) and visually examined the quality of segmentation. Typical segmentation results for obese, slim, and slim appearing but viscerally obese subjects are shown in Fig. 5. The algorithm took about 5 min to complete on each image (80 slices) using an PC with a 2.8Ghz Intel Core 2 Duo processor. This timing was similar to that of HippoFat, and slightly better than that which was reported for the other automated algorithms (7min for 15 slices [25], 7.4min [23]).

Fig. 5(a) is a segmented fat image for an obese female subject (70 years), BMI: 34.7, WHR: 0.92, WC: 106cm, SAT Volume: 4501cm³, VAT Volume: 3731cm³, and Fig. 5(b) is the corresponding image partitioned into VAT (inner, darker region) and SAT (outer, lighter region).

Fig. 5(c) is a segmented fat image for slim male subject (71 years), BMI: 19.8, WHR: 0.86, WC: 77cm, SAT Volume: 672cm³, VAT Volume: 474cm³. Fig. 5(d) is the corresponding image partitioned into VAT and SAT. Fig. 5(e) is a segmented fat image for a male subject (71 years), BMI: 22.6, WHR: 0.89, WC=85cm, SAT Volume: 1787cm³, VAT Volume: 3732cm³. Fig. 5(f) is the

corresponding image partitioned into VAT and SAT. This subject's BMI, WHR and WC are in the normal range, despite substantial amount of visceral fat. The overwhelming majority of the images were correctly separated into VAT and SAT compartments. A small number of images (5 out of 289, approx. 1.7%) exhibited strong connections between VAT and SAT, which resulted in estimated subcutaneous fat "leaking" inside the visceral compartment. This is illustrated by Fig. 6 in which Figs 6(a) and 6(c) are thresholded fat-images, and Figs. 6(b) and 6(d) are the corresponding images separated into VAT and SAT with arrows indicating leakage due to wide connections between adipose tissue compartments. Such segmentation errors, however, typically affected only a few slices in the volume and are unlikely to cause large errors in the estimated SAT/VAT volumes.

Another issue was extremely low intensity of adipose tissue in a few slim athletic subjects (4 out of 289, 1.4%), which resulted in underestimation of both adipose tissue compartments. This is illustrated in Fig. 6(e), showing a thresholded fat image of extremely low intensity for a lean individual, causing underestimation in the separated Fig. 6(f) of both adipose tissue compartments.

Experiment 2. Volumetric Comparison Against Manually Traced Scans

For quantitative evaluation, we compared SAT and VAT volumes obtained using our algorithm with those obtained using manual tracing on a set of 9 scans. The volumes were computed by adding segmented areas of all traced slices (every 10^th slice) and multiplying the result by the slice thickness. The means (standard deviations) of manually measured volumes were 376ml (140ml) for SAT, and 217ml (93ml) for VAT. The performance results are summarized in Table 1. Table 1. Adipose tissue volume estimation - comparison between various approaches tested.

Overall, the second embodiment achieved excellent correlations between the automated and manual measurements for both adipose tissue compartments, r = 0.997 for SAT and r = 0.967 for VAT (Table 1 , row 1 ). This is illustrated also by Fig. 7 which shows associations between SAT and VAT volumes obtained using our automated algorithm and manual tracings, using scans from a sample of 9 subjects.

Fig. 8 shows Bland-Altman plots showing the difference between manual and automated measurements. According to the Bland-Altman plots, the automatically measured SAT volumes were slightly underestimated compared to the manual measurements. Excluding one outlier (an extremely lean athletic subject with very low SAT volume), the mean underestimation was 14.5ml, which constitutes 3.9% of the mean SAT volume (376 ml). After taking into account the regression line, the standard deviation of SAT volume measurements was 7.1ml or 1.9% of the mean SAT volume. The automatically measured VAT volumes were underestimated for lean subjects and .

20 overestimated for overweight subjects. After taking into account the regression line, the standard deviation of VAT volume measurements was 29ml or 13.3% of the mean VAT volume (217ml). The standard deviations of SAT and VAT volumes obtained using our algorithm and expressed as percentages of mean measured volumes were similar to those reported for Kullberg's method [23] (2.7% for SAT and 14% for VAT), but worse than those reported for Liou's method [25] (1-1.2% for SAT and 1.7-2.7% for VAT, depending on the acquisition sequence used).

For further comparison, we applied the publicly available HippoFat algorithm [22] to the same data set, giving results also shown in Table 1. Since HippoFat was designed for T1 weighted images, it was applied to the in-phase images from our water and fat acquisitions. We experienced several issues while running the algorithm. In four out of 9 subjects, the algorithm failed to run fully automatically and required manual adjustment of the initial SAT contour. One subject had to be excluded from evaluation because the algorithm would not run on it even after the manual adjustment. Our experience is consistent with that reported in a comparison study [26]. However, it is also possible that at least some of the problems we encountered were due to the fact that in-phase images from water and fat acquisition were not optimized for best fat/lean contrast. Even after manual interventions, HippoFat's measurements correlated less well with expert segmentation compared to our algorithm ( r = 0.966 for SAT, r = 0.860 for VAT), see Table 1. The SAT volumes were overestimated by mean 13.9ml or 3.7% of mean SAT volume, while VAT volumes were considerably underestimated by mean 65.2ml or 30% of mean VAT volume. The standard deviations of SAT and VAT volume measurements were 26.4ml and 54.7ml (7% and 25.2%) respectively, which is considerably worse compared to the second embodiment.

We also analysed the quality of the estimated volumes for two preliminary versions of the second embodiment. The only difference between these preliminary versions and the second embodiment concerns how the initial adipose/non-adipose segmentation is obtained. As explained above, in the second embodiment the fat-only image was thresholded, where the threshold was chosen by maximizing similarity with the previously thresholded fat-fraction image (Fig. 2G). In the first preliminary version of the embodiment, the segmentation was obtained by thresholding the fat-fraction image (Fig. 2E). In the second preliminary version of the second embodiment, we thresholded the fat-only image without the help of the fat-fraction image, determining the threshold automatically using Otsu method [41]. The first preliminary version performed similarly to the second embodiment, but the performance of the second preliminary version was much poorer, due to failure to select an appropriate threshold on two out of nine subjects, most likely due to reasons previously discussed above in relation to Fig. 3.

Experiment 3. Test-Retest Reproducibility

To evaluate the measurement reproducibility we calculated the coefficient of variation (CV) of SAT and VAT volumes between all pairs of repeated scans, acquired on 10 healthy young volunteers. All methods performed well in terms of SAT volume reproducibility, but there were substantial differences between the methods in terms of VAT volume reproducibility (the last column of Table 1). The second embodiment, and its second preliminary version, achieved the best VAT volume reproducibility (CV=2- 2.3%), while the reproducibility of HippoFat and the first preliminary version were much worse (CV=8.5-8.6%). The poor HippoFat reproducibility is consistent with the results obtained in comparison study [26], which estimated HippoFat volume reproducibility at CV=13.4%. Poor performance of the first preliminary version of our algorithm was likely due to noisier appearance of the segmentation results (compare Fig. 2G and Fig. 2E).

The reproducibility of VAT volume measurements obtained using the second

embodiment (CV=2.3%) was also much better than that reported for Kullberg's method [23] (CV=7.1 %), and there were no reproducibility values reported in Liou's work [25]. 3. Discussion

We have proposed novel algorithms for automated segmentation and quantification of abdominal adipose tissue distribution from water and fat (Dixon) MRI acquisitions. The processing pipeline involves intensity thresholding for adipose/non-adipose tissue separation followed by a graph cuts approach to separate the adipose tissue into subcutaneous and visceral compartments. The processing does not involve any user interaction and, in experimental tests for the second embodiment, takes about 5min on. a typical scan (80 slices).

Our approach achieves, class-leading reproducibility in visceral adipose tissue volume measurements (CV=2.3%). This was better than reproducibility reported for all five methods compared in the study [26] (CV= 5.2-13.4%), four of which involved varying degrees of human interaction, and is comparable to that of CT-based VAT volume measurements (2.8-4%) [27, 28]. The sensitivity of our measurements to even small reduction in the volume of adipose tissue suggests possible deployment in short-term weight loss studies.

In terms of accuracy, there was a strong agreement between automated and manual measurements (r = 0.998 for SAT and r = 0.965 for VAT). The error standard deviations, expressed as proportions of mean volumes (1.9% for SAT and 13.3% for VAT), were similar to those reported for Kullberg's method [23] but worse than those reported for Liou's method [25]. However, in our opinion, the accuracy results reported by Liou et al were not entirely realistic. For example, the thresholds to separate adipose and non- adipose structures were chosen the same for both automatic and manual

segmentations. This means that if the automatic threshold was chosen too low and hence underestimated the adipose tissue volume, the same underestimation would occur in the manual segmentation. Moreover, the authors themselves acknowledged that high-intensity nonfat pixels arising from bone marrow and fatty intestinal contents were not edited out, and were considered erroneously as part of visceral adipose tissue. The automated measurements of VAT volumes obtained by the second embodiment were slightly underestimated for lean subjects and slightly overestimated for overweight subjects. This is a consequence of the design choice to fix the fat-fraction threshold at 0.6 for initial fat/non-fat separation. The visceral adipose tissue in lean subjects has darker appearance on fat-only and fat-fraction images (as shown in Fig. 5(c)), due to larger relative proportion of penetrating blood vessels, and hence ideally requires lower threshold values (0.5-0.55)..

Note that, if necessary, our approach can also be combined with water-suppressed MRI, by acquiring a pair of separate fat+water and water-suppressed scans [42].

However, this would require the scans to be aligned to each other, complicating the processing pipeline. Moreover, the resultant fat-fraction images will be less uniform, because the division of water-suppressed and fat+water scans removes only one source of intensity nonuniformity, due to coil sensitivity. In comparison, the fat-fraction images generated using the Dixon technique do not need alignment and further reduce intensity nonuniformity by alleviating the effect of B1 transmit field inhomogeneity, which is usually the largest source of intensity variations in MR images.

Similar to the majority of existing methods, the second embodiment was not designed to edit out high-intensity non-fat voxels arising from bone marrow, inter-muscular fat, intestinal fat and motion artefacts. While this could lead to overestimation of VAT volume, this is often compensated by darker appearance of VAT that leads to underestimation of its volume. We are aware of only one existing approach that attempted to remove misclassified bone marrow/inter-muscular fat using pelvis and vertebra models [23]. However, incorporation of such correction techniques relies on appropriate positioning, e.g., through automated finding of the spinal cord position, which complicates the algorithm and may make it less reliable.

Our experiments also highlight the importance of thresholding the fat-only image and choosing the threshold with the help of the fat-fraction image. Relying on the fat-only image alone (without the fat-fraction image) may lead to inappropriate threshold selection (caused by intensity nonuniformity and heterogeneous appearance of visceral adipose tissue). This was highlighted by the poor performance of the second

preliminary version. On the other hand, relying on a fat-fraction image only leads to poor reproducibility due to increased segmentation noise. The second embodiment optimally combines the strengths of fat-fraction and fat-only images.

In summary, we propose a method that can automatically and reliably obtain the volumes of abdominal adipose compartments that yields segmentation comparable to that performed by a human expert that additionally exhibits reproducibility rivalling CT- based measurements. Possible applications relate to cardio-metabolic risk reduction assessment and intervention programs.

References

The following references are incorporated by reference in their entirety:

1. Vague, J., La differentiation sexuelle; facteur determinant des formes de

I'obesite. Presse Med, 1947. 55(30): p. 339.

2. Shen, W., et al., Adipose tissue quantification by imaging methods: a proposed classification. Obes Res, 2003. 11(1): p. 5-16.

3. Bjorntorp, P., Body fat distribution, insulin resistance, and metabolic diseases.

Nutrition, 1997. 13(9): p. 795-803.

4. Conway, J.M., et al., Visceral adipose tissue differences in black and white

women. Am J Clin Nutr, 1995. 61 (4): p. 765-71. *

5. Wajchenberg, B.L., Subcutaneous and visceral adipose tissue: their relation to the metabolic syndrome. Endocr Rev, 2000. 21 (6): p. 697-738.

6. Rosito, G.A., et al., Pericardial fat, visceral abdominal fat, cardiovascular disease risk factors, and vascular calcification in a community-based sample: the

Framingham Heart Study. Circulation, 2008. 117(5): p. 605-13.

7. Ross, R., et al., Abdominal adiposity and insulin resistance in obese men. Am J Physiol Endocrinol Metab, 2002. 282(3): p. E657-63.

8. Brochu, M., et al., Visceral adipose tissue is an independent correlate of glucose disposal in older obese postmenopausal women. J Clin Endocrinol Metab, 2000.

85(7): p. 2378-84.

9. Randle, P.J., et al., The glucose fatty-acid cycle. Its role in insulin sensitivity and the metabolic disturbances of diabetes mellitus. Lancet, 1963. 1(7285): p. 785-9.

10. Misra, A. and N.K. Vikram, Clinical and pathophysiological consequences of abdominal adiposity and abdominal adipose tissue depots. Nutrition, 2003. 19(5): p. 457-66. 11. Garg, A., Regional adiposity and insulin resistance. J Glin Endocrinol Metab, 2004. 89(9): p. 4206-10.

12. Grundy, S.M., Does a diagnosis of metabolic syndrome have value in clinical practice? Am J Clin Nutr, 2006. 83(6): p. 1248-51.

13. Alberti, K.G., P. Zimmet, and J. Shaw, Metabolic syndrome-a new world-wide definition. A Consensus Statement from the International Diabetes Federation. Diabetic medicine : a journal of the British Diabetic Association, 2006. 23(5): p. 469-80.

14. Machann, J., et al., Standardized assessment of whole body adipose tissue

topography by MRI. J Magn Reson Imaging, 2005. 21 (4): p. 455-62.

15. van der Kooy, K. and J.C. Seidell, Techniques for the measurement of visceral fat: a practical guide. Int J Obes Relat Metab Disord, 1993. 17(4): p. 187-96.

16. Thomas, E.L., et al., Preferential loss of visceral fat following aerobic exercise, measured by magnetic resonance imaging. Lipids, 2000. 35(7): p. 769-76.

17. Abate, N., et al., Estimation of adipose tissue mass by magnetic resonance

imaging: validation against dissection in human cadavers. J Lipid Res, 1994. 35(8): p. 1490-6.

18. Rossner, S., et al., Adipose tissue determinations in cadavers-a comparison between cross-sectional planimetry and computed tomography. Int J Obes, 1990. 14(10): p. 893-902.

19. Goodpaster, B.H., et al., Subcutaneous abdominal fat and thigh muscle

composition predict insulin sensitivity independently of visceral fat. Diabetes, 1997. 46(10): p. 1579-85.

20. Thomas, E.L. and J.D. Bell, Influence of undersampling on magnetic resonance imaging measurements of intra-abdominal adipose tissue. Int J Obes Relat

Metab Disord, 2003. 27(2): p. 211-8.

21. Kanaley, J.A., I. Giannopoulou, and L.L. Ploutz-Snyder, Regional differences in abdominal fat loss. Int J Obes (Lond), 2007. 31(1): p. 147-52.

22. Positano, V., et al., An accurate and robust method for unsupervised assessment of abdominal fat by MRI. J Magn Reson Imaging, 2004. 20(4): p. 684-9.

23. Kullberg, J., et al., Automated and reproducible segmentation of visceral and subcutaneous adipose tissue from abdominal MRI. Int J Obes (Lond), 2007. 31(12): p. 1806-17.

24. Kullberg, J., et al., Adipose tissue distribution in children: automated

quantification using water and fat MRI. J Magn Reson Imaging, 2010. 32(1): p.

204-10.

25. Liou, T.H., et al., Fully automated large-scale assessment of visceral and

subcutaneous abdominal adipose tissue by magnetic resonance imaging. Int J Obes (Lond), 2006. 30(5): p. 844-52.

26. Bonekamp, S., et al., Quantitative comparison and evaluation of software

packages for assessment of abdominal adipose tissue distribution by magnetic resonance imaging. Int J Obes (Lond), 2008. 32(1): p. 100-11.

27. Yoshizumi, T., et al., Abdominal fat: standardized technique for measurement at CT. Radiology, 1999. 211 (1): p. 283-6. 28. Thaete, F.L., et al., Reproducibility of computed tomography measurement of visceral adipose tissue area. Int J Obes Relat Metab Disord, 1995. 19(7): p. 464- 7.

29. Dixon, W.T., Simple proton spectroscopic imaging. Radiology, 1984, 153(1 ): p.

189-94.

30. Ma, J., Dixon techniques for water and fat imaging. J Magn Reson Imaging, 2008. 28(3): p. 543-58.

31. Sadananthan, S.A., et aL, Skull stripping using graph cuts. Neurolmage, 2010.

49(1 ): p. 225-39.

32. Gray, D.S., et al., Magnetic-resonance imaging used for determining fat

distribution in obesity and diabetes. Am J Clin Nutr, 1991. 54(4): p. 623-7.

33. Gronemeyer, S.A., et al., Fast adipose tissue (FAT) assessment by MRI. Magn Reson Imaging, 2000. 18(7): p. 815-8.

34. Armao, D., et al., Accurate quantification of visceral adipose tissue (VAT) using water-saturation MRI and computer segmentation: preliminary results. J Magn

Reson Imaging, 2006. 23(5): p. 736-41.

35. Peng, Q., et al., Water-saturated three-dimensional balanced steady-state free precession for fast abdominal fat quantification. J Magn Reson Imaging, 2005. 21 (3): p. 263-71.

36. Grady, L. Fast, quality, segmentation of large volumes - Isoperimetric distance trees, in ECCV 2006, Lecture Notes in Computer Science. 2006.

37. Peng, Q., et al., Automated method for accurate abdominal fat quantification on water-saturated magnetic resonance images. J Magn Reson Imaging, 2007. 26(3): p. 738-46.

38. Li, X., et al., Technical evaluation of in vivo abdominal fat and IMCL quantification using MRI and MRSI at 3 T. Magn Reson Imaging, 2008. 26(2): p. 188-97.

39. Jin, Y.P., et al., Segmentation and evaluation of adipose tissue from whole body

MRI scans. Medical Image Computing and Computer-Assisted Intervention -

Miccai 2003, Pt 1 , 2003. 2878: p. 635-642.

40. Boykov, Y. and V. Kolmogorov, An experimental comparison of min-cut/max-flow algorithms for energy minimization in vision. IEEE Transactions on Pattern

Analysis and Machine Intelligence, 2004. 26(9): p. 1 124-1 137.

41. Otsu, N., Threshold Selection Method from Gray-Level Histograms, leee

Transactions on Systems Man and Cybernetics, 1979. 9(1): p. 62-66.

42. Johnson, D.H., et al., Reproducible MRI measurement of adipose tissue volumes in genetic and dietary rodent obesity models. J Magn Reson Imaging, 2008.

28(4): p. 915-27.

Claims

1. A computer-implemented method of segmenting an abdominal image of the adipose tissue of a subject into subcutaneous abdominal adipose tissue (SAT) and visceral abdominal adipose tissue (VAT), the abdominal image comprising a plurality of voxels, the voxels including adipose tissue voxels corresponding to adipose tissue and non-adipose tissue voxels which do not correspond to adipose tissue, the method comprising:

(a) using the abdominal image to generate a graph comprising:

(i) vertices corresponding to respective voxels of the abdominal image, and (ii) edges connecting neighboring said vertices, each said edge being assigned a respective weight;

(b) defining a SAT seed region as a set of voxels in the abdominal image proximate a surface of the abdomen;

(c) defining a background seed region; (d) partitioning the graph into background and SAT portions, said SAT portion of the graph including said SAT seed region and said background portion of the graph including the background seed region, said background and SAT portions of the graph being selected to minimize the sum of the weights of those of the edges which connect a said vertex of the SAT portion of the graph to a said vertex of the background portion of the graph; and

(e) labeling adipose tissue voxels corresponding to vertices within the

background portion of the graph as VAT, and adipose tissue voxels of the abdominal image corresponding to vertices within the SAT portion of the graph as SAT.

2. A method according to claim 1 in which said vertices comprise adipose tissue vertices corresponding to adipose tissue voxels, and the weight of each edge which connects two adipose tissue vertices depends upon at least one of the distances of the corresponding adipose tissue voxels from the nearest non-adipose tissue voxel.

3. A method according to claim 2 in which the weight of each edge which connects two adipose tissue vertices is proportional to the greater of the respective distances of each of the two corresponding adipose tissue voxels from the nearest non-adipose tissue voxel.

4. A method according to any preceding claim further comprising defining the background seed region as the non-adipose tissue voxels.

5. A method according to any preceding claim in which the SAT seed region is obtained by defining an abdominal outer boundary, and generating the SAT seed region as voxels which are within the outer boundary and not more than a certain distance from the outer boundary.

6. A method according to any preceding claim in which the abdominal image is obtained from an MRI image set obtained by the Dixon technique and comprising an in- phase image, an out-phase image, a water-only image and a fat-only image.

7. A method according to claim 6, when dependent on claim 5, comprising a step of identifying air voxels, defining a mask excluding the air voxels, and defining the abdominal outer boundary based on the air mask.

8. A method according to claim 7 further comprising generating the abdominal image by: forming a ratio image having an intensity in each of a plurality of voxels representative of a ratio of (i) an intensity of a corresponding voxel of the fat-only image and (ii) an intensity of a corresponding voxel of a normalization image obtained from the set of images; identifying and labeling voxels of the ratio image for which the intensity is high and which were not identified as air voxels, thereby obtaining a labeled ratio image; and thresholding the fat-only image to form a thresholded fat-only image which constitutes the abdominal image, using a threshold which maximizes a measure of similarity between the thresholded fat-only image and the labeled ratio image.

9. A computer-implemented method of identifying adipose tissue from an abdominal MRI image set obtained by the Dixon technique and comprising an in-phase image, an out-phase image, a water-only image and a fat-only image of the abdomen of a subject, the method comprising: forming a ratio image having an intensity in each of a plurality of voxels representative of a ratio of (i) an intensity of a corresponding voxel of the fat-only image and (ii) an intensity of a corresponding voxel of a normalization image obtained from the set of images; and identifying and labeling voxels of the ratio image for which the intensity is high, thereby obtaining a labeled ratio image.

10. A method according to claim 9 in which the identified voxels of the ratio image are voxels for which the intensity is above a threshold.

11. A method according to claim 9 or claim 10 comprising a step of identifying air voxels, the identified voxels of the ratio image being voxels which are not identified as air voxels.

12. A method according to claim 9, claim 0 or claim 11 comprising selecting a threshold with which to threshold the fat-only image as the threshold which maximizes a measure of similarity between the thresholded fat-only image and the labeled ratio image.

13. A method according to any of claims 8 to 12 in which the intensity of each voxel of the normalization image is the sum of the intensity of a corresponding voxel of the fat- only image and an intensity of a corresponding voxel of the water-only image.

14. A method according to any of claims 8 to 12 in which the ratio image is a fat- fraction image having an intensity in each of the plurality of voxels proportional to a ratio of (i) an intensity of a corresponding voxel of the fat-only image and (ii) the sum of the intensity of the corresponding voxel of the fat-only image and an intensity of a corresponding voxel of the water-only image. 5. A computer system having a processor and a memory, the memory storing program instructions which, when performed by the processor, cause the processor to carry out a method according to any of claims 1-14.

16. A computer program product storing program instructions operative by a processor of a computer system to cause the processor to carry out a method according to any of claims 1-14.